This narrative review systematically catalogues AI algorithms published between 2018 and February 2021 for prostate cancer classification and detection on MRI. It covers 59 studies and assesses not just their technical performance but also their clinical utility -- asking whether AI actually improves patient outcomes, not just diagnostic accuracy metrics.
The review distinguishes two fundamentally different tasks: Computer-Aided Diagnosis (CADx) -- classifying a pre-identified region of interest as benign or malignant -- and Computer-Aided Detection (CADe) -- automatically finding suspicious lesions across the entire prostate without prior human annotation. CADx systems are more common (39 studies, 66%) than CADe systems (20 studies, 34%).
A key framework used throughout is a five-level clinical efficacy scale adapted from the radiology literature. Levels range from technical performance metrics (level 1) through diagnostic accuracy (level 2) and clinical workflow integration (level 3) to demonstrating patient outcome improvement (level 5). Most studies reach only level 2.
The review was motivated by the fact that prostate MRI interpretation is complex, time-consuming, and subject to significant inter-reader variability, especially for the intermediate PI-RADS 3 category. AI tools could potentially standardize reading, reduce workload, and improve detection of clinically significant cancer.
The majority of CADx studies (25 of 39) used conventional machine learning (ML) with a radiomics pipeline: extract handcrafted features from manually annotated regions of interest, select the most informative features, then train a classifier such as Random Forest, SVM, or Logistic Regression.
Input features were derived from multiple MRI sequences: T2-weighted (T2W) images (which show anatomical detail), Diffusion-Weighted Imaging (DWI) and its derived Apparent Diffusion Coefficient (ADC) maps (which reflect tissue cellularity), and Dynamic Contrast-Enhanced (DCE) sequences (which show vascularity). Most studies used multiparametric MRI combining two or more of these.
Cohort sizes for ML CADx studies ranged from 20 to 346 patients (median approximately 140). The majority (84%) used retrospective data from a single center, limiting assessment of generalizability. Only four studies (16%) used multicenter data, which is essential for testing whether algorithms perform consistently across different scanners and protocols.
AUC values for two-class lesion classification (distinguishing benign from malignant, or clinically insignificant from clinically significant cancer) ranged widely from 0.668 to 0.989 across studies. This large spread reflects differences in cohort composition, definition of clinically significant cancer (ISUP grade cutoff), and ground truth method (biopsy vs. radical prostatectomy).
Ten studies used deep learning (DL), primarily convolutional neural networks (CNNs), for two-class lesion classification. Unlike ML pipelines, DL does not require manual feature engineering -- the network learns relevant features directly from image patches or volumes of interest fed as input.
Several studies used transfer learning to overcome the challenge of small prostate MRI datasets. Pre-trained networks originally designed for other tasks (such as diabetic retinopathy detection or general image classification using VGG-16 or AlexNet) were fine-tuned on prostate MRI data. Zhong et al. showed that transfer learning improved AUC from 0.687 to 0.726 compared to training from scratch.
Four studies performed multi-class classification, attempting to assign lesions to specific ISUP grade groups (1 through 5) rather than a binary malignant/benign label. This harder task yielded lower AUC values -- the CNN by Abraham et al. achieved only 0.379 for ISUP grade 3, reflecting how difficult it is to distinguish intermediate grades from imaging alone.
Only two DL studies compared algorithm performance directly against radiologist readings on the same data. Both showed similar performance between the algorithm and readers, neither conclusively demonstrating superiority. This parity with expert radiologists is encouraging but not sufficient to justify replacing human review without prospective clinical validation.
CADe systems operate at the voxel level rather than on pre-annotated ROIs: they scan the entire prostate volume and generate probability maps highlighting suspicious areas. This is more clinically powerful than CADx because it removes the need for a radiologist to first identify and outline a lesion before AI analysis.
Seventeen studies for two-class lesion detection were included, with cohort sizes from 16 to 360 patients (median 163). DL dominated this category (11 of 17 studies, 65%), with U-Net architectures being particularly common for generating voxel-level segmentation maps. AUC values ranged from 0.645 to 0.997 across studies.
Three multicenter studies assessed CADe within a clinical workflow (efficacy level 3) by having radiologists read MRI first without AI assistance, then repeat with AI probability maps. The results were mixed: Zhu et al. showed AUC improvement from 0.83 to 0.89 with CADe assistance, while Mehralivand et al. found no significant improvement. These inconsistent findings make it unclear under what conditions AI detection tools help radiologists.
Three additional studies assessed multi-class detection, combining lesion localization with Gleason/ISUP grading. This remains the hardest task: one study achieved a weighted kappa of only 0.13 for simultaneous detection and grading, while another comparing AI to PI-RADS scoring showed moderate agreement (kappa 0.42), highlighting how much room for improvement remains.
As of February 2021, six AI products for prostate MRI had received regulatory clearance: three with FDA 510(k) clearance, one with CE marking, and two with both. Products include OnQ Prostate (Cortechs.ai), PROView (GE Medical Systems), Quantib Prostate, qp-Prostate (Quibim), JPC-01K (JLK Inc.), and Prostate MR (Siemens Healthineers).
Notably, most approved commercial products focus on prostate segmentation and volumetry -- measuring prostate size to calculate PSA density -- rather than cancer detection or lesion classification. Only one product (Siemens Prostate MR) offers AI-based lesion detection and PI-RADS classification, and its evidence base derives from a single published validation study.
The large gap between academic algorithms and commercial products reflects the high bar for regulatory approval and clinical deployment. Academic studies typically report internal validation on retrospective data; commercialization requires prospective validation, demonstrated clinical utility, and quality assurance processes to maintain performance over time as patient populations and scanner software evolve.
Only 15% of all 59 reviewed studies achieved efficacy level 3 or higher by testing AI within an actual clinical workflow or performing decision curve analysis to estimate the impact on patient management, such as reducing unnecessary biopsies. The vast majority remain at level 2 -- benchmarking accuracy metrics without demonstrating real-world benefit.
The most consistent finding across all 59 studies is insufficient external validation. When algorithms trained at one institution are tested at another, performance typically drops substantially. The most striking example: one study showed AUC declining from 0.83 on internal validation to 0.67 on an external cohort -- a 19% drop that could mean the difference between a useful and a harmful clinical tool.
Dataset sizes across the reviewed studies (median 162 patients) are far smaller than those used for AI systems in other medical imaging domains. A commercially deployed AI system for breast cancer detection was trained on 189,000 mammograms; the prostate CADe prototype by Winkel et al. used only 2,170 exams. Insufficient training data limits both the accuracy and the robustness of learned models.
The radiomics pipeline has a transparency advantage over deep learning -- each feature has a clear biological meaning -- but its reproducibility is compromised by inter-reader variability in ROI annotation, scanner differences, and correlated feature clusters. These factors mean that a radiomics model tuned at one institution may not reproduce its performance elsewhere.
Standardization of reporting and benchmarking is urgently needed. The review authors advocate for centralized challenge platforms (such as grand-challenge.org) where multiple algorithms are all evaluated on the same standardized datasets, enabling fair head-to-head comparisons that are currently impossible due to heterogeneous study designs.
The review concludes that AI algorithms for prostate MRI show promising stand-alone performance but are not yet ready for routine clinical use. The critical missing ingredients are larger and more diverse training datasets, rigorous external validation, and prospective studies that measure patient outcomes rather than just AUC values.
Future research should prioritize benchmarking against expert readers on the same cases -- not separate datasets -- and should include decision curve analysis to quantify how much unnecessary biopsy AI tools can prevent. Studies showing that AI reduces unnecessary procedures without missing clinically significant cancers would provide the clearest evidence of clinical value.
Integrating AI within actual clinical reading workflows -- where radiologists use AI outputs in real time -- is essential for understanding how AI assistance changes diagnostic decisions. Reader studies of this kind are far more informative than offline algorithm evaluations on historical data.
The authors predict continued rapid development in this field, with increasing multi-institutional collaboration to build larger datasets. Federated learning and shared challenge datasets offer promising paths to closing the gap between academic AI performance and the clinical robustness standards required for regulatory approval and patient safety.