Access the full text.
Sign up today, get DeepDyve free for 14 days.
J. Finsterbusch (2012)Improving the performance of diffusion‐weighted inner field‐of‐view echo‐planar imaging based on 2D‐Selective radiofrequency excitations by tilting the excitation plane
Journal of Magnetic Resonance Imaging, 35
A. Rosenkrantz, H. Chandarana, J. Pfeuffer, Michael Triolo, M. Shaikh, D. Mossa, C. Geppert (2014)Zoomed echo-planar imaging using parallel transmission: impact on image quality of diffusion-weighted imaging of the prostate at 3T
Abdominal Imaging, 40
D. Bonekamp, Patrick Schelb, M. Wiesenfarth, T. Kuder, Fenja Deister, A. Stenzinger, J. Nyarangi-Dix, M. Röthke, M. Hohenfellner, H. Schlemmer, J. Radtke (2018)Histopathological to multiparametric MRI spatial mapping of extended systematic sextant and MR/TRUS-fusion-targeted biopsy of the prostate
European Radiology, 29
Hamed Kordbacheh, R. Seethamraju, E. Weiland, B. Kiefer, M. Nickel, Thitinan Chulroek, M. Cecconi, V. Baliyan, M. Harisinghani (2019)Image quality and diagnostic accuracy of complex-averaged high b value images in diffusion-weighted MRI of prostate cancer
Abdominal Radiology, 44
DJ Winkel (2021)605
Multicase Study Invest Radiol, 56
X Yu (2020)1355
Mengting Xu, T. Zhang, Zhongnian Li, Mingxia Liu, Daoqiang Zhang (2021)Towards Evaluating the Robustness of Deep Diagnostic Models by Adversarial Attack
Medical image analysis, 69
Hokuto Hirano, Akinori Minagi, Kazuhiro Takemoto (2020)Universal adversarial attacks on deep neural networks for medical image classification
BMC Medical Imaging, 21
J. Allyn, N. Allou, Charles Vidal, Amélie Renou, C. Ferdynus (2020)Adversarial attack on deep learning-based dermatoscopic image recognition systems
P. Vidnerová, Roman Neruda (2020)Vulnerability of classifiers to evolutionary generated adversarial examples
Neural networks : the official journal of the International Neural Network Society, 127
W. Bulten, H. Pinckaers, H. Boven, R. Vink, Thomas Bel, B. Ginneken, J. Laak, C. Kaa, G. Litjens (2019)Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study.
The Lancet. Oncology
D Yang (2017)507
L. Hu, D. Zhou, C. Fu, T. Benkert, C. Jiang, Rui Li, L. Wei, J. Zhao (2020)Advanced zoomed diffusion-weighted imaging vs. full-field-of-view diffusion-weighted imaging in prostate cancer detection: a radiomic features study
European Radiology, 31
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations
H. Vargas, O. Akin, T. Franiel, Y. Mazaheri, Junting Zheng, C. Moskowitz, Kazuma Udo, J. Eastham, H. Hricak (2011)Diffusion-weighted endorectal MR imaging at 3 T for prostate cancer: tumor detection and assessment of aggressiveness.
Radiology, 259 3
Naveed Akhtar, A. Mian (2018)Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey
IEEE Access, 6
T. Tamada, V. Prabhu, Jianhong Li, J. Babb, S. Taneja, A. Rosenkrantz (2017)Assessment of prostate cancer aggressiveness using apparent diffusion coefficient values: impact of patient race and age
Abdominal Radiology, 42
Sebastian Feuerlein, Matthew Davenport, A. Krishnaraj, Elmar Merkle, Rajan Gupta (2015)Computed high b-value diffusion-weighted imaging improves lesion contrast and conspicuity in prostate cancer
Prostate Cancer and Prostatic Disease, 18
G. Nketiah, K. Selnæs, Elise Sandsmark, J. Teruel, Brage Krüger-Stokke, H. Bertilsson, T. Bathen, M. Elschot (2018)Geometric distortion correction in prostate diffusion‐weighted MRI and its effect on quantitative apparent diffusion coefficient analysis
Magnetic Resonance in Medicine, 79
C. Brendle, P. Martirosian, N. Schwenzer, S. Kaufmann, S. Kruck, U. Kramer, M. Notohamiprodjo, K. Nikolaou, C. Schraml (2016)Diffusion-weighted imaging in the assessment of prostate cancer: Comparison of zoomed imaging and conventional technique.
European journal of radiology, 85 5
O. Donati, Y. Mazaheri, A. Afaq, H. Vargas, Junting Zheng, C. Moskowitz, H. Hricak, O. Akin (2014)Prostate cancer aggressiveness: assessment with whole-lesion histogram analysis of the apparent diffusion coefficient.
Radiology, 271 1
Babak Bejnordi, M. Veta, Paul Diest, B. Ginneken, N. Karssemeijer, G. Litjens, J. Laak, M. Hermsen, Quirine Manson, M. Balkenhol, Oscar Geessink, N. Stathonikos, M. Dijk, P. Bult, F. Beca, Andrew Beck, Dayong Wang, A. Khosla, Rishab Gargeya, H. Irshad, Aoxiao Zhong, Q. Dou, Quanzheng Li, Hao Chen, Huangjing Lin, P. Heng, C. Hass, Elia Bruni, Q. Wong, U. Halici, Mustafa Öner, R. Cetin-Atalay, Matt Berseth, V. Khvatkov, A. Vylegzhanin, Oren Kraus, M. Shaban, N. Rajpoot, Ruqayya Awan, K. Sirinukunwattana, Talha Qaiser, Y. Tsang, David Tellez, Jonas Annuscheit, P. Hufnagl, Mira Valkonen, K. Kartasalo, Leena Latonen, P. Ruusuvuori, Kaisa Liimatainen, Shadi Albarqouni, Bharti Mungal, A. George, S. Demirci, N. Navab, Seiryo Watanabe, S. Seno, Y. Takenaka, H. Matsuda, Hady Phoulady, V. Kovalev, A. Kalinovsky, V. Liauchuk, G. Bueno, M. Fernández-Carrobles, I. Serrano, O. Deniz, Daniel Racoceanu, Rui Venâncio (2017)Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer
B. Nandram, Thelge Peiris (2018)Bayesian analysis of a ROC curve for categorical data using a skew-binormal model
Statistics and Its Interface, 11
Fei Yang, N. Dogan, R. Stoyanova, J. Ford (2018)Evaluation of radiomic texture feature error due to MRI acquisition and reconstruction: A simulation study utilizing ground truth.
Physica medica : PM : an international journal devoted to the applications of physics to medicine and biology : official journal of the Italian Association of Biomedical Physics, 50
B. Turkbey, A. Rosenkrantz, M. Haider, A. Padhani, G. Villeirs, K. Macura, C. Tempany, P. Choyke, F. Cornud, D. Margolis, H. Thoeny, S. Verma, J. Barentsz, J. Weinreb (2019)Prostate Imaging Reporting and Data System Version 2.1: 2019 Update of Prostate Imaging Reporting and Data System Version 2.
Dong Yang, Daguang Xu, S. Zhou, B. Georgescu, Mingqing Chen, Sasa Grbic, Dimitris Metaxas, D. Comaniciu (2017)Automatic Liver Segmentation Using an Adversarial Image-to-Image Network
T. Kuru, K. Wadhwa, Richard Chang, L. Echeverria, M. Roethke, A. Polson, G. Rottenberg, B. Koo, Edward Lawrence, Jonas Seidenader, V. Gnanapragasam, R. Axell, W. Roth, A. Warren, A. Doble, G. Muir, R. Popert, H. Schlemmer, B. Hadaschik, C. Kastner (2013)Definitions of terms, processes and a minimum dataset for transperineal prostate biopsies: a standardization approach of the Ginsburg Study Group for Enhanced Prostate Diagnostics
BJU International, 112
Ruiming Cao, Xinran Zhong, Sohrab Afshari, E. Felker, V. Suvannarerg, Teeravut Tubtawee, S. Vangala, F. Scalzo, S. Raman, K. Sung (2021)Performance of Deep Learning and Genitourinary Radiologists in Detection of Prostate Cancer Using 3‐T Multiparametric Magnetic Resonance Imaging
Journal of Magnetic Resonance Imaging, 54
(2023)Springer International Publishing; 2017
Amogh Hiremath, R. Shiradkar, H. Merisaari, P. Prasanna, O. Ettala, P. Taimen, H. Aronen, P. Boström, I. Jambor, A. Madabhushi (2020)Test-retest repeatability of a deep learning architecture in detecting and segmenting clinically significant prostate cancer on apparent diffusion coefficient (ADC) maps
European Radiology, 31
F. Giganti, A. Rosenkrantz, G. Villeirs, V. Panebianco, A. Stabile, M. Emberton, C. Moore (2019)The Evolution of MRI of the Prostate: The Past, the Present, and the Future.
AJR. American journal of roentgenology
M. Klingebiel, T. Ullrich, M. Quentin, D. Bonekamp, J. Aissa, D. Mally, C. Arsov, P. Albers, G. Antoch, L. Schimmöller (2020)Advanced diffusion weighted imaging of the prostate: Comparison of readout-segmented multi-shot, parallel-transmit and single-shot echo-planar imaging.
European journal of radiology, 130
Xin Yu, B. Lou, Bibo Shi, D. Winkel, Nacim Arrahmane, M. Diallo, T. Meng, H. Busch, R. Grimm, B. Kiefer, D. Comaniciu, A. Kamen, H. Huisman, A. Rosenkrantz, T. Penzkofer, I. Shabunin, M. Choi, Qingsong Yang, D. Szolar (2020)False Positive Reduction Using Multiscale Contextual Features for Prostate Cancer Detection in Multi-Parametric MRI Scans
2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI)
A. Rosenkrantz, H. Chandarana, N. Hindman, F. Deng, J. Babb, S. Taneja, C. Geppert (2013)Computed diffusion-weighted imaging of the prostate at 3 T: impact on image quality and tumour detection
European Radiology, 23
H. Togt (2003)Publisher's Note
J. Netw. Comput. Appl., 26
L. Hu, Liming Wei, Shuhao Wang, C. Fu, T. Benker, J. Zhao (2021)Better lesion conspicuity translates into improved prostate cancer detection: comparison of non-parallel-transmission-zoomed-DWI with conventional-DWI
Abdominal Radiology, 46
Ruiming Cao, Amirhossein Bajgiran, Sohrab Mirak, Sepideh Shakeri, Xinran Zhong, D. Enzmann, S. Raman, K. Sung (2019)Joint Prostate Cancer Detection and Gleason Score Prediction in mp-MRI via FocalNet
IEEE Transactions on Medical Imaging, 38
M. Jolly, Christoph Gütter, J. Guehring (2010)Cardiac segmentation in MR cine data using inverse consistent deformable registration
2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro
T. Hambrock, D. Somford, H. Huisman, I. Oort, J. Witjes, C. Kaa, T. Scheenen, J. Barentsz (2011)Relationship between apparent diffusion coefficients at 3.0-T MR imaging and Gleason grade in peripheral zone prostate cancer.
Radiology, 259 2
Background Deep‑learning‑based computer ‑aided diagnosis (DL ‑ CAD) systems using MRI for prostate cancer (PCa) detection have demonstrated good performance. Nevertheless, DL‑ CAD systems are vulnerable to high heterogenei‑ ties in DWI, which can interfere with DL‑ CAD assessments and impair performance. This study aims to compare PCa detection of DL‑ CAD between zoomed‑field‑ of‑ view echo‑planar DWI (z‑DWI) and full‑field‑ of‑ view DWI (f‑DWI) and find the risk factors affecting DL ‑ CAD diagnostic efficiency. Methods This retrospective study enrolled 354 consecutive participants who underwent MRI including T2WI, f‑DWI, and z‑DWI because of clinically suspected PCa. A DL ‑ CAD was used to compare the performance of f‑DWI and z‑DWI both on a patient level and lesion level. We used the area under the curve (AUC) of receiver operating characteristics analysis and alternative free‑response receiver operating characteristics analysis to compare the performances of DL ‑ CAD using f‑ DWI and z‑DWI. The risk factors affecting the DL ‑ CAD were analyzed using logistic regression analyses. P values less than 0.05 were considered statistically significant. Results DL‑ CAD with z‑DWI had a significantly better overall accuracy than that with f‑DWI both on patient level and lesion level (AUC : 0.89 vs. 0.86; AUC : 0.86 vs. 0.76; P < .001). The contrast‑to ‑noise ratio (CNR) of lesions in DWI patient lesion was an independent risk factor of false positives (odds ratio [OR] = 1.12; P < .001). Rectal susceptibility artifacts, lesion diameter, and apparent diffusion coefficients (ADC) were independent risk factors of both false positives (OR rectal = 5.46; OR = 1.12; OR = 0.998; all P < .001) and false negatives ( OR = 3.31; susceptibility artifact diameter, ADC rectal susceptibility artifact OR = 0.82; OR = 1.007; all P ≤ .03) of DL‑ CAD. diameter ADC Conclusions Z‑DWI has potential to improve the detection performance of a prostate MRI based DL ‑ CAD. Trial registration ChiCTR, NO. ChiCT R2100 041834. Registered 7 January 2021. *Correspondence: Jun‑gong Zhao firstname.lastname@example.org 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. Hu et al. Cancer Imaging (2023) 23:6 Page 2 of 12 Keywords Diffusion magnetic resonance imaging, Deep learning, Prostatic neoplasms, Artifact, Risk factor Background subjects signed the informed consent before they under- Diffusion-weighted imaging (DWI) is an indispensable went the MRI examination and allowed us to use their technique in prostate magnetic resonance imaging (MRI), data for a series of follow-up studies about AI system providing both qualitative and quantitative functional building for PCa diagnosis. All procedures performed in information of prostate tissue and lesions [1–5]. Rely- studies involving human participants were according to ing on the Prostate Imaging Reporting and Data System the 1964 Helsinki Declaration and its later amendments. (PI-RADS) , DWI combined with T2-weighted imag- ing (T2WI) has shown significantly improved accuracy of Participant selection detection and characterization of prostate cancer (PCa) Between January 2021 and January 2022, participants lesions  and plays an important role in the clinical man- with clinically suspected PCa undergone MRI examina- agement strategy of patients with suspected PCa. However, tions and subsequent MRI fusion ultrasound-guided due to differences in hardware, software, and technical targeted biopsies (2–4 cores) of MRI-suspicious lesions experience [7, 8], there is a high variation of diagnostic (PI-RADS score ≥ 3) followed by systematic biopsies accuracy and inter-observer agreement in the interpreta- (10–12 cores) were consequently enrolled from Shanghai tion of prostate MRI across medical centers [1, 2]. These Sixth People’s Hospital Affiliated to Shanghai Jiao Tong factors limit the clinic application of prostate MRI. University School of Medicine. All MRI scans were inter- Various deep-learning-based computer-aided diagnosis preted by two senior radiologists with more than 15 years (DL-CAD) systems using prostate MRI for PCa detection of experience in prostate MRI interpretation using the have demonstrated comparable or improved performance PI-RADS version 2.1. and reproducibility with less time and labor compared to The inclusion criteria included the following: (a) pros - experienced radiologists [7, 9–12]. Nevertheless, DL-CAD tate lesions with definite boundaries on three types of systems are vulnerable to high heterogeneities in DWI MR images according to according to PI-RADS version , which can interfere with DL-CAD assessments and 2.1 (T2WI, DWI, and ADC); (b) complete clinical infor- impair performance. Developing a DWI sequence that can mation and entire MRI reports including the number, improve the accuracy and reliability of DL-CAD would PI-RADS score, and location of suspected PCa lesion; improve PCa diagnosis. DWI acquisition performed using (c) complete biopsy records and results, including the full-field-of-view (FOV) ssEPI-DWI (f-DWI) is prone to number, location, Gleason score (GS) of lesions. Exclu- distortions, susceptibility artifacts, and limited spatial res- sion criteria were (a) a prior history of PCa treatment; olution. By contrast, zoomed-field-of-view echo-planar (b) biopsy within 6 months prior to the MRI examination; DWI (z-DWI) using a small FOV that only covers a specific (c) an interval of more than 2 weeks between MRI and region-of-interest (ROI) results in fewer geometric distor- the biopsy procedure; (d) unavailability of the final PCa tions, susceptibility artifacts, and higher spatial resolution diagnosis. [2, 10, 14–18]. Previous studies indicated that variation of noise, deformation and changes of resolution are impor- MRI examination tant factors interfering with the judgement of DL-CAD [13, Patients were advised to empty their bowel prior to the 19–22], therefore, we hypothesized that z-DWI might be examination. All patients underwent both T2-weighted helpful to improve the performance of DL-CAD for PCa imaging (T2WI), f-DWI, z-DWI with b-values of 50, diagnosis. 1000, and 1500 s/mm on a 3 T MRI scanner (MAG- In this study, we assessed the use of DL-CAD with f-DWI NETOM Skyra, Siemens Healthcare, Erlangen, Ger- and with z-DWI and compared each format in diagnos- many), and a phased-array 18-channel body coil in ing MRI-visible PCa. The risk factors of patient condition, combination with an integrated 32-channel spine coil was image quality, and lesion characteristics that could affect used for signal reception. Z-DWI was performed with a the DL-CAD diagnostic efficiency were analyzed. slight rotation of the field-of-excitation , motion reg - istration , and complex averaging . More detailed Methods parameters of each MRI sequence are shown in Table 1. This retrospective study was approved by the local eth - ics committee at our institution [Approve No: 2022-KY- Histopathology matching and annotation 073(K)]. As part of a prospective study aiming to build The ground truth of this study was lesion confirmation on a robust AI system for PCa diagnosis, all the enrolled histopathology after biopsy. At least one GU radiologist Hu et al. Cancer Imaging (2023) 23:6 Page 3 of 12 Table 1 MR sequence parameters Parameter T2‑ weighted imaging f‑DWI z‑DWI Sequence TSE SS‑EPI zoomed SS‑EPI Field of view (mm ) 200 × 200 380 × 280 190 × 109 Scan matrix 320 × 256 178 × 132 100 × 56 Voxel size (mm ) 0.6 × 0.6 × 3.5 2.1 × 2.1 × 3.0 1.9 × 1.9 × 3.0 PE Dir. Left‑ > Right Anterior‑ > Posterior Left‑ > Right Echo time (ms) 101 79 76 Repetition time (ms) 6000 4200 3800 Bandwidth (Hz/px) 200 1872 1612 b‑ value (s/mm ) NA 50, 1000, 1500 NEX 2 1, 3, 6 Fat suppression NA FatSat Acceleration factor 2 (GRAPPA) 2 (GRAPPA) NA Acquisition time (min) 2:08 3:05 2:37 PE Dir. Phase Encoding direction, GRAPPA Generalized auto-calibrating partial parallel acquisition, NA not applicable and one GU pathologist retrospective reviewed MRI and [9, 12, 27]. In brief, the system computes apparent dif- histopathology examinations together at a multidiscipli- fusion coefficient (ADC) maps and calculates b-value nary meeting scheduled monthly. Each lesion in MRI was images at b = 2000 s/mm using the input DWI and matched to the corresponding location on the specimen then segments whole-gland volumes using T2WI. through visual co-registration. According to the Ginsburg After that, the DWI and ADC are aligned to the T2WI. Study group method, a sextant scheme of the prostate Finally, the system identifies the clinically relevant has been used for analysis of the correct identification of lesions based on T2WI, calculated b-value images, the lesions’ localization . Prostate contours were seg- ADC maps, and prostate segmentations, and provides mented on T2WI images and partitioned into sextants detected lesion localization, a PI-RADS category, and using the midsagittal plane and four additional angulated case-based level of suspicion (LoS). The LoS represents planes according to the biopsy protocol. Sextant-specific the confidence of the software that a lesion with a PI- systematic biopsy histopathology was assigned to all MR RADS score of 3 or above is present in that patient, and sextants and augmented by calculating the maximum ranges from 3.0 to 5.0 in steps of 0.1. GS between systematic histopathology and histopathol- ogy from targeted fusion biopsy to sextants intersecting Evaluation of detection performance with MR lesions to create a sextant map of histopathol- The patient-based diagnostic performance of DL-CAD ogy ground truth . MRI reports and biopsy results, was evaluated by computing ROC curves for each case- including the number, GS, and prostate location by MRI based LoS. We evaluated two clinically relevant tasks: (a) of lesions of all selected participants, were recorded in differentiating between benign lesion and PCa (Gleason the regular retrospective review. Grade Group (GGG) ≥ 1), and (b) differentiating between benign lesion or low-grade cancer of GS 3 + 3 and clini- Deep learning‑based computer‑aided diagnosis cally significant PCa (csPCa) (GGG ≥ 2) . A prototype DL-CAD system (MR Prostate AI v1.2.5; The DL-CAD system can automatically detect clinically October 2020; Siemens Healthcare GmbH, Erlangen, relevant PCa lesions which were defined on pathology/ Germany) was used to test the performances of z-DWI histology as Gleason score ≥ 7, and/or volume ≥ 0.5 cc, and f-DWI in PCa detection. This DL-CAD system was and/or extra-prostatic extension according to PI-RADS trained using 2170 bp-MRI prostate examinations from v2.1. The clinically relevant PCa lesions-based detection 7 institutions, consisting of 944 lesion-free cases and performances of DL-CAD were evaluated using free- 1226 cases with at least 1 clinically significant lesion response receiver operating characteristics (FROC) anal- that is deemed at least equivocal as designated by PI- ysis due to PCa’s multi-focality [6, 28, 29]. In addition, RADS score 3 and higher; none of the cases included considering the FROC curve has an infinite area under in the testing data set was part of the training . The the curve (AUC), we also used alternative free-response architecture and processing steps of the DL-CAD sys- receiver operating characteristics (AFROC) analysis with tem have been described in detail in previous studies Hu et al. Cancer Imaging (2023) 23:6 Page 4 of 12 a finite AUC ranging from 0 to 1 to evaluate the lesion- deviation of the noise (S D ) in a noise-only area, a noise based detection performance of the DL-CAD . third ROI was placed in the center of the bladder with the DWI at b = 1500 s/mm . The ADC values, mean signal DWI image analysis intensities in the lesion ROI (S ), the reference ROI lesion Two radiologists with approximately 3 years of experi- (S ), and SD were recorded for further analyses. normal noise ence in prostate MRI reporting and blinded to all clinical The difference in the noise between z-DWI and details and biopsy results twice evaluated the DWI sets f-DWI was calculated using the estimated signal-to- including the image quality scoring, DWI signal intensi- noise ratio (eSNR) : ties measurements and ADC measurements. The inde - eSNR = S /SD lesion noise pendent readings of the first time were used for assessing inter-reader agreement on quality scores, and DWI sig- Lesion conspicuity was determined by the contrast- nal intensities and ADC measurements. The final image to-noise ratio (CNR): quality score of each DWI set was determined by the two CNR = (S − S )/SD noise lesion normal radiologists by consensus. The final image quality score and the mean results of DWI signal intensities measure- We compared DWI quality and the characteristics ments and ADC measurements of the two radiologists of benign lesions and malignant lesions to determine were used for the evaluation of risk factors affecting the risk factors affecting the DL-CAD diagnostic efficiency. DL-CAD diagnostic efficiency. The relationships between DL-CAD diagnostic perfor - After four weeks, the second time image analysis was mance and image quality as well as lesion characteris- performed by the two readers in a different order to test tics were also evaluated. for reproducibility. Image evaluation was performed using the Image J Software (National Institutes of Health, Bethesda, MD, USA). Statistical analyses For each time, the DWI sets in two different sessions The one-sample Kolmogorov-Smirnov test was used to at intervals of at least two weeks to minimize recognition check the assumption of a normal distribution of the bias. In each session, only 1 of the 2 DWI sets for each data. The independent t test or paired t test was used for patient was evaluated. Specifically, the DWI sets were normally distributed data. The Mann-Whitney U test reviewed in a random order and were rated in terms of was used to assess non-normally distributed continuous overall quality and anatomic distortion using a 5-point variables. Categorical variables were reported as percent- Likert scale with 5 indicating the highest quality [18, 31]. ages and compared by χ test. Comparisons of sensitivity Axial T2WI images were used as a reference for guid- and specificity were performed using the McNemar test. ing anatomical localization of findings on the DWI sets Comparisons of AUCs were performed using the Delong . In addition, the presence of artifacts including rec- test. Inter- and intra-observer agreement of overall qual- tal susceptibility artifacts, phase wrap-around, artifacts ity, anatomic distortion, and artifact evaluation were from artificial joint replacements, other artifacts from tested with a weighted κ coefficient. ADC values and outside the body (only for f-DWI) was noted, and the SI and S I measurements were assessed with the lesion normal grade of artifact influence on image quality was scored as: intraclass correlation coefficient. 1, excellent image quality; 2, mild artifact, not impacting Univariable logistic regression analyses and multivaria- diagnosis; 3, moderate artifact, mildly impacting diagno- ble logistic regression analyses with stepwise approaches sis; 4, pronounced artifact, moderately impacting diag- were applied to assess the relationship between false pos- nosis; 5, pronounced artifact, non-diagnostic. Artifacts itives and potential risk factors and between false nega- scored ≥3 were considered to have an influence on the tives and potential risk factors. The multicollinearity of diagnosis . variables in the multivariable analysis was determined To evaluate noise, lesion conspicuity, and ADC values using a variance inflation factor (VIF) of greater than 10. of each DWI set, the radiologists were also asked to draw Statistical evaluations were performed using R v4.10 ROIs on the ADC map, which were then copied to the (R Foundation for Statistical Computing, Vienna, Aus- DWI (b = 1500 s/mm ) image. One ROI was placed in the tria; https:// www.R- proje ct. org/). The VIFs were cal - center of the lesion in the slice with the largest extent of culated using the “car” package. Comparison of Binary the lesion. The second ROI was placed in the correspond - Diagnostic Tests in a Paired Study Design was performed ing contralateral normal tissue as a reference ROI. If the using “DTComPair” package. The ROC curves were plot - contralateral tissue was also abnormal, the reference ted using the “pROC” package. The FROC curves and ROI was placed in healthy appearing tissue of the same AFROC curves were plotted using the “BayesianFROC” anatomical zone as the lesion. To calculate the standard Hu et al. Cancer Imaging (2023) 23:6 Page 5 of 12 package. Forest plots of the logistic regression analyses Table 2 Demographic and Clinical Characteristics of Included Patients were performed using “forestmodel” package. P values less than 0.05 were considered statistically significant. Variable Patients without Patients cancer (n = 180) with cancer (n = 174) Results Participant and lesion baseline characteristics a Median age (year) 69 (65 ‑ 75) 72 (66 ‑ 79) Initially, a total of 389 participants were enrolled. Of a Total PSA (ng/ml) 8.9 (5.1 ‑ 13.0) 14.0 (6.7 ‑ 29.6) these, 35 were excluded according to the inclusion and a Free PSA (ng/ml) 1.3 (0.8 ‑ 2.3) 1.6 (0.8 ‑ 3.5) exclusion criteria. The detailed reasons for exclusion a Free PSA/total PSA 0.2 (0.1 ‑ 0.2) 0.1 (0.1 ‑ 0.2) are listed in Fig. 1. A total of 354 patients (median age, b No. of patients with MRI‑ detected lesions 65 years; interquartile range [IQR], 71–77 years) with 486 1 lesion 145 (81) 107 (61) lesions (250 cancer lesions and 236 benign lesions) were 2 lesions 21 (12) 59 (34) included in the final study. Baseline epidemiologic and 3 lesions 8 (4) 7 (4) clinical characteristics of the participants are shown in More than 3 lesions 6 (3) 1 (1) Table 2. PSA prostate-specific antigen, PI-RADS Prostate Imaging Reporting and Data There were no significant differences in the mean System patient age between subjects with and without PCa Data in parentheses are the interquartile range (P = 0.23). Participants with PCa had higher levels of Data in parentheses are percentages total PSA and free PSA and lower free PSA ratio than those without PCa (All P < .001). Detailed information PCa of GS 3 + 3 and csPCa (Sensitivity: 0.81 [95% CI: 0.74 about lesions, including lesion location, pathologic find - - 0.87] vs. 0.78 [95% CI: 0.70-0.84]; Specificity: 0.85 [95% ings, and clinical assessment, is shown in Table 3. CI: 0.79-0.80] vs. 0.82 [95% CI:0.76-0.87]; AUC: 0.88 [95% CI:0.84-0.91] vs. 0.85 [95% CI: 0.81-0.88], P = 0.024). Patient‑based performance Figure 2 shows that, compared with f-DWI, DL-CAD Lesion‑based detection performance based on z-DWI had better performance in differenti - Lesion-based detection performance of the DL-CAD sys- ating between benign lesion and PCa (Sensitivity: 0.79 tem is shown in Table 4 and Fig. 3. [95% CI: 0.73-0.85] vs. 0.78 [95% CI: 0.71-0.84]; Specific - Compared with f-DWI, z-DWI had significantly higher ity: 0.89 [95% CI: 0.83-0.93] vs. 0.83[95% CI: 0.77-0.89]; sensitivity but lower specificity for lesion detection at AUC: 0.89 [95% CI:0.85-0.92] vs. 0.86 [95% CI: 0.81-0.89], PI-RADS category greater than or equal to 3 (Sensitiv- P = 0.007) and in differentiating between benign tissue or ity: 0.93 [95% CI: 0.90-0.96] vs. 0.73 [95% CI:0.68-0.79]; Fig. 1 Flowchart of participant inclusion and exclusion Hu et al. Cancer Imaging (2023) 23:6 Page 6 of 12 Table 3 Lesion characteristics Variable Benign lesions (n = 236) Malignant lesions (n = 250) Lesion location Transition zone 73 (31) 118 (47) Peripheral zone 163 (69) 132 (53) Median Diameter (mm) 7.5 (4.6–11.8) 21.3 (13.1–37.3) Clinical PI‑RADS score† PI‑RADS 2 123 (52) 23 (9) PI‑RADS 3 52 (22) 25 (10) PI‑RADS 4 47 (20) 110 (44) PI‑RADS 5 14 (6) 92 (37) Gleason Group group† Gleason grade group 1 (GS 3 + 3) NA 17 (7) Gleason grade group 2 (GS 3 + 4) NA 37 (15) Gleason grade group 3 (GS 4 + 3) NA 80 (32) Gleason grade group 4 (GS = 8) NA 33 (13) Gleason grade group 5 (GS > 8) NA 83 (33) Data in parentheses are the interquartile range Data in parentheses are percentages Fig. 2 Comparison of receiver operating characteristics curves for analysis for LoS of DL‑ CAD based on f‑DWI and z‑DWI performance on benign versus grade group 1 and above (a) and on benign and grade group 1 versus grade group 2 and above (b). LoS = patient‑based level of suspicion. DL‑ CAD = Deep Learning Based Computer‑Aided Diagnosis Specificity: 0.61 [95% CI: 0.55-0.67] vs. 0.76 [95% CI:0.70- Table 4 Prostate cancer lesion detection performance of DL‑ CAD using f‑DWI and z‑DWI 0.81]; P < .001 for all comparisons). At a detection sensitivity > 0.1, DL-CAD using PI‑RADS No. of true‑positive No. of false‑positive z-DWI provided lower False Positive Fractions per detections detections patient than DL-CAD using f-DWI (Fig. 4a). DL- f‑DWI z‑DWI f‑DWI z‑DWI (n = 91) CAD using z-DWI had better performance for PCa (n = 182) (n = 230) (n = 52) lesion detection with less false-positive detections 3 1 4 1 1 per patient than DL-CAD f-DWI (AUC, 0.855 [95% 4 47 66 22 57 CI:0.825-0.883] vs. 0.760 [95% CI: 0.714-0.799]; 5 134 160 29 33 P < .001) (Fig . 4b). Hu et al. Cancer Imaging (2023) 23:6 Page 7 of 12 Fig. 3 Alluvial diagrams showed all lesions detected and given PI‑RADS score 3 (Yellow), PI‑RADS score 4 (Red) and PI‑RADS score 5 (Blue) by DL‑ CAD using f‑DWI (a) and z‑DWI (b). The detected lesions included both true ‑positive detections and false ‑positive detections. Compared with f‑DWI, z‑DWI has both more true ‑positive detections (182 vs. 230) and more false ‑positive detections (52 vs. 91) Fig. 4 FROC analysis (a) and AFROC analysis (b) for detection sensitivity on prostate cancer lesions. The number of false positives (x‑axis) of Fig. 4a is shown in log‑scale. According to the FROC curves, for a detection sensitivity > 0.1, using z‑DWI shows a lower rate of false ‑positive detections per patient. The AFROC curves (b) show that compared with using f‑DWI, using z‑DWI can result in higher AUC in terms of PCa lesion detections per patient (0.855 [95%CI:0.825‑0.883] vs. 0.760 [95%CI:0.714‑0.799]). FROC=Free ‑response receiver operating characteristics analysis. AFROC = alternative free‑response receiver operating characteristics. FPF = false positive fraction DWI image analysis Risk factors evaluation The inter- and intra-observer agreement for image qual - Examples of DL-CAD diagnosis using f-DWI and z-DWI ity scores, DWI signal intensities, and ADC measure- are shown in Fig. 5. Based on subjective visual evaluation, ments were concordant (Suppl. material). prostate deformation, artifacts, and lesion signal intensity As shown in Table 5, DL-CAD using z-DWI has sig- on DWI and ADC values are possible reasons resulting in nificantly higher scores for overall image quality and DL-CAD misdiagnosis. distortion of the prostate, and lower scores of the As shown in Fig. 6, CNR was an independent risk factor severity of artifacts compared with DL-CAD using for false positives of DL-CAD (odds ratio [OR], 1.12; 95% f-DWI (P ≤ .035). For both benign lesions and malig- CI, 1.05-1.21; P < .001) and rectal susceptibility artifacts, nant lesions, DL-CAD using z-DWI had lower ADC diameter, and ADC were independent risk factors associ- values and higher CNR and eSNR than those in DL- ated with both false positive detections (OR rectal susceptibil- CAD f-DWI (P ≤ .011). 5.46; 95% CI, 2.77-10.96; OR , 1.12; 95% CI, ity artifact, diameter Hu et al. Cancer Imaging (2023) 23:6 Page 8 of 12 Table 5 Comparison of subjective image quality and the main 1.07-1.17; OR 0.998; 95% CI, 0.997-0.999; all P < .001) ADC, lesion between different diffusion‑ weighted imaging sequences and false negative detections (OR , rectal susceptibility artifacts of the prostate 3.31; 95% CI, 1.15-9.68; OR , 0.82; 95% CI, 0.75-0.89; diameter OR , 1.007; 95% CI, 1.004-1.009; all P ≤ .03) of DL-CAD. Feature f‑DWI z‑DWI P value ADC Subjective image quality score (n = 354) Discussion Overall image qualit y 4.0 ± 0.8 4.4 ± 0.7 <.001 Our study has two main contributions. First, we com- Distortion of prostate 4.1 ± 0.7 4.5 ± 0.7 <.001 pared PCa detection performance of the DL-CAD sys- Rectal susceptibility 1.6 ± 0.9 1.5 ± 0.8 0.035 tem with the use of z-DWI and f-DWI, finding that the artifact DL-CAD system exhibited significantly better PCa detec - Artificial joint 1.1 ± 0.6 1.0 ± 0.2 <.001 tion performance based on z-DWI than using f-DWI. It replacements indicates that z-DWI may be a way towards more con- Phase wrap‑around 1.3 ± 0.7 1.0 ± 0.1 <.001 sistent and better image quality. Second, risk factors that Other artifacts out of 1.2 ± 0.6 NA NA body affected the diagnostic performance of the DL-CAD sys - Main lesion characteristics tem in the assessment of PCa were identified. As these Benign lesion (n = 236) types of image artifacts are common in prostate MRI, eSNR 21.9 (17.5‑29.2) 30.7 (25.1‑40.4) <.001 the risk factors that interfere with one DL-CAD system CNR 1.6(1.0‑3.6) 2.1(1.0‑4.6) 0.011 may have similar effects in other DL-CAD systems with ADC 1027.6 ± 262.5 977.9 ± 272.7 <.001 different network structures and parameter settings. Malignant lesion (n = 250) Understanding these risk factors might be helpful for eSNR 27.7(22.3‑37.6) 44.2(33.1‑57.7) 0.233 standardizing prostate MRI scanning guidelines for DL- CNR 7.6 (2.5‑15.6) 8.0 (4.0‑16.7) 0.004 CAD analysis, customizing the corresponding DL-CAD training strategies, and improving the diagnostic accu- ADC 751.5 ± 205.3 698.3 ± 158.8 <.001 racy and generalization of DL-CAD. a b Data in parentheses are the interquartile range. Numbers are means ± standard deviations DWI acquired by zoomed-FOV technology has pro- vided better image quality than that obtained with other technologies [10, 15, 18, 32, 34]. However, due in part to the physiological limitations of visually identifying sub- tle differences among lesions, the improvement in image Fig. 5 Examples of DL‑ CAD diagnosis using f‑DWI (a) and z‑DWI (b). The red or orange ROIs indicate highly suspected PCa lesions detected by the system. Cases without ROIs indicate no suspected PCa lesion was detected. Case 1. A 77‑ year‑ old man with an inflammatory lesion in the central zone (CZ). DL‑ CAD using f‑DWI detected a fake cancer lesion in PZ, which was a rectal artifact. Case 2. A 62‑ year‑ old man with a PCa lesion (Gs = 3 + 4) in the right PZ, which was missed by DL‑ CAD with the use of f‑DWI possibly due to rectal distortion and artifacts. Case 3. A 71‑ year‑ old man with an inflammatory lesion in the right PZ, which was detected by DL ‑ CAD using z‑DWI. Case 4. A 73‑ year‑ old man with a PCa lesion (Gs = 4 + 3) in the transition zone ( TZ) which was missed by DL‑ CAD using z‑DWI Hu et al. Cancer Imaging (2023) 23:6 Page 9 of 12 Fig. 6 Logistic regression analyses showing variables associated with wrong detections. a and b is the univariable logistic regression analysis of benign and malignant lesions, respectively. c and d is the multivariable logistic regression analyses of benign and malignant lesions, respectively. PSA, prostate‑specific antigen; PZ, peripheral zone; TZ, transition zone; eSNR, estimated signal‑to ‑noise ratio quality did not significantly improve the subjective eval - sensitivity of DL-CAD using z-DWI for detecting lesions uation performance of radiologists for PCa detection improved by 20%. However, its specificity for detecting in many previous studies [15, 34]. Given the ability of lesions was reduced by 15%. Our results indicate that the DL-CAD to mine the sub-pixel level, previous radiomic observed operating point of DL-CAD using z-DWI was study found that a radiomics model based on z-DWI had shifted in favor of higher sensitivity, but considering the a higher diagnostic accuracy, sensitivity, and specificity superior ROC curves, z-DWI achieved superior specific - than a model based on f-DWI . Partly differing from ity at given sensitivity levels. previous result, the improvement of PCa detection per- In contrast to the radiomic model which was con- formance of the DL-CAD using z-DWI primarily comes structed with explicable texture feature information, the from the improvement in sensitivity. We found that the training and diagnosis process of DL-CAD is much more Hu et al. Cancer Imaging (2023) 23:6 Page 10 of 12 complex. We used common clinical research methods to multi-vendor datasets obtained from multiple imaging find risk factors contributing to diagnostic errors from centers are needed to verify our results. Third, because in the macroscopic level of DL-CAD. We found that CNR our clinic routine, active surveillance instead of biopsy is was positively associated with false positives of DL-CAD, typically selected for patients with elevated PSA but neg- whilst ADC was negatively associated with false positives ative MRIs, it is hard for us to evaluate the performances of DL-CAD. It indicates that parameter settings produc- of DL-CAD on PCa patients with negative MRIs. There - ing on average lower ADC values and higher CNR in PCa fore, only patients with prostate lesions with definite lesions in DWI might be helpful to improve the perfor- boundaries on all MR images were included, there might mance of DL-CAD using DWI sets. be a potential source of selection bias. Fourth, a reduced Because the lesions in z-DWI have lower ADC values FOV may prevent the visualization of lymph nodes, a but higher CNR in both benign and malignant lesions full-FOV DWI is still needed to study lymph nodes. than those in f-DWI, it is not surprising that DL-CAD Finally, in our study, targeted biopsies were used as refer- using z-DWI detected more true positives but had ence standards. Whole-mount histopathology may have increased false positives. Therefore, strategies to over - improved the accuracy of the agreement between the MR come these problems will need to be determined before images and the histopathology. applying z-DWI to existing f-DWI based DL-CAD sys- tems, e.g., by further training the DL-CAD with z-DWI Conclusions data in the future. In conclusion, z-DWI has the potential to improve the Consistent with our clinical observations, we found detection performance of a prostate MRI based DL-CAD that the severity of rectal susceptibility artifacts is an system. independent high-risk factor for both false positives and false negatives of DL-CAD. Severe artifacts in the rectal Abbreviations region led to signal gain or loss in adjacent prostate tis- DL‑ CAD Deep learning‑based computer aided diagnosis sue and to local gland deformation of the prostate, which PCa Prostate cancer ssEPI Single‑shot echo ‑planar imaging impairs the performance of DL-CAD which relies on f‑DWI Full‑field‑of‑view DWI an accurate co-registration of T2WI and DWI. There - z‑DWI Zoomed‑field‑ of‑ view echo‑planar DWI fore, we hold that the reduction of rectal susceptibility AUC Ar ea under the curve CNR Contrast‑to‑noise ratio artifacts was one of the main reasons why the DL-CAD OR Odds ratio diagnostic accuracy using z-DWI was improved. It is also ADC Apparent diffusion coefficient indicated that good bowel preparation before prostate FOV Field‑of‑view PI‑RADS Prostate Imaging Reporting and Data System MRI examination may help to maintain the accuracy and GGG Gleason Grade Group stability of DL-CAD. LoS Case‑based level of suspicion Another interesting finding of our study which is incon - ROI Region‑of‑interest FROC F ree‑response receiver operating characteristics sistent with our initial expectations is that noise and phase VIF Variance inflation factor wrap-around were not found to be independent risk fac- GS Gleason score tors that interfered with the diagnostic efficiency of DL- SD Standard deviation eSNR Estimated signal‑to ‑noise ratio CAD. These aspects may not be important considerations for developing improved DWI scanning strategies of DL- Acknowledgements CAD for PCa diagnosis. According to the results, artifacts None. caused by artificial joint replacements and other artifacts Authors’ contributions out-of-body in DWI also were not key factors resulting in LH, CF: Conceptualization, Methodology, Investigation, Visualization, Writing − the misdiagnoses of DL-CAD. However, considering few original draft. XS: Methodology. AK, BL, HH, AT, TP, MC, IS, DW, PX, DS, FC, SS, ES: Methodology, Software, Formal analysis. JZ, LL: Resources, Supervision. RG, HB, patients suffering these artifacts, this result still needs to TB: Resources, Writing − review & editing. JG, YL: Supervision, Formal analysis. be further verified by larger samples. All authors read and approved the final manuscript. Our study had limitations. First, only one trained DL- Funding CAD based on full-FOV DWI was used for the com- This study received funding by the National Natural Science Foundation of parison of DWI sets. Although the effects of reduced China (Nos. 81901845, 81671791), Science Foundation of Shanghai Jiaotong image quality are typically not limited to a single model, University Affiliated Sixth People’s Hospital (No. 201818), and Shanghai key discipline of medical imaging (No: 2017ZZ02005). whether the risk factors we have identified affect the per - formance of other models still needs further verification. Availability of data and materials Second, only DWI images obtained from a single man- The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. ufacturer were included for comparison. Results from Hu et al. Cancer Imaging (2023) 23:6 Page 11 of 12 8. Yang F, Dogan N, Stoyanova R, Ford JC. Evaluation of radiomic texture Declarations feature error due to MRI acquisition and reconstruction: a simulation study utilizing ground truth. Physica Medica. 2018;50:26–36. Ethics approval and consent to participate 9. Winkel DJ, Tong A, Lou B, Kamen A, Comaniciu D, Disselhorst JA, et al. A This retrospective study was approved by the local ethics committee at novel deep learning based computer‑aided diagnosis system improves our institution [Approve No: 2022‑KY ‑073(K)]. As part of a prospective study the accuracy and efficiency of radiologists in Reading Biparametric mag‑ aiming to build a robust AI system for PCa diagnosis, all the enrolled subjects netic resonance images of the prostate: results of a multireader. Multicase signed the informed consent before they underwent the MRI examination Study Invest Radiol. 2021;56(10):605–13. and allowed us to use their data for a series of follow‑up studies about AI 10. Hu L, Wei L, Wang S, Fu C, Benker T, Zhao J. Better lesion conspicuity system building for PCa diagnosis. translates into improved prostate cancer detection: comparison of non‑parallel‑transmission‑zoomed‑DWI with conventional‑DWI. Abdom Consent for publication Radiol (NY ). 2021;46(12):5659–68. Consent for publication have been obtained from that person whose images 11. Bulten W, Pinckaers H, van Boven H, Vink R, de Bel T, van Ginneken B, et al. were contained in this article. Automated deep‑learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study. Lancet Oncol. 2020;21(2):233–41. Competing interests 12. Yu X, Lou B, Shi B, Winkel D, Szolar D. False Positive Reduction Using The authors declare that they have no competing interests. Multiscale Contextual Features for Prostate Cancer Detection in Multi‑ Parametric MRI Scans. In: In: 2020 IEEE 17th International Symposium on Author details Biomedical Imaging (ISBI). Iowa City: IEEE; 2020. p. 1355–9. Department of Diagnostic and Interventional Radiology, Shanghai 13. Hirano H, Minagi A, Takemoto K. Universal adversarial attacks on deep Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School neural networks for medical image classification. BMC Med Imaging. of Medicine, No. 600, Yi Shan Road, Shanghai 200233, China. MR Application 2021;21(1):9. Development, Siemens Shenzhen magnetic Resonance Ltd., Shenzhen, China. 14. Feuerlein S, Davenport MS, Krishnaraj A, Merkle EM, Gupta RT. Computed Department of Radiology, Xiangyang No.1 People’s Hospital, Hubei University high b‑ value diffusion‑ weighted imaging improves lesion contrast and con‑ of Medicine, Xiangyang 441000, China. MR Application Predevelopment, spicuity in prostate cancer. Prostate Cancer Prostatic Dis. 2015;18(2):155–60. Siemens Healthcare GmbH, Erlangen, Germany. Innovation Owner Artificial 15. Brendle C, Martirosian P, Schwenzer NF, Kaufmann S, Kruck S, Kramer U, Intelligence for Oncology, Siemens Healthcare GmbH, Erlangen, Germany. et al. Diffusion‑ weighted imaging in the assessment of prostate cancer: Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ, 7 8 comparison of zoomed imaging and conventional technique. Eur J USA. Radboud University Medical Center, Nijmegen, Netherlands. New Y ork Radiol. 2016;85(5):893–900. University, New York City, NY, USA. Charité, Universitätsmedizin Berlin, Berlin, 16. Finsterbusch J. Improving the performance of diffusion‑ weighted inner Germany. Eunpyeong St. Mary’s Hospital, Catholic University of Korea, Seoul, 11 12 field‑ of‑ view echo‑planar imaging based on 2D ‑selective radiofrequency Republic of Korea. Patero Clinic, Moscow, Russia. Universitätsspital Basel, 13 14 excitations by tilting the excitation plane. J Magn Reson Imaging. Basel, Switzerland. Changhai Hospital, Shanghai, China. Diagnostikum Graz 2012;35(4):984–92. Süd‑ West, Graz, Austria. Oregon Health and Science University, Portland, 16 17 17. Hu L, Zhou DW, Fu CX, Benkert T, Jiang CY, Li RT, et al. Advanced zoomed OR, USA. Loyola University Medical Center, Maywood, IL, USA. Medical diffusion‑ weighted imaging vs. full‑field‑ of‑ view diffusion‑ weighted University of Gdansk, Gdansk, Poland. Clinical Research Center, Shanghai imaging in prostate cancer detection: a radiomic features study. Eur Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School Radiol. 2021;31(3):1760–9. of Medicine, Shanghai 200233, China. Department of Radiology, Renmin 18. Rosenkrantz AB, Chandarana H, Pfeuffer J, Triolo MJ, Shaikh MB, Mossa DJ, Hospital of Wuhan University, Wuhan 430060, China. et al. Zoomed echo‑planar imaging using parallel transmission: impact on image quality of diffusion‑ weighted imaging of the prostate at 3T. Received: 16 November 2022 Accepted: 11 January 2023 Abdom Imaging. 2015;40(1):120–6. 19. Xu M, Zhang T, Li Z, Liu M, Zhang D. Towards evaluating the robustness of deep diagnostic models by adversarial attack. Med Image Anal. 2021;69:101977. 20. Allyn J, Allou N, Vidal C, Renou A, Ferdynus C. Adversarial attack on deep learning‑based dermatoscopic image recognition systems: risk of misdi‑ References agnosis due to undetectable image perturbations. Medicine (Baltimore). 1. Giganti F, Rosenkrantz AB, Villeirs G, Panebianco V, Stabile A, Emberton M, 2020;99(50):e23568. et al. The evolution of MRI of the prostate: the past, the present, and the 21. Akhtar N, Mian A. Threat of adversarial attacks on deep learning in com‑ future. AJR Am J Roentgenol. 2019;213(2):384–96. puter vision: a survey. Ieee Access. 2018;6:14410–30. 2. Turkbey B, Rosenkrantz AB, Haider MA, Padhani AR, Villeirs G, Macura 22. Vidnerova P, Neruda R. Vulnerability of classifiers to evolutionary gener ‑ KJ, et al. Prostate imaging reporting and data system version 2.1: 2019 ated adversarial examples. Neural Netw. 2020;127:168–81. update of prostate imaging reporting and data system version 2. Eur Urol. 23. Jolly MPD, Guetter C, Guehring J. Cardiac segmentation in MR cine data 2019;76(3):340–51. using inverse consistent deformable registration. In: In: Proceedings of 3. Donati OF, Mazaheri Y, Afaq A, Vargas HA, Zheng J, Moskowitz CS, et al. Prostate the 2010 IEEE International Symposium on Biomedical Imaging: From cancer aggressiveness: assessment with wholelesion hist ‑ ogram analysis of Nano to Macro. Rotterdam: IEEE; 2010. p. 484–7. the apparent diffusion coefficient. Radiology. 2014;271(1):143–52. 24. Kordbacheh H, Seethamraju RT, Weiland E, Kiefer B, Nickel MD, Chulroek T, 4. Hambrock T, Somford DM, Huisman HJ, van Oort IM, Witjes JA, Huls‑ et al. Image quality and diagnostic accuracy of complex‑averaged high b bergen‑ van de Kaa CA, et al. Relationship between apparent diffusion value images in diffusion‑ weighted MRI of prostate cancer. Abdom Radiol coefficients at 3.0‑ T MR imaging and Gleason grade in peripheral zone (NY ). 2019;44(6):2244–53. prostate cancer. Radiology. 2011;259(2):453–61. 25. Kuru TH, Wadhwa K, Chang RT, Echeverria LM, Roethke M, Polson A, et al. 5. Vargas HA, Akin O, Franiel T, Mazaheri Y, Zheng J, Moskowitz C, et al. Diffusion‑ Definitions of terms, processes and a minimum dataset for transperineal weighted endorectal MR imaging at 3 T for prostate cancer: tumor detection prostate biopsies: a standardization approach of the Ginsburg study and assessment of aggressiveness. Radiology. 2011;259(3):775–84. Group for Enhanced Prostate Diagnostics. BJU Int. 2013;112(5):568–77. 6. Cao R, Zhong X, Afshari S, Felker E, Suvannarerg V, Tubtawee T, et al. Per‑ 26. Bonekamp D, Schelb P, Wiesenfarth M, Kuder TA, Deister F, Stenzinger formance of deep learning and genitourinary radiologists in detection of A, et al. Histopathological to multiparametric MRI spatial mapping of prostate Cancer using 3‑ T multiparametric magnetic resonance imaging. extended systematic sextant and MR/TRUS‑fusion‑targeted biopsy of the J Magn Reson Imaging. 2021;54(2):474–83. prostate. Eur Radiol. 2019;29(4):1820–30. 7. Hiremath A, Shiradkar R, Merisaari H, Prasanna P, Ettala O, Taimen P, et al. 27. Yang D, Xu D, Zhou SK, Georgescu B, Chen M, Grbic S, et al. Automatic Test‑retest repeatability of a deep learning architecture in detecting and Liver Segmentation Using an Adversarial Image‑to ‑Image Network. segmenting clinically significant prostate cancer on apparent diffusion Cham: Springer International Publishing; 2017. p. 507–15. coefficient (ADC) maps. Eur Radiol. 2021;31(1):379–91. Hu et al. Cancer Imaging (2023) 23:6 Page 12 of 12 28. Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens G, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast Cancer. JAMA. 2017;318(22):2199–210. 29. Cao R, Mohammadian Bajgiran A, Afshari Mirak S, Shakeri S, Zhong X, Enzmann D, et al. Joint prostate Cancer detection and Gleason score prediction in mp‑MRI via FocalNet. IEEE Trans Med Imaging. 2019;38(11):2496–506. 30. Nandram B, Peiris T. Bayesian analysis of a ROC curve for categori‑ cal data using a skew‑binormal model. Statistics and Its Interface. 2018;11(2):369–84. 31. Nketiah G, Selnaes KM, Sandsmark E, Teruel JR, Kruger‑Stokke B, Bertilsson H, et al. Geometric distortion correction in prostate diffusion‑ weighted MRI and its effect on quantitative apparent diffusion coefficient analysis. Magn Reson Med. 2018;79(5):2524–32. 32. Rosenkrantz AB, Chandarana H, Hindman N, Deng FM, Babb JS, Taneja SS, et al. Computed diffusion‑ weighted imaging of the prostate at 3 T: impact on image quality and tumour detection. Eur Radiol. 2013;23(11):3170–7. 33. Klingebiel M, Ullrich T, Quentin M, Bonekamp D, Aissa J, Mally D, et al. Advanced diffusion weighted imaging of the prostate: comparison of readout‑segmented multi‑shot, parallel‑transmit and single ‑shot echo ‑ planar imaging. Eur J Radiol. 2020;130:109161. 34. Tamada T, Prabhu V, Li J, Babb JS, Taneja SS, Rosenkrantz AB. Assess‑ ment of prostate cancer aggressiveness using apparent diffusion coefficient values: impact of patient race and age. Abdom Radiol (NY ). 2017;42(6):1744–51. 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
Cancer Imaging – Springer Journals
Published: Jan 17, 2023
Keywords: Diffusion magnetic resonance imaging; Deep learning; Prostatic neoplasms; Artifact; Risk factor
Access the full text.
Sign up today, get DeepDyve free for 14 days.