Prostate cancer is typically detected after an elevated prostate-specific antigen (PSA) test or abnormal rectal exam, followed by biopsy. However, PSA is not cancer-specific -- it can rise due to benign conditions like prostate enlargement or inflammation -- and traditional ultrasound-guided biopsy samples predefined locations rather than suspicious regions, leading to missed cancers and unnecessary biopsies.
Multiparametric MRI (mpMRI) changed prostate cancer diagnostics by enabling radiologists to visualize suspicious lesions before biopsy, allowing targeted sampling of MRI-visible areas. The PI-RADS scoring system (version 2.1) was developed to standardize how radiologists score MRI lesions on a 1-to-5 scale. Despite this framework, significant variation persists between radiologists, especially less experienced readers.
Deep learning AI offers a potential solution to this variability. By training on large annotated datasets where every lesion's location and biopsy outcome are known, an AI model can learn to flag suspicious regions automatically and consistently, independent of reader experience. This study from the National Cancer Institute (NCI) and NVIDIA presents a fully automated deep learning system for this purpose.
A key practical design choice was to use biparametric MRI (bpMRI) rather than full multiparametric MRI. Biparametric MRI omits the dynamic contrast-enhanced (DCE) sequence, making it faster, less expensive, and free of contrast agent injection. This reflects the growing clinical trend toward bpMRI-first protocols and makes the system more broadly deployable.
The study combined two datasets: the publicly available PROSTATEx dataset from Radboud University in the Netherlands (145 biopsied patients) and a National Cancer Institute (NCI/NIH) institutional dataset of 380 patients from clinical trials conducted between 2015 and 2019. Together, 525 patients provided 844 lesions with biopsy-confirmed histopathology as ground truth.
The dataset was split into training (70 percent), validation (15 percent), and test (15 percent) cohorts. All suspicious lesion regions were manually contoured by a single expert genitourinary radiologist with over 15 years of experience reading more than 1,500 prostate MRI scans annually. Lesion masks were assigned a positive label only when the corresponding targeted biopsy showed cancer at any ISUP grade group (grade 1 or higher).
Two separate deep learning architectures were trained and compared. The first was a standard 3D U-Net, a widely used encoder-decoder architecture for volumetric medical image segmentation. The second was an AH-Net (Anisotropic Hybrid Network), which transfers 2D convolutional features to 3D anisotropic volumes, potentially better suited to the non-isotropic resolution of prostate MRI slices. Both models took three-channel input: T2-weighted imaging, ADC maps, and high b-value diffusion-weighted imaging.
Models were trained using a combination of Dice loss and cross-entropy loss with the Adam optimizer, run for 5,000 epochs on four NVIDIA V100 GPUs. Data augmentation techniques (random intensity shifts, flipping, zooming, Gaussian noise and smoothing) were applied to prevent overfitting. Performance was measured separately at the lesion level (sensitivity, positive predictive value, false positives per patient) and patient level (overall cancer detection sensitivity and specificity).
At the patient level, both models detected cancer with high sensitivity. In the validation set, AH-Net achieved 95.3 percent sensitivity and UNet achieved 92.2 percent. This means the models correctly identified that a patient had prostate cancer in approximately 9 out of 10 cases, comparable to expert radiologist performance for cancer-positive patients.
However, specificity was low -- 26.7 percent for AH-Net and only 6.7 percent for UNet in validation. This means the models frequently flagged cancer in patients who were actually cancer-free. This high false positive rate at the patient level reflects a fundamental limitation of current AI detection systems: they are tuned for sensitivity at the cost of specificity.
At the lesion level, the picture was similar. AH-Net achieved 74.4 percent sensitivity with a mean of 0.87 false positive lesions per patient in validation, outperforming UNet's 70.9 percent sensitivity with 1.41 false positives per patient. In the independent test set, UNet slightly outperformed AH-Net in sensitivity (72.8 vs. 63.0 percent), but AH-Net produced fewer false positives per patient (1.40 vs. 1.90).
Analysis of the false positives revealed a clinically meaningful pattern: approximately 15 percent of false positive lesions detected by AI corresponded to regions that radiologists had also flagged as suspicious but that were benign on targeted biopsy. These tend to be PI-RADS 3-5 lesions that looked suspicious on imaging but lacked cancer on sampling -- a known challenge in prostate MRI interpretation even for experienced human readers.
The most directly comparable published study, by Saha et al. from Radboud University, trained a deep attention model on 1,950 patients and achieved sensitivity of 92.3 percent with 1.69 false positives per patient on an internal test set, and 84.6 percent sensitivity with 2.22 false positives on an external set. Human radiologists in that comparison achieved 90.7 percent sensitivity with only 0.3 false positives per patient -- highlighting that false positive rate remains the key gap between AI and expert radiologists.
The current study's slightly lower lesion-level sensitivity (63 to 73 percent) compared to Saha et al. is partly attributed to the more inclusive ground truth definition: the NCI model was trained to detect all cancer grades including ISUP grade group 1 (Gleason 6), which is less visible on MRI for both humans and machines. Most other studies exclude low-grade cancers and report performance only for clinically significant cancer (ISUP grade 2 or higher).
The study's key strengths are its multi-center, multi-scanner training data spanning two continents, and its use of MRI-targeted biopsy (rather than systematic biopsy) as ground truth, which provides more reliable lesion-level correspondence. The decision to build on biparametric MRI rather than full mpMRI also increases practical applicability as bpMRI becomes more standard in clinical practice.
The authors frame the system not as a replacement for radiologists, but as a computer-aided detection tool to be used alongside human readers. Published evidence suggests that AI assistance most benefits less experienced radiologists while having minimal impact on already expert readers -- which reflects the appropriate role of such tools in a tiered clinical workflow.
The study's design was retrospective and "in silico" -- the AI model was never tested in a live clinical reading session where radiologists used it in real time. This is a critical gap, because the practical value of an AI detection aid depends on whether it changes radiologist decisions and whether those changes improve or worsen patient outcomes.
All lesion annotations were made by a single expert radiologist. While this ensures consistency, it introduces a risk of single-annotator bias. A model trained to replicate one radiologist's interpretation may not generalize to other reading styles or clinical environments. Multi-reader annotation studies would strengthen confidence in the ground truth labels.
The use of targeted biopsy rather than whole-mount surgical pathology as ground truth introduces some uncertainty about true negatives -- the study could not formally calculate specificity at the lesion level because no patients underwent systematic whole-prostate pathologic mapping. This is a trade-off between clinical realism and scientific rigor.
The next planned step is a prospective study comparing radiologist performance with and without AI assistance across readers of different experience levels. This design would provide the clearest evidence of whether the AI system translates to improved patient-level cancer detection, reduced unnecessary biopsies, or better triage of patients to the most appropriate diagnostic pathway.
The study demonstrates that a deep learning model trained on biparametric MRI data from two international institutions can detect prostate cancer lesions with sensitivity above 90 percent at the patient level, representing a meaningful achievement for fully automated lesion detection.
The primary remaining challenge is false positive lesion calls, averaging 1.4 to 1.9 per patient. Reducing this rate without sacrificing sensitivity is the central technical goal for the next generation of these models. Incorporating additional clinical context, combining lesion-level and patient-level predictions, and applying post-processing false positive filters are promising strategies.
The authors conclude that while the system is not ready for fully autonomous use, it is well-suited to assist radiologists as a second-reader tool, particularly in settings where highly experienced prostate MRI specialists are not available, and as the bpMRI-based workflow becomes increasingly standard in prostate cancer diagnosis.