Fully Automatic Deep Learning in Bi-institutional Prostate Magnetic Resonance Imaging

Netzer, Nils; Weißer, Cedric; Schelb, Patrick; Wang, Xianfeng; Qin, Xiaoyan; Görtz, Magdalena; Schütz, Viktoria; Radtke, Jan Philipp; Hielscher, Thomas; Schwab, Constantin; Kuder, Tristan Anselm; Gnirs, Regula; Hohenfellner, Markus; Maier‐Hein, Klaus H.; Bonekamp, David

doi:10.1097/rli.0000000000000791

Cited by 36 publications

(39 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Therefore, we put more data in the testing set (69 csPCa, 71 non-csPCa cases) to better evaluate and verify the generalization ability of the model. Our proposed biparametric model yielded high sensitivity of 98.6% for csPCa detection based on patients, as well as other studies with the sensitivity ranging from 82.9% to 97% ( 13 , 14 , 39 – 41 ). However, it is worth noting that the specificity was not as high as expected in our study (64.8%) and in other studies (47%–76%) ( 12 – 14 , 39 ).…”

Section: Discussionsupporting

confidence: 76%

Fully automated detection and localization of clinically significant prostate cancer on MR images using a cascaded convolutional neural network

et al. 2022

View full text Add to dashboard Cite

PurposeTo develop a cascaded deep learning model trained with apparent diffusion coefficient (ADC) and T2-weighted imaging (T2WI) for fully automated detection and localization of clinically significant prostate cancer (csPCa).MethodsThis retrospective study included 347 consecutive patients (235 csPCa, 112 non-csPCa) with high-quality prostate MRI data, which were randomly selected for training, validation, and testing. The ground truth was obtained using manual csPCa lesion segmentation, according to pathological results. The proposed cascaded model based on Res-UNet takes prostate MR images (T2WI+ADC or only ADC) as inputs and automatically segments the whole prostate gland, the anatomic zones, and the csPCa region step by step. The performance of the models was evaluated and compared with PI-RADS (version 2.1) assessment using sensitivity, specificity, accuracy, and Dice similarity coefficient (DSC) in the held-out test set.ResultsIn the test set, the per-lesion sensitivity of the biparametric (ADC + T2WI) model, ADC model, and PI-RADS assessment were 95.5% (84/88), 94.3% (83/88), and 94.3% (83/88) respectively (all p > 0.05). Additionally, the mean DSC based on the csPCa lesions were 0.64 ± 0.24 and 0.66 ± 0.23 for the biparametric model and ADC model, respectively. The sensitivity, specificity, and accuracy of the biparametric model were 95.6% (108/113), 91.5% (665/727), and 92.0% (773/840) based on sextant, and were 98.6% (68/69), 64.8% (46/71), and 81.4% (114/140) based on patients. The biparametric model had a similar performance to PI-RADS assessment (p > 0.05) and had higher specificity than the ADC model (86.8% [631/727], p< 0.001) based on sextant.ConclusionThe cascaded deep learning model trained with ADC and T2WI achieves good performance for automated csPCa detection and localization.

show abstract

Section: Discussionsupporting

confidence: 76%

Fully automated detection and localization of clinically significant prostate cancer on MR images using a cascaded convolutional neural network

et al. 2022

View full text Add to dashboard Cite

show abstract

“…Eight studies were included in the final selection [ 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 ]. An overview of the selected studies is shown in Table 1 .…”

Section: Resultsmentioning

confidence: 99%

“…Four studies (50%) used only two-dimensional (2D) operations within their model architecture, while two studies (25%) used three-dimensional (3D) operations. One study [ 23 ] used a combination of 2D and 3D architectures. One study used a 2.5 approach [ 25 ] by supplying two slices adjacent to the target slice as additional inputs to the model while using 2D operations within the model architecture.…”

Section: Resultsmentioning

confidence: 99%

“…All other studies trained their models from scratch. Cao et al [ 25 ] and Netzer et al [ 23 ] both evaluated the performance of a previously developed algorithm (‘FocalNet’ and ‘nnUNet’, respectively). All but one of the proposed methods provided methods for the visual inspection of the DL predictions, which included prediction heatmaps (five studies, 62.5%), gradient-weighted class activation mapping (Grad-CAM) [ 22 , 31 ], and class-activation maps [ 21 ].…”

Section: Resultsmentioning

confidence: 99%

“…The quality of reporting in the methods sections was significantly lower than that in the introduction ( p = 0.018) and discussion ( p = 0.018) sections. The assessment with QUADAS-2 found four studies at risk of bias [ 23 , 25 , 26 , 28 ] and two studies with concerns regarding applicability [ 22 , 25 ]. Specifically, a high risk of bias in the index test and reference standard was found in Cao et al [ 26 ] due to the exclusion of MRI invisible lesions from the analysis.…”

Section: Resultsmentioning

confidence: 99%

See 2 more Smart Citations

Comparative Performance of Deep Learning and Radiologists for the Diagnosis and Localization of Clinically Significant Prostate Cancer at MRI: A Systematic Review

Roest

Fransen

Kwee

et al. 2022

Life

View full text Add to dashboard Cite

Background: Deep learning (DL)-based models have demonstrated an ability to automatically diagnose clinically significant prostate cancer (PCa) on MRI scans and are regularly reported to approach expert performance. The aim of this work was to systematically review the literature comparing deep learning (DL) systems to radiologists in order to evaluate the comparative performance of current state-of-the-art deep learning models and radiologists. Methods: This systematic review was conducted in accordance with the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist. Studies investigating DL models for diagnosing clinically significant (cs) PCa on MRI were included. The quality and risk of bias of each study were assessed using the checklist for AI in medical imaging (CLAIM) and QUADAS-2, respectively. Patient level and lesion-based diagnostic performance were separately evaluated by comparing the sensitivity achieved by DL and radiologists at an identical specificity and the false positives per patient, respectively. Results: The final selection consisted of eight studies with a combined 7337 patients. The median study quality with CLAIM was 74.1% (IQR: 70.6–77.6). DL achieved an identical patient-level performance to the radiologists for PI-RADS ≥ 3 (both 97.7%, SD = 2.1%). DL had a lower sensitivity for PI-RADS ≥ 4 (84.2% vs. 88.8%, p = 0.43). The sensitivity of DL for lesion localization was also between 2% and 12.5% lower than that of the radiologists. Conclusions: DL models for the diagnosis of csPCa on MRI appear to approach the performance of experts but currently have a lower sensitivity compared to experienced radiologists. There is a need for studies with larger datasets and for validation on external data.

show abstract

Weakly Supervised MRI Slice‐Level Deep Learning Classification of Prostate Cancer Approximates Full Voxel‐ and Slice‐Level Annotation: Effect of Increasing Training Set Size

Weißer,

Netzer,

Görtz

et al. 2023

Magnetic Resonance Imaging

Self Cite

View full text Add to dashboard Cite

BackgroundWeakly supervised learning promises reduced annotation effort while maintaining performance.PurposeTo compare weakly supervised training with full slice‐wise annotated training of a deep convolutional classification network (CNN) for prostate cancer (PC).Study TypeRetrospective.SubjectsOne thousand four hundred eighty‐nine consecutive institutional prostate MRI examinations from men with suspicion for PC (65 ± 8 years) between January 2015 and November 2020 were split into training (N = 794, enriched with 204 PROSTATEx examinations) and test set (N = 695).Field Strength/Sequence1.5 and 3T, T2‐weighted turbo‐spin‐echo and diffusion‐weighted echo‐planar imaging.AssessmentHistopathological ground truth was provided by targeted and extended systematic biopsy. Reference training was performed using slice‐level annotation (SLA) and compared to iterative training utilizing patient‐level annotations (PLAs) with supervised feedback of CNN estimates into the next training iteration at three incremental training set sizes (N = 200, 500, 998). Model performance was assessed by comparing specificity at fixed sensitivity of 0.97 [254/262] emulating PI‐RADS ≥ 3, and 0.88–0.90 [231–236/262] emulating PI‐RADS ≥ 4 decisions.Statistical TestsReceiver operating characteristic (ROC) and area under the curve (AUC) was compared using DeLong and Obuchowski test. Sensitivity and specificity were compared using McNemar test. Statistical significance threshold was P = 0.05.ResultsTest set (N = 695) ROC‐AUC performance of SLA (trained with 200/500/998 exams) was 0.75/0.80/0.83, respectively. PLA achieved lower ROC‐AUC of 0.64/0.72/0.78. Both increased performance significantly with increasing training set size. ROC‐AUC for SLA at 500 exams was comparable to PLA at 998 exams (P = 0.28). ROC‐AUC was significantly different between SLA and PLA at same training set sizes, however the ROC‐AUC difference decreased significantly from 200 to 998 training exams. Emulating PI‐RADS ≥ 3 decisions, difference between PLA specificity of 0.12 [51/433] and SLA specificity of 0.13 [55/433] became undetectable (P = 1.0) at 998 exams. Emulating PI‐RADS ≥ 4 decisions, at 998 exams, SLA specificity of 0.51 [221/433] remained higher than PLA specificity at 0.39 [170/433]. However, PLA specificity at 998 exams became comparable to SLA specificity of 0.37 [159/433] at 200 exams (P = 0.70).Data ConclusionWeakly supervised training of a classification CNN using patient‐level‐only annotation had lower performance compared to training with slice‐wise annotations, but improved significantly faster with additional training data.Evidence Level3Technical EfficacyStage 2

show abstract

Fully Automatic Deep Learning in Bi-institutional Prostate Magnetic Resonance Imaging

Abstract: Supplemental digital content is available in the text.

Cited by 36 publications

References 41 publications

Fully automated detection and localization of clinically significant prostate cancer on MR images using a cascaded convolutional neural network

Fully automated detection and localization of clinically significant prostate cancer on MR images using a cascaded convolutional neural network

Comparative Performance of Deep Learning and Radiologists for the Diagnosis and Localization of Clinically Significant Prostate Cancer at MRI: A Systematic Review

Weakly Supervised MRI Slice‐Level Deep Learning Classification of Prostate Cancer Approximates Full Voxel‐ and Slice‐Level Annotation: Effect of Increasing Training Set Size

Contact Info

Product

Resources

About