Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

A radiomics-based deep learning approach to predict progression free-survival after tyrosine kinase inhibitor therapy in non-small cell lung cancer

A radiomics-based deep learning approach to predict progression free-survival after tyrosine... Background The epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors ( TKIs) are a first ‑line therapy for non‑small cell lung cancer (NSCLC) with EGFR mutations. Approximately half of the patients with EGFR‑mutated NSCLC are treated with EGFR‑ TKIs and develop disease progression within 1 year. Therefore, the early prediction of tumor progression in patients who receive EGFR‑ TKIs can facilitate patient management and development of treat‑ ment strategies. We proposed a deep learning approach based on both quantitative computed tomography (CT ) characteristics and clinical data to predict progression‑free survival (PFS) in patients with advanced NSCLC after EGFR‑ TKI treatment. Methods A total of 593 radiomic features were extracted from pretreatment chest CT images. The DeepSurv models for the progression risk stratification of EGFR‑ TKI treatment were proposed based on CT radiomic and clinical features from 270 stage IIIB‑IV EGFR‑mutant NSCLC patients. Time ‑ dependent PFS predictions at 3, 12, 18, and 24 months and estimated personalized PFS curves were calculated using the DeepSurv models. Results The model combining clinical and radiomic features demonstrated better prediction performance than the clinical model. The model achieving areas under the curve of 0.76, 0.77, 0.76, and 0.86 can predict PFS at 3, 12, 18, and 24 months, respectively. The personalized PFS curves showed significant differences (p < 0.003) between groups with good (PFS > median) and poor (PFS < median) tumor control. Conclusions The DeepSurv models provided reliable multi‑time ‑point PFS predictions for EGFR‑ TKI treatment. The personalized PFS curves can help make accurate and individualized predictions of tumor progression. The proposed deep learning approach holds promise for improving the pre‑ TKI personalized management of patients with EGFR‑ mutated NSCLC. Keywords Computer tomography imaging, EGFR TKI, Deep learning, Radiomics, Prognostic Chia‑Feng Lu and Chien‑ Yi Liao contributed equally to this work. *Correspondence: Yuh‑Min Chen ymchen@vghtpe.gov.tw Yu‑Te Wu ytwu@nycu.edu.tw Full list of author information is available at the end of the article © The Author(s) 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons. org/ publi cdoma in/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Lu et al. Cancer Imaging (2023) 23:9 Page 2 of 12 can also be applied as independent predictors to comple- Introduction ment clinical information. Lung cancer is the most common malignant neoplastic In this study, we proposed a deep learning-based disease worldwide and is categorized into small cell lung approach to assess the personalized probability of tumor cancer and non-small cell lung cancer (NSCLC), caus- progression in patients who had advanced NSCLC with ing nearly 2 million deaths globally each year [1]. Most EGFR mutations treated with EGFR-TKIs. The proposed NSCLC patients develop relapse of the disease after sur- models combining chest CT radiomic and clinical fea- gery or are even diagnosed as medically inoperable and tures provided a reliable prediction of PFS. We hypoth- therefore have to receive systemic therapies [2]. Patients esized that imaging features extracted from pretreatment with advanced or metastatic NSCLC have to receive chest CT could improve the prediction of tumor progres- systemic therapies for tumor control. The development sion after EGFR-TKI treatment in NSCLC patients. of targeted therapies over the last two decades has con- tributed considerably to the management of NSCLC Materials and methods patients. Epidermal growth factor receptor (EGFR) Study design mutations, mainly exon 19 deletion and exon 21 L858R The Institutional Review Board of Taipei Veterans Gen - mutations, are the most commonly detected oncogenic eral Hospital approved this retrospective study (2021– drivers in approximately 20%–50% of stage IV NSCLC 09-009BCF) and waived the requirement of acquiring patients. Previous studies have indicated that nearly half informed consent from patients. The design of this of NSCLC patients in Asia have EGFR mutations [3, 4]. study is shown in Fig.  1, which includes the inclusion of EGFR tyrosine kinase inhibitors (TKIs) have been dem- patients, the collection of clinical data and standardized onstrated to suppress the growth of NSCLC with EGFR contrast CT imaging features, creation of independent mutations [5]. EGFR mutations are more prevalent in training and testing datasets, selection of key features for Asian NSCLC patients [6]. Accordingly, the application predicting PFS in the training dataset, development of of TKI therapy in NSCLC has received a great deal of deep learning models based on clinical features alone or attention, especially in East Asia [7]. in combination with CT radiomics, and assessment of the The common first-line EGFR-TKI therapy for NSCLC effectiveness of the PFS prediction after TKI treatment includes gefitinib, erlotinib, and afatinib. In phase III in the testing data set. The estimated personalized PFS clinical trials, patients receiving these medications have curves of the model were applied to predict the progres- achieved overall response rates of 56% to 74%, progres- sion risk period and the short-term (3 months), medium- sion-free survival (PFS) of 9.7 to 11.1 months, and overall term (12  months), and long-term (18 and 24  months) survival of 22.9 to 28.2  months [8–10]. However, resist- progression status. This study was performed in accord - ance to EGFR-TKI in patients with NSCLC is frequently ance with the Declaration of Helsinki [22]. observed within 1  year after treatment [11, 12]. There - fore, early identification of patients with a high prob - Patient cohort and image data ability of tumor progression after EGFR-TKI therapy This study retrospectively included 270 EGFR-mutated can facilitate the development of appropriate treatment NSCLC patients treated with EGFR-TKIs from 2017 strategies and is therefore crucial for the management of to 2020. The patient data were collected in accordance advanced NSCLC. Additionally, intra-tumor heteroge- with the following inclusion criteria: (1) identification neity among the postulated molecular mechanisms have of NSCLC with a stage greater than IIIB according to been found to be associated with resistance to EGFR- the American Joint Committee on Cancer (AJCC) stag- TKI therapy [13]. However, some studies have observed ing system, edition 8 [23]; (2) evidence from histological that clinical prognostic factors for evaluating EGFR-TKI examinations of pathology samples from surgical speci- resistance only possess a limited predictive effect because mens or tissue biopsies; (3) receipt of first-line EGFR-TKI of the interplay of molecular mechanisms in NSCLC treatment without surgery, chemotherapy, and radiother- [14, 15]. In recent years, quantitative radiomics analysis apy for NSCLC in accordance with the National Compre- of medical images has been considered as a promising hensive Cancer Network (NCCN) treatment guidelines non-invasive diagnostic method for the study of primary [24]; (4) adequate quality contrast chest CT examination or metastatic lung cancer [16–18]. The proposed Image data and clinical information; and (5) patient without Biomarker Standardization Initiative (IBSI) improves the other neoplastic diseases. reliability of radiomics analysis and further accentuates The quality assessment of CT scans and delineation of its clinical applications based on image quantification the regions of interest (ROIs) was performed by a mul- [19]. Moreover, radiomic features extracted from com- tidisciplinary team of experienced radiologists and certi- puted tomography (CT) images are suggested to evaluate fied pulmonologists. The soft tissue window (width: 350, the heterogeneity of lung lesions [20, 21]. These features Lu  et al. Cancer Imaging (2023) 23:9 Page 3 of 12 Fig. 1 Flowchart for development and validation of DeepSurv model level: 50) and lung window (width: 1500, level: − 600) on using the previously published Multimodal Radiomics CT images were applied for ROI delineation. The soft tis - Platform (available online: http:// cflu. lab. nycu. edu. tw/ sue window was used to distinguish between tumors, col-MRP_ MLing lioma. html, accessed on 6 Sept 2022) [18, lapsed lungs, and fluid components, such as pleural and 27] in compliance with the IBSI on the MATLAB R2022a pericardial effusions, and the lung window was applied to environment [19]. The formulae for radiomics analysis determine the border of tumors. are listed in Table S1. To identify key radiomic and clinical features for pre- Radiomics analysis and feature selection dicting TKI outcomes, a two-step feature selection was The acquired CT images were subjected to numerous applied to the training data set (70% of cases). The initial preprocessing steps before radiomics analysis. First, the statistical method selections, including using univariate resolution of the CT images was adjusted to the identical Cox proportional regression for radiomic features and dimension with a pixel size of 1 × 1 × 1 mm . Second, the the chi-squared test for clinical features, were followed intensities of the CT images were converted into stand- by the implementation of a sequential forward selection ardized ranges (Z-score transformation) based on the (SFS) algorithm [28]. Moreover, to maintain the valid- mean and standard deviation of image data. Finally, low- ity of the deep learning model (i.e., a sufficient number pass (L) and high-pass (H) dimensional wavelet filtering of input features), we applied the selection criterion of were applied to the three axes of CT images, produc- p < 0.1 in the first step (the Cox and chi-squared meth - ing eight image sets: LLL, LLH, LHL, LHH, HLL, HLH, ods). Then, the performance of the proposed PFS predic - HHL, and HHH wavelet filtered images. tion models was evaluated using a testing data set (the Radiomic features, including histogram, geometry, and remaining 30% of cases). texture features (gray level co-occurrence matrix, GLCM; gray level run length matrix, GLRLM; and local binary Prediction models pattern, LBP) [25, 26], were extracted from all image data The DeepSurv model, a multilayer perceptron based sets (eight wavelet decomposed and original CT images). on the Cox proportional hazards (CPH), was applied to In the feature extraction process, the feature aggrega- estimate tumor progression after TKI treatment [29]. tion of GLCM and GLRLM values was performed by Conventional CPH model contains a log-linear regres- averaging over 3D directional matrices according to IBSI sion of relative hazard function that links covariates to guideline to achieve optimal rotational invariance [19]; the patient survival. The DeepSurv model substitutes slice-by-slice computation of the LBP features was fol- the CPH log-linear regression with a multi-layer percep- lowed by histogram analysis of the LBP matrix for all CT tron to estimate the nonlinear properties of the hazard slices. A total of 593 radiomic features were generated for function, and thus has the potential to achieve superior each ROI. All of the ROI delineation, image preprocess- performance in survival prediction. DeepSurv is a con- ing, and subsequent radiomics analysis were performed figurable feed-forward neural network and the input Lu et al. Cancer Imaging (2023) 23:9 Page 4 of 12 to the network is the baseline predictors. The network between models. The feature selection and subsequent propagates the input data through several hidden layers DeepSurv model training were performed on R Deep- with specified weights. The hidden layers include batch Surv package (available online: https:// rdrr. io/ cran/ normalization, nonlinear rectified linear unit (ReLU) survi valmo dels/ src/R/ deeps urv.R, accessed on 6 Sept activation, fully connected, and dropout layers. The final 2022). layer is a single node that conducts a linear combination The optimal thresholds for time-dependent ROC to generate the final output. The hyper-parameters influ - curves at each selected time point were applied to ence the performance of DeepSurv model with regard to intuitively represent personalized progression risks. A the training time, model convergence speed, and predic- Weibull probability distribution function was applied for tion accuracy. Accordingly, the optimization of hyper- curve fitting through the four time-dependent thresholds parameters is essential for model training. to construct a reference risk curve [32]. The area between In this study, the hyper-parameters of network (includ- the reference risk curve and the personalized PFS curve ing number of hidden layers, number of nodes in each was used to assess the risk of tumor progression. If the layer, initial learning rate, learning rate decay, and drop- value of this area during the observed period was nega- out rate) were determined using the grid search method tive, it indicated that the part of the personalized PFS [30]. The setup of hyper-parameters was determined curve was lower than the reference risk curve and rep- based on the prediction performance and training time resented a high risk of progression. A schematic repre- cost (Table S2). Finally, the proposed DeepSurv percep- sentation of the risk-of-progression period is shown in tron consisted of an input layer (the number of nodes Figure S2. was equal to the number of selected features), hidden layers (including batch normalization, ReLU activation, 32-node fully connected, and dropout layers), and an out- Results put layer. Moreover, an Adam optimizer with a dropout Clinical characteristics of recruited patients rate of 40%, an initial learning rate of 0.01, a learning rate In the present study, the median PFS of the recruited decay of 0.01, and L2 regularization was performed in the patients with NSCLC after TKI treatment was training process. The loss function of DeepSurv models 11.5  months. Over 74% of patients were nonsmokers. was defined as the average negative log partial likelihood The majority of patients had stage IV adenocarcinoma proposed in a previous study [29]. The architecture of (97.4%) and exhibited EGFR exon 19 deletions (43.3%) proposed DeepSurv model is shown in Figure S1. and exon 21 L858R substitutions (49.3%). Over half of the Three DeepSurv models were developed using the patients showed no adverse effects after TKI treatment. following information: (1) clinical features (e.g., AJCC Table 1 summarizes the clinical characteristics of the 270 TNM stage, smoking status, and the histopathology of recruited NSCLC patients. No significant differences in NSCLC); (2) radiomic features; and (3) a combination of clinical characteristics were identified between the train - clinical and radiomic features. The DeepSurv models are ing and test sets (Table S3). The results of clinical labora - used to estimate personalized PFS curves for each case tory tests are listed in Table S4. based on the corresponding logarithmic hazard func- tion. A log-rank test was performed to assess the statisti- cal difference in the average of personalized PFS curves Selected features for PFS prediction between the good control (PFS > 11.5  months) and poor We ultimately selected 10 features, including 5 clinical control (PFS < 11.5 months) groups. and 5 radiomic features, through the two-step feature The DeepSurv models were further applied to pre - selection process. The key clinical features for predict - dict the progression status at individual follow-up time ing tumor progression included regional lymph node points (e.g., 3, 12, 18, and 24  months). The predictive metastasis, distant metastasis of the tumor, NSCLC efficacy of the DeepSurv models in predicting pro - histology, total protein, and mean corpuscular volume. gression status was evaluated using time-dependent Selected radiomic features included textural features that receiver operating characteristic (ROC) curves, area described local homogeneity and one geometric feature under the ROC curve (AUC), index of concordance that measured the compactness of tumor shape com- (C-index), sensitivity, and specificity. A bootstrap ran - pared to a sphere. Four of the five selected radiomic fea - dom sampling method was applied to the testing data tures were GLCM features—GLCM inverse difference set to statistically compare the prediction performance moment normalized (IDMN) based on LHL, HLL, HHL, of the clinical/radiomic and combined (clinical and and HHH wavelets—and the remaining one is a geome- radiomic features) DeepSurv models [31]. A paired try feature—compactness. The details of the selected fea - t test was used to compare the difference in the AUC tures are listed in Table S5. Lu  et al. Cancer Imaging (2023) 23:9 Page 5 of 12 Table 1 Characteristics of 270 recruited NSCLC patients Table 1 (continued) Characteristics Value Characteristics Value Mean corpuscular volume Age, median(IQR) 67.5 (60–75) High, N(%) 85 (31.5) Gender Normal, N(%) 81 (30.0) Female, N(%) 158 (58.5) Low, N(%) 5 (1.9) Smoking status Not available, N(%) 99 (37.6) Smoker, N(%) 69 (25.6) White blood cells count ECOG PS score High, N(%) 54 (20.0) 0, N(%) 98 (36.3) Normal, N(%) 185 (68.5) 1, N(%) 139 (51.5) Low, N(%) 7 (2.6) 2, N(%) 22 (8.1) Not available, N(%) 24 (8.9) > 2, N(%) 11 (4.1) Definitions: total proteins: high: > 6 g/dl, low: < 6 g/dl; mean corpuscular volume: Histology of NSCLC high > 100 fl, normal:80-100 fl, low < 80 fl; white blood cells count: high: > 12,000/ Adenocarcinoma, N(%) 263 (97.4) cumm, normal 4000–12,000/cumm, low: < 4000/cumm Squamous cell carcinoma, N(%) 4 (1.5) Abbreviations: ECOG Eastern Cooperative Oncology Group, EGFR Epidermal Others, N(%) 3 (1.1) growth factor receptor, NSCLC Non-small cell lung cancer, PS Performance status, TKI Tyrosine kinase inhibitor Clinical T stage 1, N(%) 35 (13.0) 2, N(%) 78 (28.9) Performance of DeepSurv prediction models 3, N(%) 44 (16.3) The patients were divided into good con - 4, N(%) 106 (39.2) trol (PFS > 11.5  months) and poor control Not available, N(%) 7 (2.6) (PFS < 11.5  months) groups based on the median PFS, Clinical N stage and we sought to evaluate the prediction efficacy of 0, N(%) 69 (25.6) the DeepSurv prediction models. Figure  2 displays the 1, N(%) 20 (7.4) personalized PFS curves estimated using the DeepSurv 2, N(%) 75 (27.8) models based on the radiomic, clinical, and combined 3, N(%) 104 (38.5) (clinical and radiomic features) datasets, respectively. Not available, N(%) 2 (0.7) Our results demonstrated that personalized survival Clinical M stage curves generated by clinical and combined DeepSurv 0, N(%) 10 (3.7) models differed significantly between the two tumor 1a, N(%) 82 (30.4) control groups (p < 0.002), which indicated that both 1b, N(%) 41 (15.2) models provided reliable predictions in differentiat - 1c, N(%) 137 (50.7) ing tumor responses to TKI treatment. However, aver- Clinical stage age PFS curves estimated by the model solely based Stage III, N(%) 23 (8.5) on radiomic features were not significantly different Stage IVA, N(%) 110 (40.7) (p = 0.35) between the good and poor control groups. Stage IVB, N(%) 137 (50.8) We estimated the prediction performance at four EGFR mutation status follow-up time points (3, 12, 18, and 24  months) by Exon 19 deletion, N(%) 117 (43.3) using the testing data set. The time-dependent ROC Exon 21 L858R substitution, N(%) 133 (49.3) curves for each model are shown in Fig.  3. The radi- Others, N(%) 20 (7.4) omic feature-based model produced AUCs between TKI 0.49 and 0.69 with C-index of 0.57. The clinical fea- Gefitinib, N(%) 46 (17.0) ture-based model produced AUCs between 0.71 and Erlotinib, N(%) 85 (31.5) 0.72 with a C-index of 0.63. The combined model Afatinib, N(%) 139 (51.5) had an AUC range of 0.76 to 0.86 with a C-index of Adverse drug reaction to EGFR-TKI 0.66. Table  2 lists the comprehensive performance Yes, N(%) 130 (48.1) and statistical comparisons between radiomic/clini- Progression free survival, median(months) 11.5 (4.9–17.9) cal and combined models. Overall, the combined Total protein model significantly outperformed the radiomic and High, N(%) 87 (32.2) clinical models in terms of efficacy (AUC, sensitivity, Low, N(%) 6 (2.2) and specificity) in predicting progression risk at each Not available, N(%) 177 (65.6) selected time point. Lu et al. Cancer Imaging (2023) 23:9 Page 6 of 12 Fig. 2 Distribution of personalized PFS curves predicted by the DeepSurv models. The estimated personalized PFS curves of patients in training set, testing set, and average of personalized PFS curves in testing set based on (a) radiomic, (b) clinical, and (c) combined model, respectively. The red curves in the figure represented the patients with PFS better than median PFS, and blue curves indicated the patients with PFS poorer than median PFS Figure  4 illustrates the prediction of the risk-of- (1.0  month) time. For patients with a long PFS (without progression period for representative cases with long any regional lymph node metastasis or bone metastases, (24.3  months), moderate (11.9  months), and short PFS Fig.  4a), both clinical and combined models generated Lu  et al. Cancer Imaging (2023) 23:9 Page 7 of 12 Fig. 3 Results of time‑ dependent prediction of PFS after TKI treatment. The time‑ dependent ROC curves of DeepSurv models for predicting PFS after TKI treatment based on (a) radiomic, (b) clinical, and (c) combined model, respectively Table 2 Statistical comparisons between developed prediction models based on test dataset Model performance Radiomic Clinical (C-index = 0.63) Combined p-values (C-index = 0.57) (C-index = 0.66) Radiomic vs. Clinical vs. Combined Combined 3 months Original AUC 0.49 0.71 0.76 a a AUC 0.52 ± 0.05 0.70 ± 0.06 0.75 ± 0.06 < 0.001 < 0.001 a a Sensitivity 0.48 ± 0.10 0.66 ± 0.10 0.68 ± 0.05 < 0.001 0.03 Specificity 0.64 ± 0.09 0.69 ± 0.06 0.73 ± 0.13 < 0.001 < 0.001 12 months Original AUC 0.59 0.71 0.77 a a AUC 0.60 ± 0.11 0.70 ± 0.06 0.78 ± 0.05 < 0.001 < 0.001 a a Sensitivity 0.51 ± 0.04 0.67 ± 0.11 0.69 ± 0.06 < 0.001 0.02 a a Specificity 0.60 ± 0.08 0.62 ± 0.07 0.70 ± 0.11 < 0.001 < 0.001 18 months Original AUC 0.69 0.72 0.76 a a AUC 0.70 ± 0.05 0.71 ± 0.08 0.78 ± 0.05 < 0.001 < 0.001 Sensitivity 0.60 ± 0.10 0.67 ± 1.12 0.65 ± 0.06 < 0.001 0.07 a a Specificity 0.67 ± 0.03 0.62 ± 0.06 0.80 ± 0.14 < 0.001 < 0.001 24 months Original AUC 0.67 0.71 0.86 a a AUC 0.70 ± 0.10 0.69 ± 0.13 0.85 ± 0.05 < 0.001 < 0.001 a a Sensitivity 0.71 ± 0.08 0.71 ± 0.13 0.89 ± 0.06 < 0.001 < 0.001 a a Specificity 0.75 ± 0.13 0.78 ± 0.06 0.80 ± 0.14 < 0.001 0.03 Significant difference based on the paired t-test personalized PFS curves that were higher than the ref- curve and the reference risk curve during the period of erence risk curve. This indicated the models accurately 3 to 12  months. This indicated that two models accu - predicted a risk-of-progression period of longer than rately predicted a risk-of-progression period between 3 24  months. For patients with a moderate PFS (without and 12 months. As for patients with a short PFS (having any regional lymph node metastasis but having lung and regional lymph node metastases and bone and pleural pleural metastases, Fig.  4b), only the combined model metastases, Fig.  4c), both clinical and combined models identified an intersection between the personalized PFS estimated personalized PFS curves that were lower than Lu et al. Cancer Imaging (2023) 23:9 Page 8 of 12 Fig. 4 Representative cases for the predictions of PFS based on different data set. Figure shows CT images and the DeepSurv risk ‑ of‑progression period of (a) a patient with long (24.3 months) PFS, (b) a patient with moderate (11.9 months) PFS, and (c) a patient with short (1.0 month) PFS. The comparison of the selected geometric feature with the PFS of representative cases is presented in (d) the reference risk curves, indicating that the accurate adverse drug reactions and facilitate the early implemen- prediction of a risk-of-progression period was less than tation of necessary treatments. In patients with NSCLC, 3  months. Patients with long PFS had a higher value of contrast-enhanced chest CT remains the standard imag- compactness (one of the selected radiomic features) ing test for the initial diagnosis of NSCLC. Nevertheless, reflecting a rounder-shaped lesion than those with mod - according to our survey, CT images without contrast erate or short PFS (Fig.  4d). The results suggested that enhancement constitute the majority of the publicly the combined models provided reliable estimates of the accessible NSCLC imaging databases. Furthermore, to risk-of-progression period for patients with NSCLC after enrich the imaging database of NSCLC, a prognostic EGFR-TKI therapy. model based on contrast-enhanced CT images should be proposed that considers the possibility of combining Discussion multimodality CT into the data set. First-generation EGFR-TKIs (gefitinib and erlotinib) and Studies have revealed the potential of radiomic fea- the second-generation EGFR-TKI( afatinib) have been tures extracted from CT images to predict outcomes of used as the first-line treatment of NSCLC in the last dec - TKI therapy in patients with advanced NSCLC [37, 38]. ade [33, 34]. Patients with EGFR-mutant NSCLC who The multivariate CPH models, the most extensively used were treated with EGFR-TKIs had an improved PFS com- survival analysis approach, have been applied in several pared with those treated with standard chemotherapy studies. However, the calculation of linear covariance [35]. The most common reason for discontinuing EGFR- between variables using the CPH model does not provide TKI therapy is tumor progression, and therefore, per- a reliable assessment of therapeutic outcomes because sonalized prediction of EGFR-TKI resistance is notable the covariation between prognostic factors is mostly non- [36]. Hence, a reliable prediction could prevent potential linear. Moreover, this limitation becomes more evident Lu  et al. Cancer Imaging (2023) 23:9 Page 9 of 12 when high-throughput radiomic features are further red blood cell mean corpuscular volume may indicate a exploited as prognostic factors in the regression model. deficiency of folate, resulting in abnormal methylation, Therefore, we applied the DeepSurv model to predict synthesis, replication, and repair of DNA [43, 44]. How- tumor progression in NSCLC patients. The DeepSurv ever, because not all patients received hematology tests, model that features a multilayer neural network provides the association of these features with TKI treatment out- a reliable nonlinear regression of covariates between comes requires further investigation. Smoking, a known prognostic factors [29]. Furthermore, estimated person- independent prognostic factor for NSCLC, was not con- alized PFS curves from the DeepSurv model provide an sidered for the following reasons. First, the feature data intuitive approach for prognostic evaluation. Estimated set may contain indicators that are highly correlated with risk-of-progression periods are allowed for the prediction smoking; therefore, smoking was given less weightage of TKIs resistance in NSCLC patients for personalizing in the SFS algorithm. Second, only approximately 25% treatment strategies and management. of the patients in this study were smokers, which could In a previous study, clinical-based CPH models with have affected prediction accuracy due to data imbalance. a C-index of 0.62 to 0.63 were proposed to predict PFS Third, smokers were related to a high incidence of non- after EGFR-TKI treatment in NSCLC patients. A CPH EGFR-mutant lung cancers [45], implying that this factor model based on CT radiomics has been further used for had a confounding effect. Hence, tumor stage and labora - time-dependent PFS prediction. The models achieved tory data may be considered to assess the efficacy of TKI AUCs ranging from 0.70 to 0.82 in predicting PFS at 10 therapy in NSCLC patients. and 12  months. This indicated different data sets could Even though the radiomic model itself was not sufficient lead to bias in the prediction performance of the model to accurately predict the PFS, our results demonstrated [38]. Our proposed combined model exhibited a more that the synergetic effect of combined model (including accurate prediction performance than the clinical and both radiomic and clinical features) showed significant radiomic models and achieved a C-index of 0.66. Moreo- enhancement of prediction performance. Compactness ver, the model had reliable efficacy in predicting PFS at and IDMN were the selected radiomic features for PFS 3, 12, 18, and 24 months (achieving AUCs of 0.75–0.86), prediction. The results revealed that patients with poor and its high prediction performance could be attributed PFS had reduced values of compactness in their CT to two reasons. First, the radiomics process in the present testing. In addition, low compactness indicates that the study was conducted in accordance with the IBSI guide- tumor exhibits a more asymmetric geometry relative to line [19]. Standardized image quantification enhanced a spherical tumor and has been reported to be associated the stability of radiomic features and the reliability of with a highly aggressive form of tumor [46]. This find - prediction models. Second, we applied the DeepSurv ing implied that the high local aggressiveness of NSCLC model to simulate the nonlinear interactions between was one of the main causes of EGFR-TKI resistance. In predictors. This may facilitate the adaptation of models addition, patients with poor PFS exhibited high values to changes in the risk of tumor progression at different of IDMN on CT scans. High IDMN values indicate that time points. the voxel intensity of the image is locally similar. EGFR In the two-step feature selection process, the histology mutated NSCLC is recognized to be highly angiogenic of NSCLC and the AJCC pathological N and M stages and venous aggressive [47] and is linked to a low IDMN were identified as the key clinical factors for predicting value on contrast CT images. Therefore, NSCLC patients PFS after EGFR-TKI treatment. The presence of squa - with low IDMN values on CT images can be expected to mous NSCLC and lymph node metastases are known have a high level of EGFR mutations and a good EGFR- prognostic factors for advanced lung cancer [39, 40]. We TKI response. further categorized the metastatic states as M0, M1a, In this study, the application of DeepSurv model was M1b, and M1c staging based on the number of occur- suggested to evaluate the risk-of-progression period of rences location [41]. Our results indicated that patients NSCLC patients. Estimated personalized PFS curves with multiple distant metastases had a poor prognosis. describe the probability of tumor progression after Moreover, we considered two commonly used labora- EGFR-TKI treatment. As tumor progression may occur tory features, namely total protein and red blood cell at different times during follow-up, time-dependent mean corpuscular volume, in the analysis. Patients with ROC curves can be used to assess the progression sta- low total protein and high mean corpuscular volume are tus at critical follow-up time points. Figure  4 indicates associated with poor PFS. A low total protein level may that the clinical model provided a reliable PFS predic- reflect patient exhaustion, which may cause patients to tion for patients with good and poor tumor control. This have severe constitutional symptoms and the inability to implied that the prognostic effect of different stages of withstand intensive treatment [42]. Increasing values of conventional tumor staging was significant. The clinical Lu et al. Cancer Imaging (2023) 23:9 Page 10 of 12 SFS Sequential forward selection prediction model performed poorly because patients CPH Cox proportional hazards with moderate tumor control frequently presented at ROC Receiver operating characteristic similar clinical stages. The DeepSurv model incorporat - AUC Area under the curve C‑index Index of concordance ing radiomic features provided information on tumor IDMN Inverse difference moment normalized heterogeneity. The combined model also incorporated the tumor heterogeneity data from radiomics, which Supplementary Information allowed the model to more effectively differentiate the The online version contains supplementary material available at https:// doi. prognosis between patients with similar tumor stages. org/ 10. 1186/ s40644‑ 023‑ 00522‑5. Several limitations and further considerations of this study are discussed as follows. First, the CT images and Additional file 1: Table S1. The formulae for the calculation of primary radiomic features. Table S2. Grid search results of DeepSurv hyper‑param‑ therapeutic information of patients in this study were eters. Table S3. Comparisons of clinical characteristics between training acquired from a single institution. The proposed models and test sets. Table S4. Characteristics of clinical laboratory test. Table S5. should be validated with an external validation data set Identified features for the model training in each DeepSurv model. Figure S1. The architecture of applied DeepSurv model. Figure S2. Schematic from multiple centers in future research. Second, the diagram of predictive risk‑ of‑progression period in DeepSurv model. tumor segmentation in this study was performed manu- ally by a multidisciplinary team of experienced pulmo- Acknowledgements nologist and radiologists based on different CT windows. This manuscript was edited by Wallace Academic Editing. The development of an automated CT image segmenta - tion method could reduce the time required for man- Authors’ contributions Conception and design: CF Lu, CY Liao. Acquisition of data: HS Chao, HY Chiu, ual segmentation and improve the reproducibility and TH Shiao, YM Chen. Analysis and interpretation of data: CF Lu, CY Liao, TW robustness of radiomic features. Finally, clinical labora- Wang, Y Lee, JR Chen. Statistical analysis: CF Lu, CY Liao. Drafting the article: tory information of all patients was not available due to CF Lu, CY Liao, HY Chiu. Critically revising the article: all authors. Reviewed and approved submitted version of manuscript: all authors. Study supervision: HS the retrospective nature of the study. Future studies are Chao, TH Shiao, YM Chen, Y T, Wu. expected to prospectively collect the proposed key clini- cal aspects of data. Funding This work was supported by AICS, ASUSTeK Computer Incorporation, Taiwan (110J042) and Veterans General Hospitals and University System of Taiwan Joint Research Program ( VGHUST112‑ G1‑3–3). The funding sources had no Conclusions role in the design and conduct of the study; collection, management, analysis, The information on the staging, histology, and blood or interpretation of the data; preparation, review, or approval of the manu‑ analysis results of NSCLCs patients could be used to pro- script; and decision to submit the manuscript for publication. vide a reliable prediction of possible tumor progression Availability of data and materials after EGFR-TKI treatment. The additional inclusion of The raw data cannot be made publicly available for ethical and legal reasons. quantitative CT characteristics describing tumor com- However, researchers can submit inquiries for analyzed data to the corre‑ sponding authors upon reasonable request. pactness and local homogeneity further improved the predictive performance of the models. The risk-of-pro - Declarations gression period based on the DeepSurv model can pro- vide personalized predictions of therapeutic outcomes Ethics approval and consent to participate after EGFR-TKI treatment in a more intuitive man- The Institutional Review Board of Taipei Veterans General Hospital approved this retrospective study (Project Identification Number: 2021–09‑009BCF) and ner and may help personalize treatment strategies for waived the requirement of acquiring informed consent from patients. advanced NSCLC patients who have received EGFR-TKI treatment. Consent for publication Not applicable. Competing interests Abbreviations The authors declare no conflict of interest. NSCLC Non‑small cell lung cancer EGFR Epidermal growth factor receptor Author details TKIs Tyrosine kinase inhibitors Department of Biomedical Imaging and Radiological Sciences, National Yang PFS Progression‑free survival Ming Chiao Tung University, Taipei, Taiwan. Department of Chest Medicine, IBSI Image Biomarker Standardization Initiative Taipei Veteran General Hospital, Taipei, Taiwan. Institute of Biophotonics, CT Computed tomography National Yang Ming Chiao Tung University, Taipei, Taiwan. School of Medicine, AJCC American Joint Committee on Cancer National Yang Ming Chiao Tung University, Taipei, Taiwan. Brain Research NCCN National Comprehensive Cancer Network Center, National Yang Ming Chiao Tung University, Taipei, Taiwan. ROIs Regions of interest GLCM Gray level co‑ occurrence matrix Received: 21 September 2022 Accepted: 5 January 2023 GLRLM Gray level run length matrix, LBP Local binary pattern Lu  et al. Cancer Imaging (2023) 23:9 Page 11 of 12 References 19. Zwanenburg A, Vallières M, Abdalah MA, Aerts HJ, Andrearczyk V, Apte 1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global A, Ashrafinia S, Bakas S, Beukinga RJ, Boellaard R. The image biomarker cancer statistics 2018: GLOBOCAN estimates of incidence and mor‑ standardization initiative: standardized quantitative radiomics for high‑ tality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. throughput image‑based phenotyping. Radiology. 2020;295(2):328–38. 2018;68(6):394–424. 20. Cucchiara F, Del Re M, Valleggi S, Romei C, Petrini I, Lucchesi M, Crucitta 2. Arbour KC, Riely GJ. Systemic therapy for locally advanced and metastatic S, Rofi E, De Liperi A, Chella A. Integrating liquid biopsy and radiomics non–small cell lung cancer: a review. JAMA. 2019;322(8):764–74. to monitor clonal heterogeneity of EGFR‑positive non‑small cell lung 3. Hsu C‑H, Tseng C‑H, Chiang C‑ J, Hsu K‑H, Tseng J‑S, Chen K ‑ C, Wang cancer. Front Oncol. 2020;10:593831. C‑L, Chen C‑ Y, Yen S‑H, Chiu C‑H. Characteristics of young lung cancer: 21. Park BW, Kim JK, Heo C, Park KJ. Reliability of CT radiomic features Analysis of Taiwan’s nationwide lung cancer registry focusing on epider‑ reflecting tumour heterogeneity according to image quality and image mal growth factor receptor mutation and smoking status. Oncotarget. processing parameters. Sci Rep. 2020;10(1):1–13. 2016;7(29):46628. 22. Goodyear MD, Krleza‑ Jeric K, Lemmens T. The declaration of Helsinki. Br 4. Zhang Y‑L, Yuan J‑ Q, Wang K‑F, Fu X ‑H, Han X ‑R, Threapleton D, Yang Med J Publishing Group. 2007;335:624–5. Z‑ Y, Mao C, Tang J‑L. The prevalence of EGFR mutation in patients with 23. Amin MB, Greene FL, Edge SB, Compton CC, Gershenwald JE, Brookland non‑small cell lung cancer: a systematic review and meta‑analysis. Onco ‑ RK, Meyer L, Gress DM, Byrd DR, Winchester DP. The eighth edition AJCC target. 2016;7(48):78985. cancer staging manual: continuing to build a bridge from a population‐ 5. Ruiz‑ Cordero R, Devine WP. Targeted therapy and checkpoint immuno‑ based to a more “personalized” approach to cancer staging. CA Cancer J therapy in lung cancer. Surg Pathol Clin. 2020;13(1):17–33. Clin. 2017;67(2):93–9. 6. Zhou F, Zhou C. Lung cancer in never smokers—the East Asian experi‑ 24. Ettinger DS, Wood DE, Akerley W, Bazhenova LA, Borghaei H, Camidge ence. Transl Lung Cancer Res. 2018;7(4):450. DR, Cheney RT, Chirieac LR, D’Amico TA, Dilling TJ. NCCN guidelines 7. Kim ES, Melosky B, Park K, Yamamoto N, Yang JC. EGFR tyrosine kinase insights: non–small cell lung cancer, version 4.2016. J Natl Compr Cancer inhibitors for EGFR mutation‑positive non‑small‑ cell lung cancer: out‑ Netw. 2016;14(3):255–64. comes in Asian populations. Future Oncol. 2021;17(18):2395–408. 25. Dhruv B, Mittal N, Modi M. Study of Haralick’s and GLCM texture analysis 8. Yang JC‑H, Wu Y ‑L, Schuler M, Sebastian M, Popat S, Yamamoto N, Zhou C, on 3D medical images. Int J Neurosci. 2019;129(4):350–62. Hu C‑P, O’Byrne K, Feng J. Afatinib versus cisplatin‑based chemotherapy 26. García‑ Olalla Ó, Fernández‑Robles L, Alegre E, Castejón‑Limas M, Fidalgo for EGFR mutation‑positive lung adenocarcinoma (LUX ‑Lung 3 and LUX ‑ E. Boosting texture‑based classification by describing statistical informa‑ Lung 6): analysis of overall survival data from two randomised, phase 3 tion of gray‑levels differences. Sensors. 2019;19(5):1048. trials. Lancet Oncol. 2015;16(2):141–51. 27. Lu C‑F, Hsu F‑ T, Hsieh KL‑ C, Kao Y‑ CJ, Cheng S‑ J, Hsu JB‑K, Tsai PH, Chen 9. Rosell R, Carcereny E, Gervais R, Vergnenegre A, Massuti B, Felip E, R‑ J, Huang C‑ C, Yen Y. Machine learning–based radiomics for molecular Palmero R, Garcia‑ Gomez R, Pallares C, Sanchez JM. Erlotinib versus subtyping of gliomas. Clin Cancer Res. 2018;24(18):4429–36. standard chemotherapy as first ‑line treatment for European patients with 28 Mao KZ. Orthogonal forward selection and backward elimination algo‑ advanced EGFR mutation‑positive non‑small‑ cell lung cancer (EURTAC): rithms for feature subset selection. IEEE Trans Syst Man Cybern B (Cybern). a multicentre, open‑label, randomised phase 3 trial. Lancet Oncol. 2004;34(1):629–34. 2012;13(3):239–46. 29. Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: 10. Inoue A, Kobayashi K, Maemondo M, Sugawara S, Oizumi S, Isobe H, personalized treatment recommender system using a Cox proportional Gemma A, Harada M, Yoshizawa H, Kinoshita I. Updated overall survival hazards deep neural network. BMC Med Res Methodol. 2018;18(1):1–12. results from a randomized phase III trial comparing gefitinib with 30. Bergstra J, Bengio Y. Random search for hyper‑parameter optimization. J carboplatin–paclitaxel for chemo‑naïve non‑small cell lung cancer with Mach Learn Res. 2012;13(2):281–305. sensitive EGFR gene mutations (NEJ002). Ann Oncol. 2013;24(1):54–9. 31. Dixon PM. Bootstrap resampling. In: Encyclopedia of environmetrics. 2006. 11. Apicella M, Giannoni E, Fiore S, Ferrari KJ, Fernández‑Pérez D, Isella C, 32. Weibull W. A statistical distribution function of wide applicability. J Appl Granchi C, Minutolo F, Sottile A, Comoglio PM. Increased lactate secretion Mech. 1951;18:290–3. by cancer cells sustains non‑ cell‑autonomous adaptive resistance to MET 33. Cataldo VD, Gibbons DL, Pérez‑Soler R, Quintás‑ Cardama A. Treatment and EGFR targeted therapies. Cell Metab. 2018;28(6):848‑865.e846. of non–small‑ cell lung cancer with erlotinib or gefitinib. N Engl J Med. 12. Wu Y‑L, Zhou C, Liam C‑K, Wu G, Liu X, Zhong Z, Lu S, Cheng Y, Han B, 2011;364(10):947–55. Chen L. First‑line erlotinib versus gemcitabine/cisplatin in patients with 34. Bersanelli M, Tiseo M, Artioli F, Lucchi L, Ardizzoni A. Gefitinib and afatinib advanced EGFR mutation‑positive non‑small‑ cell lung cancer: analyses treatment in an advanced non‑small cell lung cancer (NSCLC) patient from the phase III, randomized, open‑label, ENSURE study. Ann Oncol. undergoing hemodialysis. Anticancer Res. 2014;34(6):3185–8. 2015;26(9):1883–9. 35. Lee CK, Brown C, Gralla RJ, Hirsh V, Thongprasert S, Tsai C‑M, Tan EH. Ho 13. Zhao Y, Wang H, He C. Drug resistance of targeted therapy for JC‑M, Chu DT, Zaatar A: Impact of EGFR inhibitor in non–small cell lung advanced non‑small cell lung cancer harbored EGFR mutation: From cancer on progression‑free and overall survival: a meta‑analysis. J Natl mechanism analysis to clinical strategy. J Cancer Res Clin Oncol. Cancer Inst. 2013;105(9):595–605. 2021;147(12):3653–64. 36. Yu HA, Arcila ME, Rekhtman N, Sima CS, Zakowski MF, Pao W, Kris MG, 14. Garg A, Batra U, Choudhary P, Jain D, Khurana S, Malik PS, Muthu V, Prasad Miller VA, Ladanyi M, Riely GJ. Analysis of Tumor Specimens at the Time K, Singh N, Suri T. Clinical predictors of response to EGFR‑tyrosine kinase of Acquired Resistance to EGFR‑ TKI Therapy in 155 Patients with EGFR‑ inhibitors in EGFR‑mutated non‑small cell lung cancer: a real‑ world multi‑ Mutant Lung CancersMechanisms of Acquired Resistance to EGFR‑ TKI centric cohort analysis from India. Curr Probl Cancer. 2020;44(3):100570. Therapy. Clin Cancer Res. 2013;19(8):2240–7. 15. Buonerba C, Iaccarino S, Dolce P, Pagliuca M, Izzo M, Scafuri L, Costabile F, 37. Li H, Zhang R, Wang S, Fang M, Zhu Y, Hu Z, Dong D, Shi J, Tian J. CT‑ Riccio V, Ribera D, Mucci B. Predictors of outcomes in patients with EGFR‑ based radiomic signature as a prognostic factor in stage IV ALK‑positive mutated non‑small cell lung cancer receiving EGFR tyrosine kinase inhibi‑ non‑small‑ cell lung cancer treated with TKI crizotinib: a proof‑ of‑ concept tors: a systematic review and meta‑analysis. Cancers. 2019;11(9):1259. study. Front Oncol. 2020;10:57. 16. Lambin P, Leijenaar RT, Deist TM, Peerlings J, De Jong EE, Van Timmeren 38. Song J, Shi J, Dong D, Fang M, Zhong W, Wang K, Wu N, Huang Y, Liu Z, J, Sanduleanu S, Larue RT, Even AJ, Jochems A. Radiomics: the bridge Cheng Y. A New Approach to Predict Progressionfr ‑ ee Survival in Stage IV between medical imaging and personalized medicine. Nat Rev Clin EGFRmutant NSCL ‑ C Patients with EGFR‑ TKI TherapyPrediction of EGFR‑ TKI Oncol. 2017;14(12):749–62. Treatment Outcome in Stage IV NSCLC. Clin Cancer Res. 2018;24(15):3583–92. 17. Scrivener M, de Jong EE, van Timmeren JE, Pieters T, Ghaye B, Geets 39. Jin R, Peng L, Shou J, Wang J, Jin Y, Liang F, Zhao J, Wu M, Li Q, Zhang X. Radiomics applied to lung cancer: a review. Transl Cancer Res. B. EGFR‑mutated squamous cell lung cancer and its association with 2016;5(4):398–409. outcomes. Front Oncol. 2021;11:2262. 18. Liao C‑ Y, Lee C‑ C, Yang H‑ C, Chen C‑ J, Chung W‑ Y, Wu H‑M, Guo W ‑ Y, Liu 40. Masters GA, Temin S, Azzoli CG, Giaccone G, Baker S Jr, Brahmer JR, Ellis R‑S, Lu C‑F. Enhancement of Radiosurgical Treatment Outcome Prediction PM, Gajra A, Rackear N, Schiller JH. Systemic therapy for stage IV non– Using MRI Radiomics in Patients with Non‑Small Cell Lung Cancer Brain small‑ cell lung cancer: American Society of Clinical Oncology clinical Metastases. Cancers. 2021;13(16):4030. practice guideline update. J Clin Oncol. 2015;33(30):3488. Lu et al. Cancer Imaging (2023) 23:9 Page 12 of 12 41. Detterbeck FC, Boffa DJ, Kim AW, Tanoue LT. The eighth edition lung cancer stage classification. Chest. 2017;151(1):193–203. 42. Watanabe T, Kinoshita T, Itoh K, Yoshimura K, Ogura M, Kagami Y, Yamagu‑ chi M, Kurosawa M, Tsukasaki K, Kasai M. Pretreatment total serum protein is a significant prognostic factor for the outcome of patients with periph‑ eral T/natural killer‑ cell lymphomas. Leuk Lymphoma. 2010;51(5):813–21. 43. Li K‑j, Gu W ‑ y, Xia X‑f. Zhang P, Zou C‑l, Fei Z ‑h: High Mean corpuscular volume as a predictor of poor overall survival in patients with esophageal cancer receiving concurrent chemoradiotherapy. Cancer Manag Res. 2020;12:7467. 44. Kim Y‑I. Will mandatory folic acid fortification prevent or promote cancer? Am J Clin Nutr. 2004;80(5):1123–8. 45. Ren JH, He WS, Yan GL, Jin M, Yang KY, Wu G. EGFR mutations in non‑ small‑ cell lung cancer among smokers and non‑smokers: A meta‑analy‑ sis. Environ Mol Mutagen. 2012;53(1):78–82. 46. Apostolova I, Rogasch J, Buchert R, Wertzel H, Achenbach HJ, Schreiber J, Riedel S, Furth C, Lougovski A, Schramm G. Quantitative assessment of the asphericity of pretherapeutic FDG uptake as an independent predic‑ tor of outcome in NSCLC. BMC Cancer. 2014;14(1):1–10. 47. van Cruijsen H, Giaccone G, Hoekman K. Epidermal growth factor recep‑ tor and angiogenesis: Opportunities for combined anticancer strategies. Int J Cancer. 2005;117(6):883–8. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ lished maps and institutional affiliations. Re Read ady y to to submit y submit your our re researc search h ? Choose BMC and benefit fr ? Choose BMC and benefit from om: : fast, convenient online submission thorough peer review by experienced researchers in your field rapid publication on acceptance support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations maximum visibility for your research: over 100M website views per year At BMC, research is always in progress. Learn more biomedcentral.com/submissions http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Cancer Imaging Springer Journals

A radiomics-based deep learning approach to predict progression free-survival after tyrosine kinase inhibitor therapy in non-small cell lung cancer

Download PDF
12 pages

Loading next page...
 
/lp/springer-journals/a-radiomics-based-deep-learning-approach-to-predict-progression-free-gz7s7adF3K

References (55)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2023
eISSN
1470-7330
DOI
10.1186/s40644-023-00522-5
Publisher site
See Article on Publisher Site

Abstract

Background The epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors ( TKIs) are a first ‑line therapy for non‑small cell lung cancer (NSCLC) with EGFR mutations. Approximately half of the patients with EGFR‑mutated NSCLC are treated with EGFR‑ TKIs and develop disease progression within 1 year. Therefore, the early prediction of tumor progression in patients who receive EGFR‑ TKIs can facilitate patient management and development of treat‑ ment strategies. We proposed a deep learning approach based on both quantitative computed tomography (CT ) characteristics and clinical data to predict progression‑free survival (PFS) in patients with advanced NSCLC after EGFR‑ TKI treatment. Methods A total of 593 radiomic features were extracted from pretreatment chest CT images. The DeepSurv models for the progression risk stratification of EGFR‑ TKI treatment were proposed based on CT radiomic and clinical features from 270 stage IIIB‑IV EGFR‑mutant NSCLC patients. Time ‑ dependent PFS predictions at 3, 12, 18, and 24 months and estimated personalized PFS curves were calculated using the DeepSurv models. Results The model combining clinical and radiomic features demonstrated better prediction performance than the clinical model. The model achieving areas under the curve of 0.76, 0.77, 0.76, and 0.86 can predict PFS at 3, 12, 18, and 24 months, respectively. The personalized PFS curves showed significant differences (p < 0.003) between groups with good (PFS > median) and poor (PFS < median) tumor control. Conclusions The DeepSurv models provided reliable multi‑time ‑point PFS predictions for EGFR‑ TKI treatment. The personalized PFS curves can help make accurate and individualized predictions of tumor progression. The proposed deep learning approach holds promise for improving the pre‑ TKI personalized management of patients with EGFR‑ mutated NSCLC. Keywords Computer tomography imaging, EGFR TKI, Deep learning, Radiomics, Prognostic Chia‑Feng Lu and Chien‑ Yi Liao contributed equally to this work. *Correspondence: Yuh‑Min Chen ymchen@vghtpe.gov.tw Yu‑Te Wu ytwu@nycu.edu.tw Full list of author information is available at the end of the article © The Author(s) 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons. org/ publi cdoma in/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Lu et al. Cancer Imaging (2023) 23:9 Page 2 of 12 can also be applied as independent predictors to comple- Introduction ment clinical information. Lung cancer is the most common malignant neoplastic In this study, we proposed a deep learning-based disease worldwide and is categorized into small cell lung approach to assess the personalized probability of tumor cancer and non-small cell lung cancer (NSCLC), caus- progression in patients who had advanced NSCLC with ing nearly 2 million deaths globally each year [1]. Most EGFR mutations treated with EGFR-TKIs. The proposed NSCLC patients develop relapse of the disease after sur- models combining chest CT radiomic and clinical fea- gery or are even diagnosed as medically inoperable and tures provided a reliable prediction of PFS. We hypoth- therefore have to receive systemic therapies [2]. Patients esized that imaging features extracted from pretreatment with advanced or metastatic NSCLC have to receive chest CT could improve the prediction of tumor progres- systemic therapies for tumor control. The development sion after EGFR-TKI treatment in NSCLC patients. of targeted therapies over the last two decades has con- tributed considerably to the management of NSCLC Materials and methods patients. Epidermal growth factor receptor (EGFR) Study design mutations, mainly exon 19 deletion and exon 21 L858R The Institutional Review Board of Taipei Veterans Gen - mutations, are the most commonly detected oncogenic eral Hospital approved this retrospective study (2021– drivers in approximately 20%–50% of stage IV NSCLC 09-009BCF) and waived the requirement of acquiring patients. Previous studies have indicated that nearly half informed consent from patients. The design of this of NSCLC patients in Asia have EGFR mutations [3, 4]. study is shown in Fig.  1, which includes the inclusion of EGFR tyrosine kinase inhibitors (TKIs) have been dem- patients, the collection of clinical data and standardized onstrated to suppress the growth of NSCLC with EGFR contrast CT imaging features, creation of independent mutations [5]. EGFR mutations are more prevalent in training and testing datasets, selection of key features for Asian NSCLC patients [6]. Accordingly, the application predicting PFS in the training dataset, development of of TKI therapy in NSCLC has received a great deal of deep learning models based on clinical features alone or attention, especially in East Asia [7]. in combination with CT radiomics, and assessment of the The common first-line EGFR-TKI therapy for NSCLC effectiveness of the PFS prediction after TKI treatment includes gefitinib, erlotinib, and afatinib. In phase III in the testing data set. The estimated personalized PFS clinical trials, patients receiving these medications have curves of the model were applied to predict the progres- achieved overall response rates of 56% to 74%, progres- sion risk period and the short-term (3 months), medium- sion-free survival (PFS) of 9.7 to 11.1 months, and overall term (12  months), and long-term (18 and 24  months) survival of 22.9 to 28.2  months [8–10]. However, resist- progression status. This study was performed in accord - ance to EGFR-TKI in patients with NSCLC is frequently ance with the Declaration of Helsinki [22]. observed within 1  year after treatment [11, 12]. There - fore, early identification of patients with a high prob - Patient cohort and image data ability of tumor progression after EGFR-TKI therapy This study retrospectively included 270 EGFR-mutated can facilitate the development of appropriate treatment NSCLC patients treated with EGFR-TKIs from 2017 strategies and is therefore crucial for the management of to 2020. The patient data were collected in accordance advanced NSCLC. Additionally, intra-tumor heteroge- with the following inclusion criteria: (1) identification neity among the postulated molecular mechanisms have of NSCLC with a stage greater than IIIB according to been found to be associated with resistance to EGFR- the American Joint Committee on Cancer (AJCC) stag- TKI therapy [13]. However, some studies have observed ing system, edition 8 [23]; (2) evidence from histological that clinical prognostic factors for evaluating EGFR-TKI examinations of pathology samples from surgical speci- resistance only possess a limited predictive effect because mens or tissue biopsies; (3) receipt of first-line EGFR-TKI of the interplay of molecular mechanisms in NSCLC treatment without surgery, chemotherapy, and radiother- [14, 15]. In recent years, quantitative radiomics analysis apy for NSCLC in accordance with the National Compre- of medical images has been considered as a promising hensive Cancer Network (NCCN) treatment guidelines non-invasive diagnostic method for the study of primary [24]; (4) adequate quality contrast chest CT examination or metastatic lung cancer [16–18]. The proposed Image data and clinical information; and (5) patient without Biomarker Standardization Initiative (IBSI) improves the other neoplastic diseases. reliability of radiomics analysis and further accentuates The quality assessment of CT scans and delineation of its clinical applications based on image quantification the regions of interest (ROIs) was performed by a mul- [19]. Moreover, radiomic features extracted from com- tidisciplinary team of experienced radiologists and certi- puted tomography (CT) images are suggested to evaluate fied pulmonologists. The soft tissue window (width: 350, the heterogeneity of lung lesions [20, 21]. These features Lu  et al. Cancer Imaging (2023) 23:9 Page 3 of 12 Fig. 1 Flowchart for development and validation of DeepSurv model level: 50) and lung window (width: 1500, level: − 600) on using the previously published Multimodal Radiomics CT images were applied for ROI delineation. The soft tis - Platform (available online: http:// cflu. lab. nycu. edu. tw/ sue window was used to distinguish between tumors, col-MRP_ MLing lioma. html, accessed on 6 Sept 2022) [18, lapsed lungs, and fluid components, such as pleural and 27] in compliance with the IBSI on the MATLAB R2022a pericardial effusions, and the lung window was applied to environment [19]. The formulae for radiomics analysis determine the border of tumors. are listed in Table S1. To identify key radiomic and clinical features for pre- Radiomics analysis and feature selection dicting TKI outcomes, a two-step feature selection was The acquired CT images were subjected to numerous applied to the training data set (70% of cases). The initial preprocessing steps before radiomics analysis. First, the statistical method selections, including using univariate resolution of the CT images was adjusted to the identical Cox proportional regression for radiomic features and dimension with a pixel size of 1 × 1 × 1 mm . Second, the the chi-squared test for clinical features, were followed intensities of the CT images were converted into stand- by the implementation of a sequential forward selection ardized ranges (Z-score transformation) based on the (SFS) algorithm [28]. Moreover, to maintain the valid- mean and standard deviation of image data. Finally, low- ity of the deep learning model (i.e., a sufficient number pass (L) and high-pass (H) dimensional wavelet filtering of input features), we applied the selection criterion of were applied to the three axes of CT images, produc- p < 0.1 in the first step (the Cox and chi-squared meth - ing eight image sets: LLL, LLH, LHL, LHH, HLL, HLH, ods). Then, the performance of the proposed PFS predic - HHL, and HHH wavelet filtered images. tion models was evaluated using a testing data set (the Radiomic features, including histogram, geometry, and remaining 30% of cases). texture features (gray level co-occurrence matrix, GLCM; gray level run length matrix, GLRLM; and local binary Prediction models pattern, LBP) [25, 26], were extracted from all image data The DeepSurv model, a multilayer perceptron based sets (eight wavelet decomposed and original CT images). on the Cox proportional hazards (CPH), was applied to In the feature extraction process, the feature aggrega- estimate tumor progression after TKI treatment [29]. tion of GLCM and GLRLM values was performed by Conventional CPH model contains a log-linear regres- averaging over 3D directional matrices according to IBSI sion of relative hazard function that links covariates to guideline to achieve optimal rotational invariance [19]; the patient survival. The DeepSurv model substitutes slice-by-slice computation of the LBP features was fol- the CPH log-linear regression with a multi-layer percep- lowed by histogram analysis of the LBP matrix for all CT tron to estimate the nonlinear properties of the hazard slices. A total of 593 radiomic features were generated for function, and thus has the potential to achieve superior each ROI. All of the ROI delineation, image preprocess- performance in survival prediction. DeepSurv is a con- ing, and subsequent radiomics analysis were performed figurable feed-forward neural network and the input Lu et al. Cancer Imaging (2023) 23:9 Page 4 of 12 to the network is the baseline predictors. The network between models. The feature selection and subsequent propagates the input data through several hidden layers DeepSurv model training were performed on R Deep- with specified weights. The hidden layers include batch Surv package (available online: https:// rdrr. io/ cran/ normalization, nonlinear rectified linear unit (ReLU) survi valmo dels/ src/R/ deeps urv.R, accessed on 6 Sept activation, fully connected, and dropout layers. The final 2022). layer is a single node that conducts a linear combination The optimal thresholds for time-dependent ROC to generate the final output. The hyper-parameters influ - curves at each selected time point were applied to ence the performance of DeepSurv model with regard to intuitively represent personalized progression risks. A the training time, model convergence speed, and predic- Weibull probability distribution function was applied for tion accuracy. Accordingly, the optimization of hyper- curve fitting through the four time-dependent thresholds parameters is essential for model training. to construct a reference risk curve [32]. The area between In this study, the hyper-parameters of network (includ- the reference risk curve and the personalized PFS curve ing number of hidden layers, number of nodes in each was used to assess the risk of tumor progression. If the layer, initial learning rate, learning rate decay, and drop- value of this area during the observed period was nega- out rate) were determined using the grid search method tive, it indicated that the part of the personalized PFS [30]. The setup of hyper-parameters was determined curve was lower than the reference risk curve and rep- based on the prediction performance and training time resented a high risk of progression. A schematic repre- cost (Table S2). Finally, the proposed DeepSurv percep- sentation of the risk-of-progression period is shown in tron consisted of an input layer (the number of nodes Figure S2. was equal to the number of selected features), hidden layers (including batch normalization, ReLU activation, 32-node fully connected, and dropout layers), and an out- Results put layer. Moreover, an Adam optimizer with a dropout Clinical characteristics of recruited patients rate of 40%, an initial learning rate of 0.01, a learning rate In the present study, the median PFS of the recruited decay of 0.01, and L2 regularization was performed in the patients with NSCLC after TKI treatment was training process. The loss function of DeepSurv models 11.5  months. Over 74% of patients were nonsmokers. was defined as the average negative log partial likelihood The majority of patients had stage IV adenocarcinoma proposed in a previous study [29]. The architecture of (97.4%) and exhibited EGFR exon 19 deletions (43.3%) proposed DeepSurv model is shown in Figure S1. and exon 21 L858R substitutions (49.3%). Over half of the Three DeepSurv models were developed using the patients showed no adverse effects after TKI treatment. following information: (1) clinical features (e.g., AJCC Table 1 summarizes the clinical characteristics of the 270 TNM stage, smoking status, and the histopathology of recruited NSCLC patients. No significant differences in NSCLC); (2) radiomic features; and (3) a combination of clinical characteristics were identified between the train - clinical and radiomic features. The DeepSurv models are ing and test sets (Table S3). The results of clinical labora - used to estimate personalized PFS curves for each case tory tests are listed in Table S4. based on the corresponding logarithmic hazard func- tion. A log-rank test was performed to assess the statisti- cal difference in the average of personalized PFS curves Selected features for PFS prediction between the good control (PFS > 11.5  months) and poor We ultimately selected 10 features, including 5 clinical control (PFS < 11.5 months) groups. and 5 radiomic features, through the two-step feature The DeepSurv models were further applied to pre - selection process. The key clinical features for predict - dict the progression status at individual follow-up time ing tumor progression included regional lymph node points (e.g., 3, 12, 18, and 24  months). The predictive metastasis, distant metastasis of the tumor, NSCLC efficacy of the DeepSurv models in predicting pro - histology, total protein, and mean corpuscular volume. gression status was evaluated using time-dependent Selected radiomic features included textural features that receiver operating characteristic (ROC) curves, area described local homogeneity and one geometric feature under the ROC curve (AUC), index of concordance that measured the compactness of tumor shape com- (C-index), sensitivity, and specificity. A bootstrap ran - pared to a sphere. Four of the five selected radiomic fea - dom sampling method was applied to the testing data tures were GLCM features—GLCM inverse difference set to statistically compare the prediction performance moment normalized (IDMN) based on LHL, HLL, HHL, of the clinical/radiomic and combined (clinical and and HHH wavelets—and the remaining one is a geome- radiomic features) DeepSurv models [31]. A paired try feature—compactness. The details of the selected fea - t test was used to compare the difference in the AUC tures are listed in Table S5. Lu  et al. Cancer Imaging (2023) 23:9 Page 5 of 12 Table 1 Characteristics of 270 recruited NSCLC patients Table 1 (continued) Characteristics Value Characteristics Value Mean corpuscular volume Age, median(IQR) 67.5 (60–75) High, N(%) 85 (31.5) Gender Normal, N(%) 81 (30.0) Female, N(%) 158 (58.5) Low, N(%) 5 (1.9) Smoking status Not available, N(%) 99 (37.6) Smoker, N(%) 69 (25.6) White blood cells count ECOG PS score High, N(%) 54 (20.0) 0, N(%) 98 (36.3) Normal, N(%) 185 (68.5) 1, N(%) 139 (51.5) Low, N(%) 7 (2.6) 2, N(%) 22 (8.1) Not available, N(%) 24 (8.9) > 2, N(%) 11 (4.1) Definitions: total proteins: high: > 6 g/dl, low: < 6 g/dl; mean corpuscular volume: Histology of NSCLC high > 100 fl, normal:80-100 fl, low < 80 fl; white blood cells count: high: > 12,000/ Adenocarcinoma, N(%) 263 (97.4) cumm, normal 4000–12,000/cumm, low: < 4000/cumm Squamous cell carcinoma, N(%) 4 (1.5) Abbreviations: ECOG Eastern Cooperative Oncology Group, EGFR Epidermal Others, N(%) 3 (1.1) growth factor receptor, NSCLC Non-small cell lung cancer, PS Performance status, TKI Tyrosine kinase inhibitor Clinical T stage 1, N(%) 35 (13.0) 2, N(%) 78 (28.9) Performance of DeepSurv prediction models 3, N(%) 44 (16.3) The patients were divided into good con - 4, N(%) 106 (39.2) trol (PFS > 11.5  months) and poor control Not available, N(%) 7 (2.6) (PFS < 11.5  months) groups based on the median PFS, Clinical N stage and we sought to evaluate the prediction efficacy of 0, N(%) 69 (25.6) the DeepSurv prediction models. Figure  2 displays the 1, N(%) 20 (7.4) personalized PFS curves estimated using the DeepSurv 2, N(%) 75 (27.8) models based on the radiomic, clinical, and combined 3, N(%) 104 (38.5) (clinical and radiomic features) datasets, respectively. Not available, N(%) 2 (0.7) Our results demonstrated that personalized survival Clinical M stage curves generated by clinical and combined DeepSurv 0, N(%) 10 (3.7) models differed significantly between the two tumor 1a, N(%) 82 (30.4) control groups (p < 0.002), which indicated that both 1b, N(%) 41 (15.2) models provided reliable predictions in differentiat - 1c, N(%) 137 (50.7) ing tumor responses to TKI treatment. However, aver- Clinical stage age PFS curves estimated by the model solely based Stage III, N(%) 23 (8.5) on radiomic features were not significantly different Stage IVA, N(%) 110 (40.7) (p = 0.35) between the good and poor control groups. Stage IVB, N(%) 137 (50.8) We estimated the prediction performance at four EGFR mutation status follow-up time points (3, 12, 18, and 24  months) by Exon 19 deletion, N(%) 117 (43.3) using the testing data set. The time-dependent ROC Exon 21 L858R substitution, N(%) 133 (49.3) curves for each model are shown in Fig.  3. The radi- Others, N(%) 20 (7.4) omic feature-based model produced AUCs between TKI 0.49 and 0.69 with C-index of 0.57. The clinical fea- Gefitinib, N(%) 46 (17.0) ture-based model produced AUCs between 0.71 and Erlotinib, N(%) 85 (31.5) 0.72 with a C-index of 0.63. The combined model Afatinib, N(%) 139 (51.5) had an AUC range of 0.76 to 0.86 with a C-index of Adverse drug reaction to EGFR-TKI 0.66. Table  2 lists the comprehensive performance Yes, N(%) 130 (48.1) and statistical comparisons between radiomic/clini- Progression free survival, median(months) 11.5 (4.9–17.9) cal and combined models. Overall, the combined Total protein model significantly outperformed the radiomic and High, N(%) 87 (32.2) clinical models in terms of efficacy (AUC, sensitivity, Low, N(%) 6 (2.2) and specificity) in predicting progression risk at each Not available, N(%) 177 (65.6) selected time point. Lu et al. Cancer Imaging (2023) 23:9 Page 6 of 12 Fig. 2 Distribution of personalized PFS curves predicted by the DeepSurv models. The estimated personalized PFS curves of patients in training set, testing set, and average of personalized PFS curves in testing set based on (a) radiomic, (b) clinical, and (c) combined model, respectively. The red curves in the figure represented the patients with PFS better than median PFS, and blue curves indicated the patients with PFS poorer than median PFS Figure  4 illustrates the prediction of the risk-of- (1.0  month) time. For patients with a long PFS (without progression period for representative cases with long any regional lymph node metastasis or bone metastases, (24.3  months), moderate (11.9  months), and short PFS Fig.  4a), both clinical and combined models generated Lu  et al. Cancer Imaging (2023) 23:9 Page 7 of 12 Fig. 3 Results of time‑ dependent prediction of PFS after TKI treatment. The time‑ dependent ROC curves of DeepSurv models for predicting PFS after TKI treatment based on (a) radiomic, (b) clinical, and (c) combined model, respectively Table 2 Statistical comparisons between developed prediction models based on test dataset Model performance Radiomic Clinical (C-index = 0.63) Combined p-values (C-index = 0.57) (C-index = 0.66) Radiomic vs. Clinical vs. Combined Combined 3 months Original AUC 0.49 0.71 0.76 a a AUC 0.52 ± 0.05 0.70 ± 0.06 0.75 ± 0.06 < 0.001 < 0.001 a a Sensitivity 0.48 ± 0.10 0.66 ± 0.10 0.68 ± 0.05 < 0.001 0.03 Specificity 0.64 ± 0.09 0.69 ± 0.06 0.73 ± 0.13 < 0.001 < 0.001 12 months Original AUC 0.59 0.71 0.77 a a AUC 0.60 ± 0.11 0.70 ± 0.06 0.78 ± 0.05 < 0.001 < 0.001 a a Sensitivity 0.51 ± 0.04 0.67 ± 0.11 0.69 ± 0.06 < 0.001 0.02 a a Specificity 0.60 ± 0.08 0.62 ± 0.07 0.70 ± 0.11 < 0.001 < 0.001 18 months Original AUC 0.69 0.72 0.76 a a AUC 0.70 ± 0.05 0.71 ± 0.08 0.78 ± 0.05 < 0.001 < 0.001 Sensitivity 0.60 ± 0.10 0.67 ± 1.12 0.65 ± 0.06 < 0.001 0.07 a a Specificity 0.67 ± 0.03 0.62 ± 0.06 0.80 ± 0.14 < 0.001 < 0.001 24 months Original AUC 0.67 0.71 0.86 a a AUC 0.70 ± 0.10 0.69 ± 0.13 0.85 ± 0.05 < 0.001 < 0.001 a a Sensitivity 0.71 ± 0.08 0.71 ± 0.13 0.89 ± 0.06 < 0.001 < 0.001 a a Specificity 0.75 ± 0.13 0.78 ± 0.06 0.80 ± 0.14 < 0.001 0.03 Significant difference based on the paired t-test personalized PFS curves that were higher than the ref- curve and the reference risk curve during the period of erence risk curve. This indicated the models accurately 3 to 12  months. This indicated that two models accu - predicted a risk-of-progression period of longer than rately predicted a risk-of-progression period between 3 24  months. For patients with a moderate PFS (without and 12 months. As for patients with a short PFS (having any regional lymph node metastasis but having lung and regional lymph node metastases and bone and pleural pleural metastases, Fig.  4b), only the combined model metastases, Fig.  4c), both clinical and combined models identified an intersection between the personalized PFS estimated personalized PFS curves that were lower than Lu et al. Cancer Imaging (2023) 23:9 Page 8 of 12 Fig. 4 Representative cases for the predictions of PFS based on different data set. Figure shows CT images and the DeepSurv risk ‑ of‑progression period of (a) a patient with long (24.3 months) PFS, (b) a patient with moderate (11.9 months) PFS, and (c) a patient with short (1.0 month) PFS. The comparison of the selected geometric feature with the PFS of representative cases is presented in (d) the reference risk curves, indicating that the accurate adverse drug reactions and facilitate the early implemen- prediction of a risk-of-progression period was less than tation of necessary treatments. In patients with NSCLC, 3  months. Patients with long PFS had a higher value of contrast-enhanced chest CT remains the standard imag- compactness (one of the selected radiomic features) ing test for the initial diagnosis of NSCLC. Nevertheless, reflecting a rounder-shaped lesion than those with mod - according to our survey, CT images without contrast erate or short PFS (Fig.  4d). The results suggested that enhancement constitute the majority of the publicly the combined models provided reliable estimates of the accessible NSCLC imaging databases. Furthermore, to risk-of-progression period for patients with NSCLC after enrich the imaging database of NSCLC, a prognostic EGFR-TKI therapy. model based on contrast-enhanced CT images should be proposed that considers the possibility of combining Discussion multimodality CT into the data set. First-generation EGFR-TKIs (gefitinib and erlotinib) and Studies have revealed the potential of radiomic fea- the second-generation EGFR-TKI( afatinib) have been tures extracted from CT images to predict outcomes of used as the first-line treatment of NSCLC in the last dec - TKI therapy in patients with advanced NSCLC [37, 38]. ade [33, 34]. Patients with EGFR-mutant NSCLC who The multivariate CPH models, the most extensively used were treated with EGFR-TKIs had an improved PFS com- survival analysis approach, have been applied in several pared with those treated with standard chemotherapy studies. However, the calculation of linear covariance [35]. The most common reason for discontinuing EGFR- between variables using the CPH model does not provide TKI therapy is tumor progression, and therefore, per- a reliable assessment of therapeutic outcomes because sonalized prediction of EGFR-TKI resistance is notable the covariation between prognostic factors is mostly non- [36]. Hence, a reliable prediction could prevent potential linear. Moreover, this limitation becomes more evident Lu  et al. Cancer Imaging (2023) 23:9 Page 9 of 12 when high-throughput radiomic features are further red blood cell mean corpuscular volume may indicate a exploited as prognostic factors in the regression model. deficiency of folate, resulting in abnormal methylation, Therefore, we applied the DeepSurv model to predict synthesis, replication, and repair of DNA [43, 44]. How- tumor progression in NSCLC patients. The DeepSurv ever, because not all patients received hematology tests, model that features a multilayer neural network provides the association of these features with TKI treatment out- a reliable nonlinear regression of covariates between comes requires further investigation. Smoking, a known prognostic factors [29]. Furthermore, estimated person- independent prognostic factor for NSCLC, was not con- alized PFS curves from the DeepSurv model provide an sidered for the following reasons. First, the feature data intuitive approach for prognostic evaluation. Estimated set may contain indicators that are highly correlated with risk-of-progression periods are allowed for the prediction smoking; therefore, smoking was given less weightage of TKIs resistance in NSCLC patients for personalizing in the SFS algorithm. Second, only approximately 25% treatment strategies and management. of the patients in this study were smokers, which could In a previous study, clinical-based CPH models with have affected prediction accuracy due to data imbalance. a C-index of 0.62 to 0.63 were proposed to predict PFS Third, smokers were related to a high incidence of non- after EGFR-TKI treatment in NSCLC patients. A CPH EGFR-mutant lung cancers [45], implying that this factor model based on CT radiomics has been further used for had a confounding effect. Hence, tumor stage and labora - time-dependent PFS prediction. The models achieved tory data may be considered to assess the efficacy of TKI AUCs ranging from 0.70 to 0.82 in predicting PFS at 10 therapy in NSCLC patients. and 12  months. This indicated different data sets could Even though the radiomic model itself was not sufficient lead to bias in the prediction performance of the model to accurately predict the PFS, our results demonstrated [38]. Our proposed combined model exhibited a more that the synergetic effect of combined model (including accurate prediction performance than the clinical and both radiomic and clinical features) showed significant radiomic models and achieved a C-index of 0.66. Moreo- enhancement of prediction performance. Compactness ver, the model had reliable efficacy in predicting PFS at and IDMN were the selected radiomic features for PFS 3, 12, 18, and 24 months (achieving AUCs of 0.75–0.86), prediction. The results revealed that patients with poor and its high prediction performance could be attributed PFS had reduced values of compactness in their CT to two reasons. First, the radiomics process in the present testing. In addition, low compactness indicates that the study was conducted in accordance with the IBSI guide- tumor exhibits a more asymmetric geometry relative to line [19]. Standardized image quantification enhanced a spherical tumor and has been reported to be associated the stability of radiomic features and the reliability of with a highly aggressive form of tumor [46]. This find - prediction models. Second, we applied the DeepSurv ing implied that the high local aggressiveness of NSCLC model to simulate the nonlinear interactions between was one of the main causes of EGFR-TKI resistance. In predictors. This may facilitate the adaptation of models addition, patients with poor PFS exhibited high values to changes in the risk of tumor progression at different of IDMN on CT scans. High IDMN values indicate that time points. the voxel intensity of the image is locally similar. EGFR In the two-step feature selection process, the histology mutated NSCLC is recognized to be highly angiogenic of NSCLC and the AJCC pathological N and M stages and venous aggressive [47] and is linked to a low IDMN were identified as the key clinical factors for predicting value on contrast CT images. Therefore, NSCLC patients PFS after EGFR-TKI treatment. The presence of squa - with low IDMN values on CT images can be expected to mous NSCLC and lymph node metastases are known have a high level of EGFR mutations and a good EGFR- prognostic factors for advanced lung cancer [39, 40]. We TKI response. further categorized the metastatic states as M0, M1a, In this study, the application of DeepSurv model was M1b, and M1c staging based on the number of occur- suggested to evaluate the risk-of-progression period of rences location [41]. Our results indicated that patients NSCLC patients. Estimated personalized PFS curves with multiple distant metastases had a poor prognosis. describe the probability of tumor progression after Moreover, we considered two commonly used labora- EGFR-TKI treatment. As tumor progression may occur tory features, namely total protein and red blood cell at different times during follow-up, time-dependent mean corpuscular volume, in the analysis. Patients with ROC curves can be used to assess the progression sta- low total protein and high mean corpuscular volume are tus at critical follow-up time points. Figure  4 indicates associated with poor PFS. A low total protein level may that the clinical model provided a reliable PFS predic- reflect patient exhaustion, which may cause patients to tion for patients with good and poor tumor control. This have severe constitutional symptoms and the inability to implied that the prognostic effect of different stages of withstand intensive treatment [42]. Increasing values of conventional tumor staging was significant. The clinical Lu et al. Cancer Imaging (2023) 23:9 Page 10 of 12 SFS Sequential forward selection prediction model performed poorly because patients CPH Cox proportional hazards with moderate tumor control frequently presented at ROC Receiver operating characteristic similar clinical stages. The DeepSurv model incorporat - AUC Area under the curve C‑index Index of concordance ing radiomic features provided information on tumor IDMN Inverse difference moment normalized heterogeneity. The combined model also incorporated the tumor heterogeneity data from radiomics, which Supplementary Information allowed the model to more effectively differentiate the The online version contains supplementary material available at https:// doi. prognosis between patients with similar tumor stages. org/ 10. 1186/ s40644‑ 023‑ 00522‑5. Several limitations and further considerations of this study are discussed as follows. First, the CT images and Additional file 1: Table S1. The formulae for the calculation of primary radiomic features. Table S2. Grid search results of DeepSurv hyper‑param‑ therapeutic information of patients in this study were eters. Table S3. Comparisons of clinical characteristics between training acquired from a single institution. The proposed models and test sets. Table S4. Characteristics of clinical laboratory test. Table S5. should be validated with an external validation data set Identified features for the model training in each DeepSurv model. Figure S1. The architecture of applied DeepSurv model. Figure S2. Schematic from multiple centers in future research. Second, the diagram of predictive risk‑ of‑progression period in DeepSurv model. tumor segmentation in this study was performed manu- ally by a multidisciplinary team of experienced pulmo- Acknowledgements nologist and radiologists based on different CT windows. This manuscript was edited by Wallace Academic Editing. The development of an automated CT image segmenta - tion method could reduce the time required for man- Authors’ contributions Conception and design: CF Lu, CY Liao. Acquisition of data: HS Chao, HY Chiu, ual segmentation and improve the reproducibility and TH Shiao, YM Chen. Analysis and interpretation of data: CF Lu, CY Liao, TW robustness of radiomic features. Finally, clinical labora- Wang, Y Lee, JR Chen. Statistical analysis: CF Lu, CY Liao. Drafting the article: tory information of all patients was not available due to CF Lu, CY Liao, HY Chiu. Critically revising the article: all authors. Reviewed and approved submitted version of manuscript: all authors. Study supervision: HS the retrospective nature of the study. Future studies are Chao, TH Shiao, YM Chen, Y T, Wu. expected to prospectively collect the proposed key clini- cal aspects of data. Funding This work was supported by AICS, ASUSTeK Computer Incorporation, Taiwan (110J042) and Veterans General Hospitals and University System of Taiwan Joint Research Program ( VGHUST112‑ G1‑3–3). The funding sources had no Conclusions role in the design and conduct of the study; collection, management, analysis, The information on the staging, histology, and blood or interpretation of the data; preparation, review, or approval of the manu‑ analysis results of NSCLCs patients could be used to pro- script; and decision to submit the manuscript for publication. vide a reliable prediction of possible tumor progression Availability of data and materials after EGFR-TKI treatment. The additional inclusion of The raw data cannot be made publicly available for ethical and legal reasons. quantitative CT characteristics describing tumor com- However, researchers can submit inquiries for analyzed data to the corre‑ sponding authors upon reasonable request. pactness and local homogeneity further improved the predictive performance of the models. The risk-of-pro - Declarations gression period based on the DeepSurv model can pro- vide personalized predictions of therapeutic outcomes Ethics approval and consent to participate after EGFR-TKI treatment in a more intuitive man- The Institutional Review Board of Taipei Veterans General Hospital approved this retrospective study (Project Identification Number: 2021–09‑009BCF) and ner and may help personalize treatment strategies for waived the requirement of acquiring informed consent from patients. advanced NSCLC patients who have received EGFR-TKI treatment. Consent for publication Not applicable. Competing interests Abbreviations The authors declare no conflict of interest. NSCLC Non‑small cell lung cancer EGFR Epidermal growth factor receptor Author details TKIs Tyrosine kinase inhibitors Department of Biomedical Imaging and Radiological Sciences, National Yang PFS Progression‑free survival Ming Chiao Tung University, Taipei, Taiwan. Department of Chest Medicine, IBSI Image Biomarker Standardization Initiative Taipei Veteran General Hospital, Taipei, Taiwan. Institute of Biophotonics, CT Computed tomography National Yang Ming Chiao Tung University, Taipei, Taiwan. School of Medicine, AJCC American Joint Committee on Cancer National Yang Ming Chiao Tung University, Taipei, Taiwan. Brain Research NCCN National Comprehensive Cancer Network Center, National Yang Ming Chiao Tung University, Taipei, Taiwan. ROIs Regions of interest GLCM Gray level co‑ occurrence matrix Received: 21 September 2022 Accepted: 5 January 2023 GLRLM Gray level run length matrix, LBP Local binary pattern Lu  et al. Cancer Imaging (2023) 23:9 Page 11 of 12 References 19. Zwanenburg A, Vallières M, Abdalah MA, Aerts HJ, Andrearczyk V, Apte 1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global A, Ashrafinia S, Bakas S, Beukinga RJ, Boellaard R. The image biomarker cancer statistics 2018: GLOBOCAN estimates of incidence and mor‑ standardization initiative: standardized quantitative radiomics for high‑ tality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. throughput image‑based phenotyping. Radiology. 2020;295(2):328–38. 2018;68(6):394–424. 20. Cucchiara F, Del Re M, Valleggi S, Romei C, Petrini I, Lucchesi M, Crucitta 2. Arbour KC, Riely GJ. Systemic therapy for locally advanced and metastatic S, Rofi E, De Liperi A, Chella A. Integrating liquid biopsy and radiomics non–small cell lung cancer: a review. JAMA. 2019;322(8):764–74. to monitor clonal heterogeneity of EGFR‑positive non‑small cell lung 3. Hsu C‑H, Tseng C‑H, Chiang C‑ J, Hsu K‑H, Tseng J‑S, Chen K ‑ C, Wang cancer. Front Oncol. 2020;10:593831. C‑L, Chen C‑ Y, Yen S‑H, Chiu C‑H. Characteristics of young lung cancer: 21. Park BW, Kim JK, Heo C, Park KJ. Reliability of CT radiomic features Analysis of Taiwan’s nationwide lung cancer registry focusing on epider‑ reflecting tumour heterogeneity according to image quality and image mal growth factor receptor mutation and smoking status. Oncotarget. processing parameters. Sci Rep. 2020;10(1):1–13. 2016;7(29):46628. 22. Goodyear MD, Krleza‑ Jeric K, Lemmens T. The declaration of Helsinki. Br 4. Zhang Y‑L, Yuan J‑ Q, Wang K‑F, Fu X ‑H, Han X ‑R, Threapleton D, Yang Med J Publishing Group. 2007;335:624–5. Z‑ Y, Mao C, Tang J‑L. The prevalence of EGFR mutation in patients with 23. Amin MB, Greene FL, Edge SB, Compton CC, Gershenwald JE, Brookland non‑small cell lung cancer: a systematic review and meta‑analysis. Onco ‑ RK, Meyer L, Gress DM, Byrd DR, Winchester DP. The eighth edition AJCC target. 2016;7(48):78985. cancer staging manual: continuing to build a bridge from a population‐ 5. Ruiz‑ Cordero R, Devine WP. Targeted therapy and checkpoint immuno‑ based to a more “personalized” approach to cancer staging. CA Cancer J therapy in lung cancer. Surg Pathol Clin. 2020;13(1):17–33. Clin. 2017;67(2):93–9. 6. Zhou F, Zhou C. Lung cancer in never smokers—the East Asian experi‑ 24. Ettinger DS, Wood DE, Akerley W, Bazhenova LA, Borghaei H, Camidge ence. Transl Lung Cancer Res. 2018;7(4):450. DR, Cheney RT, Chirieac LR, D’Amico TA, Dilling TJ. NCCN guidelines 7. Kim ES, Melosky B, Park K, Yamamoto N, Yang JC. EGFR tyrosine kinase insights: non–small cell lung cancer, version 4.2016. J Natl Compr Cancer inhibitors for EGFR mutation‑positive non‑small‑ cell lung cancer: out‑ Netw. 2016;14(3):255–64. comes in Asian populations. Future Oncol. 2021;17(18):2395–408. 25. Dhruv B, Mittal N, Modi M. Study of Haralick’s and GLCM texture analysis 8. Yang JC‑H, Wu Y ‑L, Schuler M, Sebastian M, Popat S, Yamamoto N, Zhou C, on 3D medical images. Int J Neurosci. 2019;129(4):350–62. Hu C‑P, O’Byrne K, Feng J. Afatinib versus cisplatin‑based chemotherapy 26. García‑ Olalla Ó, Fernández‑Robles L, Alegre E, Castejón‑Limas M, Fidalgo for EGFR mutation‑positive lung adenocarcinoma (LUX ‑Lung 3 and LUX ‑ E. Boosting texture‑based classification by describing statistical informa‑ Lung 6): analysis of overall survival data from two randomised, phase 3 tion of gray‑levels differences. Sensors. 2019;19(5):1048. trials. Lancet Oncol. 2015;16(2):141–51. 27. Lu C‑F, Hsu F‑ T, Hsieh KL‑ C, Kao Y‑ CJ, Cheng S‑ J, Hsu JB‑K, Tsai PH, Chen 9. Rosell R, Carcereny E, Gervais R, Vergnenegre A, Massuti B, Felip E, R‑ J, Huang C‑ C, Yen Y. Machine learning–based radiomics for molecular Palmero R, Garcia‑ Gomez R, Pallares C, Sanchez JM. Erlotinib versus subtyping of gliomas. Clin Cancer Res. 2018;24(18):4429–36. standard chemotherapy as first ‑line treatment for European patients with 28 Mao KZ. Orthogonal forward selection and backward elimination algo‑ advanced EGFR mutation‑positive non‑small‑ cell lung cancer (EURTAC): rithms for feature subset selection. IEEE Trans Syst Man Cybern B (Cybern). a multicentre, open‑label, randomised phase 3 trial. Lancet Oncol. 2004;34(1):629–34. 2012;13(3):239–46. 29. Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: 10. Inoue A, Kobayashi K, Maemondo M, Sugawara S, Oizumi S, Isobe H, personalized treatment recommender system using a Cox proportional Gemma A, Harada M, Yoshizawa H, Kinoshita I. Updated overall survival hazards deep neural network. BMC Med Res Methodol. 2018;18(1):1–12. results from a randomized phase III trial comparing gefitinib with 30. Bergstra J, Bengio Y. Random search for hyper‑parameter optimization. J carboplatin–paclitaxel for chemo‑naïve non‑small cell lung cancer with Mach Learn Res. 2012;13(2):281–305. sensitive EGFR gene mutations (NEJ002). Ann Oncol. 2013;24(1):54–9. 31. Dixon PM. Bootstrap resampling. In: Encyclopedia of environmetrics. 2006. 11. Apicella M, Giannoni E, Fiore S, Ferrari KJ, Fernández‑Pérez D, Isella C, 32. Weibull W. A statistical distribution function of wide applicability. J Appl Granchi C, Minutolo F, Sottile A, Comoglio PM. Increased lactate secretion Mech. 1951;18:290–3. by cancer cells sustains non‑ cell‑autonomous adaptive resistance to MET 33. Cataldo VD, Gibbons DL, Pérez‑Soler R, Quintás‑ Cardama A. Treatment and EGFR targeted therapies. Cell Metab. 2018;28(6):848‑865.e846. of non–small‑ cell lung cancer with erlotinib or gefitinib. N Engl J Med. 12. Wu Y‑L, Zhou C, Liam C‑K, Wu G, Liu X, Zhong Z, Lu S, Cheng Y, Han B, 2011;364(10):947–55. Chen L. First‑line erlotinib versus gemcitabine/cisplatin in patients with 34. Bersanelli M, Tiseo M, Artioli F, Lucchi L, Ardizzoni A. Gefitinib and afatinib advanced EGFR mutation‑positive non‑small‑ cell lung cancer: analyses treatment in an advanced non‑small cell lung cancer (NSCLC) patient from the phase III, randomized, open‑label, ENSURE study. Ann Oncol. undergoing hemodialysis. Anticancer Res. 2014;34(6):3185–8. 2015;26(9):1883–9. 35. Lee CK, Brown C, Gralla RJ, Hirsh V, Thongprasert S, Tsai C‑M, Tan EH. Ho 13. Zhao Y, Wang H, He C. Drug resistance of targeted therapy for JC‑M, Chu DT, Zaatar A: Impact of EGFR inhibitor in non–small cell lung advanced non‑small cell lung cancer harbored EGFR mutation: From cancer on progression‑free and overall survival: a meta‑analysis. J Natl mechanism analysis to clinical strategy. J Cancer Res Clin Oncol. Cancer Inst. 2013;105(9):595–605. 2021;147(12):3653–64. 36. Yu HA, Arcila ME, Rekhtman N, Sima CS, Zakowski MF, Pao W, Kris MG, 14. Garg A, Batra U, Choudhary P, Jain D, Khurana S, Malik PS, Muthu V, Prasad Miller VA, Ladanyi M, Riely GJ. Analysis of Tumor Specimens at the Time K, Singh N, Suri T. Clinical predictors of response to EGFR‑tyrosine kinase of Acquired Resistance to EGFR‑ TKI Therapy in 155 Patients with EGFR‑ inhibitors in EGFR‑mutated non‑small cell lung cancer: a real‑ world multi‑ Mutant Lung CancersMechanisms of Acquired Resistance to EGFR‑ TKI centric cohort analysis from India. Curr Probl Cancer. 2020;44(3):100570. Therapy. Clin Cancer Res. 2013;19(8):2240–7. 15. Buonerba C, Iaccarino S, Dolce P, Pagliuca M, Izzo M, Scafuri L, Costabile F, 37. Li H, Zhang R, Wang S, Fang M, Zhu Y, Hu Z, Dong D, Shi J, Tian J. CT‑ Riccio V, Ribera D, Mucci B. Predictors of outcomes in patients with EGFR‑ based radiomic signature as a prognostic factor in stage IV ALK‑positive mutated non‑small cell lung cancer receiving EGFR tyrosine kinase inhibi‑ non‑small‑ cell lung cancer treated with TKI crizotinib: a proof‑ of‑ concept tors: a systematic review and meta‑analysis. Cancers. 2019;11(9):1259. study. Front Oncol. 2020;10:57. 16. Lambin P, Leijenaar RT, Deist TM, Peerlings J, De Jong EE, Van Timmeren 38. Song J, Shi J, Dong D, Fang M, Zhong W, Wang K, Wu N, Huang Y, Liu Z, J, Sanduleanu S, Larue RT, Even AJ, Jochems A. Radiomics: the bridge Cheng Y. A New Approach to Predict Progressionfr ‑ ee Survival in Stage IV between medical imaging and personalized medicine. Nat Rev Clin EGFRmutant NSCL ‑ C Patients with EGFR‑ TKI TherapyPrediction of EGFR‑ TKI Oncol. 2017;14(12):749–62. Treatment Outcome in Stage IV NSCLC. Clin Cancer Res. 2018;24(15):3583–92. 17. Scrivener M, de Jong EE, van Timmeren JE, Pieters T, Ghaye B, Geets 39. Jin R, Peng L, Shou J, Wang J, Jin Y, Liang F, Zhao J, Wu M, Li Q, Zhang X. Radiomics applied to lung cancer: a review. Transl Cancer Res. B. EGFR‑mutated squamous cell lung cancer and its association with 2016;5(4):398–409. outcomes. Front Oncol. 2021;11:2262. 18. Liao C‑ Y, Lee C‑ C, Yang H‑ C, Chen C‑ J, Chung W‑ Y, Wu H‑M, Guo W ‑ Y, Liu 40. Masters GA, Temin S, Azzoli CG, Giaccone G, Baker S Jr, Brahmer JR, Ellis R‑S, Lu C‑F. Enhancement of Radiosurgical Treatment Outcome Prediction PM, Gajra A, Rackear N, Schiller JH. Systemic therapy for stage IV non– Using MRI Radiomics in Patients with Non‑Small Cell Lung Cancer Brain small‑ cell lung cancer: American Society of Clinical Oncology clinical Metastases. Cancers. 2021;13(16):4030. practice guideline update. J Clin Oncol. 2015;33(30):3488. Lu et al. Cancer Imaging (2023) 23:9 Page 12 of 12 41. Detterbeck FC, Boffa DJ, Kim AW, Tanoue LT. The eighth edition lung cancer stage classification. Chest. 2017;151(1):193–203. 42. Watanabe T, Kinoshita T, Itoh K, Yoshimura K, Ogura M, Kagami Y, Yamagu‑ chi M, Kurosawa M, Tsukasaki K, Kasai M. Pretreatment total serum protein is a significant prognostic factor for the outcome of patients with periph‑ eral T/natural killer‑ cell lymphomas. Leuk Lymphoma. 2010;51(5):813–21. 43. Li K‑j, Gu W ‑ y, Xia X‑f. Zhang P, Zou C‑l, Fei Z ‑h: High Mean corpuscular volume as a predictor of poor overall survival in patients with esophageal cancer receiving concurrent chemoradiotherapy. Cancer Manag Res. 2020;12:7467. 44. Kim Y‑I. Will mandatory folic acid fortification prevent or promote cancer? Am J Clin Nutr. 2004;80(5):1123–8. 45. Ren JH, He WS, Yan GL, Jin M, Yang KY, Wu G. EGFR mutations in non‑ small‑ cell lung cancer among smokers and non‑smokers: A meta‑analy‑ sis. Environ Mol Mutagen. 2012;53(1):78–82. 46. Apostolova I, Rogasch J, Buchert R, Wertzel H, Achenbach HJ, Schreiber J, Riedel S, Furth C, Lougovski A, Schramm G. Quantitative assessment of the asphericity of pretherapeutic FDG uptake as an independent predic‑ tor of outcome in NSCLC. BMC Cancer. 2014;14(1):1–10. 47. van Cruijsen H, Giaccone G, Hoekman K. Epidermal growth factor recep‑ tor and angiogenesis: Opportunities for combined anticancer strategies. Int J Cancer. 2005;117(6):883–8. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ lished maps and institutional affiliations. Re Read ady y to to submit y submit your our re researc search h ? Choose BMC and benefit fr ? Choose BMC and benefit from om: : fast, convenient online submission thorough peer review by experienced researchers in your field rapid publication on acceptance support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations maximum visibility for your research: over 100M website views per year At BMC, research is always in progress. Learn more biomedcentral.com/submissions

Journal

Cancer Imaging – Springer Journals

Published: Jan 20, 2023

Keywords: Computer tomography imaging; EGFR TKI; Deep learning; Radiomics; Prognostic

There are no references for this article.