Multiparametric MRI (mpMRI) has become central to the detection and staging of prostate cancer, offering valuable information about tumor location and extent without radiation. However, the quality of MRI interpretation depends heavily on radiologist experience, and significant variation exists between expert and non-expert readers.
The prostate MRI standard includes three key imaging sequences: T2-weighted imaging (T2WI), which reveals soft tissue anatomy; diffusion-weighted imaging (DWI), which highlights areas where water molecule movement is restricted by dense cancer cells; and apparent diffusion coefficient (ADC) maps, which quantify diffusion restriction. Together, these three sequences form the practical backbone of clinical prostate MRI without requiring contrast injection.
Prior deep learning studies have often used only one or two of these sequences, making it difficult to compare their performance against real clinical practice. Additionally, most studies labeled cancer regions using biopsy locations mapped by radiologists -- a method that can introduce inaccuracy because needle biopsies sample limited tissue and the labeled location may not precisely correspond to the true tumor extent.
This study aimed to close these gaps by building a deep learning model that uses all three clinically relevant MRI sequences and, critically, uses whole-mount histopathology -- a complete cross-sectional slice of the entire prostate specimen -- as the reference standard for labeling cancer regions. This approach produces far more accurate training labels than conventional biopsy-based delineation.
Whole-mount histopathology involves slicing the entire surgically removed prostate gland at 3-mm intervals and preparing complete cross-sectional tissue slides. Unlike standard biopsy, which samples only small cores from selected regions, whole-mount sections allow pathologists to see every tumor focus across the full gland in a single comprehensive review.
For deep learning training, the quality of labels determines the quality of the resulting model. When cancer region boundaries are drawn on MRI images by radiologists using whole-mount sections as a visual reference, the labels reflect the true spatial extent of cancer with far greater accuracy than labels derived from biopsy location estimates alone.
This labeling approach addresses a problem that has quietly undermined many published prostate MRI machine learning studies: if training labels are inaccurate -- either missing small lesions or mislocalizing tumor boundaries -- the model learns from flawed ground truth and its performance ceiling is artificially low.
In this study, all cancer regions on T2WI, DWI, and ADC images were manually outlined by two experienced radiologists under the supervision of a senior 15-year specialist, using the whole-mount histopathology as an explicit visual guide for every delineation. This rigorous labeling protocol is a key methodological strength that distinguishes this work from earlier studies.
The study enrolled 739 prostate cancer patients from Nanjing Drum Tower Hospital, including 466 confirmed cancer cases and 273 confirmed non-cancer cases. All patients underwent mpMRI before either radical prostatectomy (for cancer confirmation via whole-mount pathology) or targeted biopsy (for benign disease confirmation). An independent test cohort of 200 patients collected in a later period was used to evaluate real-world clinical performance.
The deep learning system operates as a three-stage sequential pipeline. First, a prostate gland segmentation network automatically identifies and outlines the entire prostate on T2WI. Second, a classification network analyzes image patches from all three sequences to determine whether the gland contains cancer. Third, if cancer is detected, a separate segmentation network delineates the exact location and extent of the lesion.
The gland segmentation network was built on a V-Net architecture -- a 3D convolutional network designed for volumetric medical image segmentation. The cancer classification network used DenseNet, an architecture where every layer receives feature information from all preceding layers, enabling efficient learning of complex image patterns with fewer parameters. The cancer segmentation network incorporated an SE-Block (Squeeze-and-Excitation Block) that automatically recalibrates feature maps to emphasize the most informative channels.
Images from both MRI scanners used in the study were acquired using standardized protocols. DWI was acquired at b-values of 50, 1,000, and 1,500 s/mm2 -- with the high b-value images providing strong diffusion contrast for cancer detection. ADC maps were computed from the b-1500 images. Lesion labeling was performed using the open-source ITK-SNAP software with whole-mount histopathology images displayed simultaneously as reference.
The prostate classification network achieved an AUC of 0.871 on the internal validation set and 0.797 on the independent test dataset -- the latter representing performance in a true clinical simulation with unseen patients. Standalone DL model diagnostic accuracy was 0.700 with sensitivity 0.710, specificity 0.690, and precision 0.696.
In direct comparison, the junior radiologist (5 years of experience) achieved accuracy of 0.645, sensitivity of 0.590, specificity of 0.700, and precision of 0.663 when reading scans alone. The senior radiologist (10 years of experience) performed better with accuracy 0.730, sensitivity 0.690, and specificity 0.770 -- but still below the DL model on sensitivity.
When the DL model was used as an assistant tool, junior radiologist performance rose significantly: accuracy improved from 0.645 to 0.755, sensitivity from 0.590 to 0.790, and specificity from 0.700 to 0.720. The DL assistance improved sensitivity by 0.200 absolute points -- meaning the junior reader missed far fewer cancers when aided by the model.
The senior radiologist with DL assistance achieved the best overall performance: accuracy 0.825, sensitivity 0.810, specificity 0.840, and precision 0.835. All improvements with DL assistance were statistically significant (p less than 0.05). The results demonstrate that even experienced radiologists benefit from AI-assisted review, catching cases they would otherwise miss.
The most important finding from a clinical perspective is not the DL model's standalone performance -- it is the magnitude of improvement DL assistance provides to less experienced readers. A junior radiologist with DL assistance outperformed the unassisted senior radiologist on sensitivity, suggesting that AI can help close the expertise gap that produces inconsistent prostate MRI interpretation across institutions.
The study deliberately excluded dynamic contrast-enhanced (DCE) MRI sequences, consistent with evidence that DCE adds limited diagnostic value beyond T2WI, DWI, and ADC in most clinical scenarios. According to PI-RADS v2.1, the DCE sequence provides meaningful contribution only for a narrow subset of PI-RADS score 3 lesions in the peripheral zone. By omitting DCE, the system avoids the cost and time of contrast injection while maintaining clinically comparable diagnostic performance.
A representative case illustrated in the paper shows a patient with a subtle right peripheral zone lesion rated PI-RADS score 3 that both radiologists read as negative. The DL model correctly flagged this region, and the DL-identified area overlapped with the true cancer location on the whole-mount specimen. This case exemplifies the model's potential value for detecting low-conspicuity lesions that human readers may overlook.
The study's approach of evaluating the DL model alongside actual radiologists in a real clinical simulation -- where readers interpreted scans without foreknowledge of the pathological results -- is more methodologically sound than studies that only report classification metrics in isolation. The head-to-head comparison directly addresses the question of whether the AI is genuinely useful in practice, not just technically accurate on a test set.
The study was conducted at a single center using two specific 3.0 Tesla MRI scanners. Deep learning models are sensitive to differences in scanner hardware, magnetic field strength, and acquisition protocol, meaning the model may not perform equally well if deployed on different machines at other hospitals. Multi-center validation would be required before widespread clinical adoption.
Because the study was retrospective, it did not incorporate clinical variables such as PSA trends, patient age stratification, or prior biopsy history into the model. Future versions that combine imaging-derived features with clinical biomarkers could improve performance further, particularly in challenging cases like high-PSA benign disease where the model may be more prone to false positives.
The authors emphasize that this DL model is designed as an adjunct tool, not a replacement for radiologist judgment. The intended workflow is for the model to provide a probabilistic overlay highlighting suspicious regions, which the radiologist then reviews and interprets in clinical context. This approach preserves the physician's decision-making authority while reducing cognitive load and providing a consistent second opinion.
The study demonstrates that whole-mount histopathology-referenced labeling produces training data of sufficient quality to build a clinically useful model even from a single-center dataset of modest size. This methodology offers a template for generating high-quality prostate MRI training datasets that future multi-center studies could adopt to improve model generalizability and performance.