Deep Learning-Based Interpretable AI for Prostate T2W MRI Quality Evaluation

RATIONALE AND OBJECTIVES

Prostate MRI quality is essential in guiding prostate biopsies. However, assessment of MRI quality is subjective with variation. Quality degradation sources exert varying impacts based on the sequence under consideration, such as T2W versus DWI. As a result, employing sequence-specific techniques for quality assessment could yield more advantageous outcomes. This study aims to develop an AI tool that offers a more consistent evaluation of T2W prostate MRI quality, efficiently identifying suboptimal scans while minimizing user bias.

MATERIALS AND METHODS

This retrospective study included 1046 patients from three cohorts (ProstateX [n = 347], All-comer in-house [n = 602], enriched bad-quality MRI in-house [n = 97]) scanned between January 2011 and May 2022. An expert reader assigned T2W MRIs a quality score. A train-validation-test split of 70:15:15 was applied, ensuring equal distribution of MRI scanners and protocols across all partitions. T2W quality AI classification model was based on 3D DenseNet121 architecture using MONAI framework. In addition to multiclassification, binary classification was utilized (Classes 0/1 vs. 2). A score of 0 was given to scans considered non-diagnostic or unusable, a score of 1 was given to those with acceptable diagnostic quality with some usability but with some quality distortions present, and a score of 2 was given to those considered optimal diagnostic quality and usability. Partial occlusion sensitivity maps were generated for anatomical correlation. Three body radiologists assessed reproducibility within a subgroup of 60 test cases using weighted Cohen Kappa.

RESULTS

The best validation multiclass accuracy of 77.1% (121/157) was achieved during training. In the test dataset, multiclassification accuracy was 73.9% (116/157), whereas binary accuracy was 84.7% (133/157). Sub-class sensitivity for binary quality distortion classification for class 0 was 100% (18/18), and sub-class specificity for T2W classification of absence/minimal quality distortions for class 2 was 90.5% (95/105). All three readers showed moderate to substantial agreement with ground truth (R1-R3 κ = 0.588, κ = 0.649, κ = 0.487, respectively), moderate to substantial agreement with each other (R1-R2 κ = 0.599, R1-R3 κ = 0.612, R2-R3 κ = 0.685), fair to moderate agreement with AI (R1-R3 κ = 0.445, κ = 0.410, κ = 0.292, respectively). AI showed substantial agreement with ground truth (κ = 0.704). 3D quality heatmap evaluation revealed that the most critical non-diagnostic quality imaging features from an AI perspective related to obscuration of the rectoprostatic space (94.4%, 17/18).

CONCLUSION

The 3D AI model can assess T2W prostate MRI quality with moderate accuracy and translate whole sequence-level classification labels into 3D voxel-level quality heatmaps for interpretation. Image quality has a significant downstream impact on ruling out clinically significant cancers. AI may be able to help with reproducible identification of MRI sequences requiring re-acquisition with explainability.

Automated prostate gland segmentation in challenging clinical cases: comparison of three artificial intelligence methods

OBJECTIVE

Automated methods for prostate segmentation on MRI are typically developed under ideal scanning and anatomical conditions. This study evaluates three different prostate segmentation AI algorithms in a challenging population of patients with prior treatments, variable anatomic characteristics, complex clinical history, or atypical MRI acquisition parameters.

MATERIALS AND METHODS

A single institution retrospective database was queried for the following conditions at prostate MRI: prior prostate-specific oncologic treatment, transurethral resection of the prostate (TURP), abdominal perineal resection (APR), hip prosthesis (HP), diversity of prostate volumes (large ≥ 150 cc, small ≤ 25 cc), whole gland tumor burden, magnet strength, noted poor quality, and various scanners (outside/vendors). Final inclusion criteria required availability of axial T2-weighted (T2W) sequence and corresponding prostate organ segmentation from an expert radiologist. Three previously developed algorithms were evaluated: (1) deep learning (DL)-based model, (2) commercially available shape-based model, and (3) federated DL-based model. Dice Similarity Coefficient (DSC) was calculated compared to expert. DSC by model and scan factors were evaluated with Wilcox signed-rank test and linear mixed effects (LMER) model.

RESULTS

683 scans (651 patients) met inclusion criteria (mean prostate volume 60.1 cc [9.05-329 cc]). Overall DSC scores for models 1, 2, and 3 were 0.916 (0.707-0.971), 0.873 (0-0.997), and 0.894 (0.025-0.961), respectively, with DL-based models demonstrating significantly higher performance (p < 0.01). In sub-group analysis by factors, Model 1 outperformed Model 2 (all p < 0.05) and Model 3 (all p < 0.001). Performance of all models was negatively impacted by prostate volume and poor signal quality (p < 0.01). Shape-based factors influenced DL models (p < 0.001) while signal factors influenced all (p < 0.001).

CONCLUSION

Factors affecting anatomical and signal conditions of the prostate gland can adversely impact both DL and non-deep learning-based segmentation models.

Deep Learning-Based Detection and Classification of Bone Lesions on Staging Computed Tomography in Prostate Cancer: A Development Study

RATIONALE AND OBJECTIVES

Efficiently detecting and characterizing metastatic bone lesions on staging CT is crucial for prostate cancer (PCa) care. However, it demands significant expert time and additional imaging such as PET/CT. We aimed to develop an ensemble of two automated deep learning AI models for 1) bone lesion detection and segmentation and 2) benign vs. metastatic lesion classification on staging CTs and to compare its performance with radiologists.

MATERIALS AND METHODS

This retrospective study developed two AI models using 297 staging CT scans (81 metastatic) with 4601 benign and 1911 metastatic lesions in PCa patients. Metastases were validated by follow-up scans, bone biopsy, or PET/CT. Segmentation AI (3DAISeg) was developed using the lesion contours delineated by a radiologist. 3DAISeg performance was evaluated with the Dice similarity coefficient, and classification AI (3DAIClass) performance on AI and radiologist contours was assessed with F1-score and accuracy. Training/validation/testing data partitions of 70:15:15 were used. A multi-reader study was performed with two junior and two senior radiologists within a subset of the testing dataset (n = 36).

RESULTS

In 45 unseen staging CT scans (12 metastatic PCa) with 669 benign and 364 metastatic lesions, 3DAISeg detected 73.1% of metastatic (266/364) and 72.4% of benign lesions (484/669). Each scan averaged 12 extra segmentations (range: 1-31). All metastatic scans had at least one detected metastatic lesion, achieving a 100% patient-level detection. The mean Dice score for 3DAISeg was 0.53 (median: 0.59, range: 0-0.87). The F1 for 3DAIClass was 94.8% (radiologist contours) and 92.4% (3DAISeg contours), with a median false positive of 0 (range: 0-3). Using radiologist contours, 3DAIClass had PPV and NPV rates comparable to junior and senior radiologists: PPV (semi-automated approach AI 40.0% vs. Juniors 32.0% vs. Seniors 50.0%) and NPV (AI 96.2% vs. Juniors 95.7% vs. Seniors 91.9%). When using 3DAISeg, 3DAIClass mimicked junior radiologists in PPV (pure-AI 20.0% vs. Juniors 32.0% vs. Seniors 50.0%) but surpassed seniors in NPV (pure-AI 93.8% vs. Juniors 95.7% vs. Seniors 91.9%).

CONCLUSION

Our lesion detection and classification AI model performs on par with junior and senior radiologists in discerning benign and metastatic lesions on staging CTs obtained for PCa.

PI-RADS Version 2.0 Versus Version 2.1: Comparison of Prostate Cancer Gleason Grade Upgrade and Downgrade Rates From MRI-Targeted Biopsy to Radical Prostatectomy

BACKGROUND. Precise risk stratification through MRI/ultrasound (US) fusion-guided targeted biopsy (TBx) can guide optimal prostate cancer (PCa) management. OBJECTIVE. The purpose of this study was to compare PI-RADS version 2.0 (v2.0) and PI-RADS version 2.1 (v2.1) in terms of the rates of International Society of Urological Pathology (ISUP) grade group (GG) upgrade and downgrade from TBx to radical prostatectomy (RP). METHODS. This study entailed a retrospective post hoc analysis of patients who underwent 3-T prostate MRI at a single institution from May 2015 to March 2023 as part of three prospective clinical trials. Trial participants who underwent MRI followed by MRI/US fusion-guided TBx and RP within a 1-year interval were identified. A single genitourinary radiologist performed clinical interpretations of the MRI examinations using PI-RADS v2.0 from May 2015 to March 2019 and PI-RADS v2.1 from April 2019 to March 2023. Upgrade and downgrade rates from TBx to RP were compared using chi-square tests. Clinically significant cancer was defined as ISUP GG2 or greater. RESULTS. The final analysis included 308 patients (median age, 65 years; median PSA density, 0.16 ng/mL2). The v2.0 group (n = 177) and v2.1 group (n = 131) showed no significant difference in terms of upgrade rate (29% vs 22%, respectively; p = .15), downgrade rate (19% vs 21%, p = .76), clinically significant upgrade rate (14% vs 10%, p = .27), or clinically significant downgrade rate (1% vs 1%, p > .99). The upgrade rate and downgrade rate were also not significantly different between the v2.0 and v2.1 groups when stratifying by index lesion PI-RADS category or index lesion zone, as well as when assessed only in patients without a prior PCa diagnosis (all p > .01). Among patients with GG2 or GG3 at RP (n = 121 for v2.0; n = 103 for v2.1), the concordance rate between TBx and RP was not significantly different between the v2.0 and v2.1 groups (53% vs 57%, p = .51). CONCLUSION. Upgrade and downgrade rates from TBx to RP were not significantly different between patients whose MRI examinations were clinically interpreted using v2.0 or v2.1. CLINICAL IMPACT. Implementation of the most recent PI-RADS update did not improve the incongruence in PCa grade assessment between TBx and surgery.

Comparison of MRI-Based Staging and Pathologic Staging for Predicting Biochemical Recurrence of Prostate Cancer After Radical Prostatectomy

BACKGROUND. Currently most clinical models for predicting biochemical recurrence (BCR) of prostate cancer (PCa) after radical prostatectomy (RP) incorporate staging information from RP specimens, creating a gap in preoperative risk assessment. OBJECTIVE. The purpose of our study was to compare the utility of presurgical staging information from MRI and postsurgical staging information from RP pathology in predicting BCR in patients with PCa. METHODS. This retrospective study included 604 patients (median age, 60 years) with PCa who underwent prostate MRI before RP from June 2007 to December 2018. A single genitourinary radiologist assessed MRI examinations for extraprostatic extension (EPE) and seminal vesicle invasion (SVI) during clinical interpretations. The utility of EPE and SVI on MRI and RP pathology for BCR prediction was assessed through Kaplan-Meier and Cox proportional hazards analyses. Established clinical BCR prediction models, including the University of California San Francisco Cancer of the Prostate Risk Assessment (UCSF-CAPRA) model and the Cancer of the Prostate Risk Assessment Postsurgical (CAPRA-S) model, were evaluated in a subset of 374 patients with available Gleason grade groups from biopsy and RP pathology; two CAPRA-MRI models (CAPRA-S model with modifications to replace RP pathologic staging features with MRI staging features) were also assessed. RESULTS. Univariable predictors of BCR included EPE on MRI (HR = 3.6), SVI on MRI (HR = 4.4), EPE on RP pathology (HR = 5.0), and SVI on RP pathology (HR = 4.6) (all p < .001). Three-year BCR-free survival (RFS) rates for patients without versus with EPE were 84% versus 59% for MRI and 89% versus 58% for RP pathology, and 3-year RFS rates for patients without versus with SVI were 82% versus 50% for MRI and 83% versus 54% for RP histology (all p < .001). For patients with T3 disease on RP pathology, 3-year RFS rates were 67% and 41% for patients without and with T3 disease on MRI. AUCs of CAPRA models, including CAPRA-MRI models, ranged from 0.743 to 0.778. AUCs were not significantly different between CAPRA-S and CAPRA-MRI models (p > .05). RFS rates were significantly different between low- and intermediate-risk groups for only CAPRA-MRI models (80% vs 51% and 74% vs 44%; both p < .001). CONCLUSION. Presurgical MRI-based staging features perform comparably to postsurgical pathologic staging features for predicting BCR. CLINICAL IMPACT. MRI staging can preoperatively identify patients at high BCR risk, helping to inform early clinical decision-making. TRIAL REGISTRATION. ClinicalTrials.gov NCT00026884 and NCT02594202.

Deep Learning-Based T2-weighted MR Image Quality Assessment and Its Impact on Prostate Cancer Detection Rates

BACKGROUND

Image quality evaluation of prostate MRI is important for successful implementation of MRI into localized prostate cancer diagnosis.

PURPOSE

To examine the impact of image quality on prostate cancer detection using an in-house previously developed artificial intelligence (AI) algorithm.

STUDY TYPE

Retrospective.

SUBJECTS

615 consecutive patients (median age 67 [interquartile range [IQR]: 61-71] years) with elevated serum PSA (median PSA 6.6 [IQR: 4.6-9.8] ng/mL) prior to prostate biopsy.

FIELD STRENGTH/SEQUENCE

3.0T/T2-weighted turbo-spin-echo MRI, high b-value echo-planar diffusion-weighted imaging, and gradient recalled echo dynamic contrast-enhanced.

ASSESSMENTS

Scans were prospectively evaluated during clinical readout using PI-RADSv2.1 by one genitourinary radiologist with 17 years of experience. For each patient, T2-weighted images (T2WIs) were classified as high-quality or low-quality based on evaluation of both general distortions (eg, motion, distortion, noise, and aliasing) and perceptual distortions (eg, obscured delineation of prostatic capsule, prostatic zones, and excess rectal gas) by a previously developed in-house AI algorithm. Patients with PI-RADS category 1 underwent 12-core ultrasound-guided systematic biopsy while those with PI-RADS category 2-5 underwent combined systematic and targeted biopsies. Patient-level cancer detection rates (CDRs) were calculated for clinically significant prostate cancer (csPCa, International Society of Urological Pathology Grade Group ≥2) by each biopsy method and compared between high- and low-quality images in each PI-RADS category.

STATISTICAL TESTS

Fisher's exact test. Bootstrap 95% confidence intervals (CI). A P value <0.05 was considered statistically significant.

RESULTS

385 (63%) T2WIs were classified as high-quality and 230 (37%) as low-quality by AI. Targeted biopsy with high-quality T2WIs resulted in significantly higher clinically significant CDR than low-quality images for PI-RADS category 4 lesions (52% [95% CI: 43-61] vs. 32% [95% CI: 22-42]). For combined biopsy, there was no significant difference in patient-level CDRs for PI-RADS 4 between high- and low-quality T2WIs (56% [95% CI: 47-64] vs. 44% [95% CI: 34-55]; P = 0.09).

DATA CONCLUSION

Higher quality T2WIs were associated with better targeted biopsy clinically significant cancer detection performance for PI-RADS 4 lesions. Combined biopsy might be needed when T2WI is lower quality.

LEVEL OF EVIDENCE

2 TECHNICAL EFFICACY: Stage 1.

The Low Rate of Adherence to Checklist for Artificial Intelligence in Medical Imaging Criteria Among Published Prostate MRI Artificial Intelligence Algorithms

OBJECTIVE

To determine the rigor, generalizability, and reproducibility of published classification and detection artificial intelligence (AI) models for prostate cancer (PCa) on MRI using the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) guidelines, a 42-item checklist that is considered a measure of best practice for presenting and reviewing medical imaging AI research.

MATERIALS AND METHODS

This review searched English literature for studies proposing PCa AI detection and classification models on MRI. Each study was evaluated with the CLAIM checklist. The additional outcomes for which data were sought included measures of AI model performance (eg, area under the curve [AUC], sensitivity, specificity, free-response operating characteristic curves), training and validation and testing group sample size, AI approach, detection versus classification AI, public data set utilization, MRI sequences used, and definition of gold standard for ground truth. The percentage of CLAIM checklist fulfillment was used to stratify studies into quartiles. Wilcoxon's rank-sum test was used for pair-wise comparisons.

RESULTS

In all, 75 studies were identified, and 53 studies qualified for analysis. The original CLAIM items that most studies did not fulfill includes item 12 (77% no): de-identification methods; item 13 (68% no): handling missing data; item 15 (47% no): rationale for choosing ground truth reference standard; item 18 (55% no): measurements of inter- and intrareader variability; item 31 (60% no): inclusion of validated interpretability maps; item 37 (92% no): inclusion of failure analysis to elucidate AI model weaknesses. An AUC score versus percentage CLAIM fulfillment quartile revealed a significant difference of the mean AUC scores between quartile 1 versus quartile 2 (0.78 versus 0.86, P = .034) and quartile 1 versus quartile 4 (0.78 versus 0.89, P = .003) scores. Based on additional information and outcome metrics gathered in this study, additional measures of best practice are defined. These new items include disclosure of public dataset usage, ground truth definition in comparison to other referenced works in the defined task, and sample size power calculation.

CONCLUSION

A large proportion of AI studies do not fulfill key items in CLAIM guidelines within their methods and results sections. The percentage of CLAIM checklist fulfillment is weakly associated with improved AI model performance. Additions or supplementations to CLAIM are recommended to improve publishing standards and aid reviewers in determining study rigor.

A Cascaded Deep Learning-Based Artificial Intelligence Algorithm for Automated Lesion Detection and Classification on Biparametric Prostate Magnetic Resonance Imaging

RATIONALE AND OBJECTIVES

Prostate MRI improves detection of clinically significant prostate cancer; however, its diagnostic performance has wide variation. Artificial intelligence (AI) has the potential to assist radiologists in the detection and classification of prostatic lesions. Herein, we aimed to develop and test a cascaded deep learning detection and classification system trained on biparametric prostate MRI using PI-RADS for assisting radiologists during prostate MRI read out.

MATERIALS AND METHODS

T2-weighted, diffusion-weighted (ADC maps, high b value DWI) MRI scans obtained at 3 Tesla from two institutions (n = 1043 in-house and n = 347 Prostate-X, respectively) acquired between 2015 to 2019 were used for model training, validation, testing. All scans were retrospectively reevaluated by one radiologist. Suspicious lesions were contoured and assigned a PI-RADS category. A 3D U-Net-based deep neural network was used to train an algorithm for automated detection and segmentation of prostate MRI lesions. Two 3D residual neural network were used for a 4-class classification task to predict PI-RADS categories 2 to 5 and BPH. Training and validation used 89% (n = 1290 scans) of the data using 5 fold cross-validation, the remaining 11% (n = 150 scans) were used for independent testing. Algorithm performance at lesion level was assessed using sensitivities, positive predictive values (PPV), false discovery rates (FDR), classification accuracy, Dice similarity coefficient (DSC). Additional analysis was conducted to compare AI algorithm's lesion detection performance with targeted biopsy results.

RESULTS

Median age was 66 years (IQR = 60-71), PSA 6.7 ng/ml (IQR = 4.7-9.9) from in-house cohort. In the independent test set, algorithm correctly detected 111 of 198 lesions leading to 56.1% (49.3%-62.6%) sensitivity. PPV was 62.7% (95% CI 54.7%-70.7%) with FDR of 37.3% (95% CI 29.3%-45.3%). Of 79 true positive lesions, 82.3% were tumor positive at targeted biopsy, whereas of 57 false negative lesions, 50.9% were benign at targeted biopsy. Median DSC for lesion segmentation was 0.359. Overall PI-RADS classification accuracy was 30.8% (95% CI 24.6%-37.8%).

CONCLUSION

Our cascaded U-Net, residual network architecture can detect, classify cancer suspicious lesions at prostate MRI with good detection, reasonable classification performance metrics.