Prostate cancer is the most common cancer and second leading cause of cancer-related death in American men. The standard approach to diagnosis -- measuring blood levels of prostate-specific antigen (PSA) -- is imprecise, leading to large numbers of unnecessary biopsies and significant overdiagnosis of low-risk tumors that may never cause harm.
Multiparametric MRI (mpMRI) of the prostate has substantially improved diagnosis by providing detailed tissue images that help target biopsies at suspicious areas. Radiologists score MRI findings using the PI-RADS system, a five-tier scale that estimates the probability that a lesion is clinically significant cancer. Scores of 4 or 5 generally prompt biopsy, while scores of 1 or 2 suggest the findings are likely benign.
Despite the PI-RADS standardization effort, studies consistently show significant variability between radiologists reading the same scan. The interpretation requires years of specialized training and remains highly subjective, meaning the same lesion can receive different scores from different readers. AI assistance could potentially reduce this variability and improve the consistency of prostate MRI reading across practice settings.
This study, conducted at the National Cancer Institute (NIH) in collaboration with NVIDIA, developed a cascaded deep learning system for prostate MRI that mimics how a radiologist approaches the scan. Rather than a single model trying to do everything, three specialized neural networks are chained together, each handling one step of the diagnostic process.
The first network detects and segments all suspicious lesions. The second acts as a filter, identifying and removing false positive detections that correspond to benign prostatic hyperplasia (BPH) nodules -- a common and harmless enlargement of the prostate that can mimic cancer on imaging. The third network then classifies each confirmed lesion with a PI-RADS score from 2 to 5.
This architecture is described as cascaded because the output of each stage feeds as input to the next, progressively refining the analysis. Using biparametric MRI (bpMRI) -- T2-weighted images plus diffusion-weighted imaging and ADC maps, without the contrast injection used in full multiparametric MRI -- makes the approach more widely applicable and less invasive.
Training data came from two sources: 1,043 clinical MRI scans from the NIH National Cancer Institute and 347 scans from the publicly available PROSTATEx dataset collected at Radboud University Medical Center in the Netherlands. Using data from two institutions with different scanners, imaging protocols, and patient populations makes the resulting model more robust than single-center studies.
All scans were re-read by a single highly experienced genitourinary radiologist with over 14 years of prostate MRI experience who reads more than 1,500 cases per year. Suspicious lesions were manually outlined and assigned a PI-RADS v2 category. Using a single expert reader ensured annotation consistency but also means the model was trained to replicate one expert's interpretive style rather than a consensus view.
The dataset was split into a training and validation set (89% of cases, 1,290 scans) and a held-out independent test set (11%, 150 scans). The test set was constructed to reflect the real distribution of PI-RADS categories seen in clinical practice, making performance estimates representative of expected real-world behavior.
The lesion detection component uses a 3D U-Net architecture, a convolutional encoder-decoder network specifically designed for volumetric medical image segmentation. Unlike 2D approaches that process one slice at a time, the 3D structure allows the network to recognize lesions based on their full three-dimensional appearance across all image slices simultaneously.
The BPH filter and PI-RADS classifier both use 3D residual neural networks, a deep learning architecture well-suited for classification tasks. Residual networks use skip connections that allow gradients to flow more easily during training, enabling much deeper networks to be trained effectively than was previously possible.
Performance was evaluated using several complementary metrics: sensitivity (the percentage of real lesions detected), positive predictive value (PPV) (the percentage of AI-flagged lesions that are truly lesions), and false discovery rate (FDR) (the percentage of AI detections that are false alarms). The Dice similarity coefficient measured how well the AI-drawn lesion boundaries matched the radiologist's outlines. For classification, overall accuracy in PI-RADS category assignment was calculated.
Without the BPH filter, the baseline detection system found 61% of lesions but produced an average of nearly one false alarm per patient -- a rate too high for practical clinical use. Adding the BPH filter made a meaningful difference: sensitivity dropped modestly to 56% while the false alarm rate fell from about 1.0 to 0.44 false positives per patient, and PPV improved from 45% to 63%.
The tradeoff is that the BPH filter reduces false positives at the cost of missing some real lesions -- those that the filter incorrectly classifies as BPH. Given that BPH nodules and some prostate cancers can look similar on MRI, this is an inherent challenge that reflects the ambiguity radiologists themselves face with these lesions.
A key finding from the biopsy correlation analysis: of the 79 lesions correctly detected by the AI, 82% were confirmed cancer-positive at targeted biopsy. Conversely, of the 57 lesions the AI missed (false negatives), 51% were benign at biopsy. This suggests the AI preferentially detects lesions that are more likely to be true cancers and tends to miss lesions that are less clinically significant.
Classifying detected lesions by PI-RADS category proved substantially harder than finding them. Overall classification accuracy in the independent test set was 30.8% across all four classes (PI-RADS 2 through 5) -- a result that appears modest but must be interpreted in context. A four-class classification problem has a random baseline of 25%, and exact category agreement is far harder than clinically meaningful agreement.
Performance varied significantly by PI-RADS category. The model completely failed on PI-RADS 2 lesions (0% correct), performed moderately on PI-RADS 3 and 5 lesions (around 52-67% correct), and struggled with PI-RADS 4 (15% correct in the test set). PI-RADS 2 and 4 lesions are generally smaller and have more ambiguous imaging features, while PI-RADS 5 lesions are by definition larger, better defined, and easier for both humans and algorithms to recognize.
When considering the clinically most important distinction -- whether a lesion is likely to harbor significant cancer (PI-RADS 4 or 5) versus not -- the performance picture improves. Most false negatives were PI-RADS 2, 3, and 4 lesions, consistent with these categories being the most visually ambiguous. Only 10% of the high-confidence PI-RADS 5 lesions were missed entirely.
The authors' honest appraisal is that exact PI-RADS category prediction may not be an appropriate goal for deep learning, at least not with current approaches. The PI-RADS system is based on subjective image features and even expert radiologists show substantial inter-reader disagreement on category assignments. Training an AI to replicate a single reader's scoring will inevitably reproduce that reader's particular interpretive style rather than a generalizable ground truth.
A more clinically valuable goal identified for future work is training directly on biopsy-confirmed cancer outcomes rather than on PI-RADS scores. If an AI could predict the probability of cancer based on histopathology as the ground truth, it would generate a calibrated risk score independent of subjective radiological categories -- potentially more useful than categorical PI-RADS assignment.
Three key limitations are acknowledged: training labels came from a single expert reader (introducing that reader's biases), dynamic contrast enhancement imaging was not included (this sequence helps distinguish PI-RADS 3 from 4 in the peripheral zone), and the system has not yet been prospectively tested against human radiologists reading the same cases in real time.
Despite its limitations, the system shows genuine clinical potential as a reading assistant rather than an autonomous reader. By pre-highlighting suspicious regions before the radiologist begins, the AI could reduce reading time, focus attention on high-risk areas, and potentially reduce the rate at which lesions are missed entirely -- particularly for radiologists with less specialized prostate MRI experience.
The BPH filter component specifically addresses a practical problem that developers encountered during testing: without it, the algorithm flagged BPH nodules as suspicious lesions at an unacceptably high rate. This real-world discovery led to a targeted architectural improvement, demonstrating the value of iterative development on clinically representative datasets.
The multi-institutional training data using the publicly available PROSTATEx dataset alongside NIH data is a strength that improves generalizability compared to many published studies that trained exclusively on single-institution data. Further external validation on datasets from other countries, imaging equipment, and clinical workflows will be essential before clinical deployment.