Prostate cancer is the second deadliest cancer among American men. Multi-parametric MRI (mpMRI), which combines T2-weighted anatomical imaging with diffusion-weighted imaging and ADC maps, has become the most sensitive non-invasive tool for detecting and localizing prostate cancer, and is increasingly used to guide targeted biopsies.
Despite its promise, prostate MRI interpretation suffers from high rates of false positives, false negatives, and substantial disagreement between radiologists. PI-RADS scores, the standardized 1-5 rating system radiologists use to characterize suspicious lesions, show wide variability in assignment, with studies reporting false positive rates of 32 to 50 percent and missed aggressive cancers in 12 to 34 percent of patients depending on reader experience.
Machine learning models that automatically detect and localize cancer on MRI could help standardize interpretation. However, these models need accurate training labels that precisely indicate cancer location and extent. The fundamental problem is that existing studies have used six different types of labels, ranging from radiologist outlines to pathologist annotations on surgical specimens to labels generated by AI algorithms on histopathology images, and no prior study had systematically compared which label type produces the best machine learning models.
This study from Stanford University directly compared four label types and four deep learning architectures to determine the optimal approach for training automated prostate cancer detection systems, with the hypothesis that digital pathologist labels derived from AI-generated Gleason pattern analysis on whole-mount histopathology images would produce the most accurate and generalizable models.
The study enrolled 390 patients across two cohorts from Stanford University Medical Center. Cohort C1 comprised 115 men who had both pre-operative mpMRI and post-operative whole-mount histopathology images after radical prostatectomy, enabling precise spatial registration between what was seen on MRI and what was found in the removed prostate. Cohort C2 included 275 men who underwent MRI-guided targeted biopsy for PI-RADS 3 or above lesions.
For the 115 radical prostatectomy patients, each lesion was labeled using four different strategies. The first, Rad (radiologist labels), used outlines drawn by radiologists during routine clinical care, confirmed by pathology to contain cancer. The second, Path (pathologist labels), used cancer outlines drawn by an expert pathologist directly on the high-resolution whole-mount histopathology images, then mapped onto the pre-operative MRI using automated registration software (RAPSODI platform).
The third and fourth label types used a previously validated digital pathologist -- a deep learning model trained on 1,133 prostate biopsies to assign pixel-level Gleason patterns (indolent Gleason 3 versus aggressive Gleason 4 and above) to every pixel of histopathology images. This model had demonstrated a kappa of 0.927 in distinguishing benign from cancerous tissue when externally validated at Stanford against expert uropathologists. Its pixel-level Gleason predictions were then converted into either lesion-level digital pathologist labels (DPath-Lesion), where an entire lesion was classified as aggressive or indolent, or pixel-level digital pathologist labels (DPath-Pixel), which preserved the spatial distribution of aggressive and indolent cancer components within each lesion.
The key difference between DPath-Pixel and all other label types is its ability to separately identify aggressive and indolent components within the same mixed lesion, rather than labeling the entire lesion by its dominant cancer type. This granularity is not achievable by any human annotation approach because the time required for manual gland-by-gland Gleason annotation of all slices is prohibitively large for clinical or research settings.
Mapping pathology labels from surgical specimens onto pre-operative MRI is technically demanding because the prostate changes shape when removed, the tissue is sectioned and stained, and the MRI is acquired weeks to months before surgery. The study used the automated RAPSODI registration platform to align whole-mount histopathology images with pre-operative MRI, with reported average registration errors of approximately 2 mm at the prostate border and 3 mm inside the prostate.
To ensure same-plane sectioning, the prostates were removed and sliced using patient-specific 3D-printed molds generated from the pre-operative MRI. These custom molds constrained the sectioning planes to match the axial T2-weighted MRI acquisition plane and slice thickness, substantially reducing the geometric distortion that would otherwise occur if tissue were cut at arbitrary orientations relative to the MRI.
All four imaging modalities (T2w MRI, ADC maps, and histopathology images from both cohorts) were preprocessed to a standardized pixel size of 0.29 x 0.29 mm with a 224 x 224 spatial dimension around the prostate. MRI intensities were standardized and normalized to account for scanner-to-scanner variation across the diverse clinical dataset. Lesions with volumes below 250 mm3 were excluded from analysis as too small to be visible on MRI and not clinically significant by PI-RADS v2 guidelines.
Four deep learning architectures previously used for prostate cancer detection were compared: SPCNet (Stanford Prostate Cancer Network), U-Net, branched U-Net, and DeepLabv3+. Each architecture was trained with all four label types, producing 16 separate digital radiologist models, all trained with identical preprocessing, class-balanced cross-entropy loss function, batch size, Adam optimizer, and 30 training epochs. This factorial design ensured that differences in detection performance could be attributed to the label type rather than architectural or training differences.
When the four label types were directly compared against each other for their accuracy in capturing cancer extent, radiologist labels (Rad) showed consistently inferior performance. Radiologist labels covered only approximately 68 percent of the lesion volume identified by pathologist labels (Path), missed cancers in 11 percent of patients who ultimately had clinically significant cancer on surgical pathology, and showed Dice overlaps of only 0.24 to 0.28 with pathology labels for cancer and aggressive cancer detection.
The poor performance of radiologist labels reflects known limitations: radiologists annotate suspicious lesions during routine clinical care without access to pathology, do not annotate every slice where a lesion continues, and cannot detect the roughly 20 percent of prostate cancers that are invisible or barely visible on MRI. These lesions, though harder to see, are often clinically significant and require treatment.
In contrast, the lesion-level digital pathologist labels (DPath-Lesion) showed high concordance with expert pathologist labels (Path), achieving Dice overlaps of 0.79 to 0.82 and lesion-level ROC-AUC of 0.94 to 1.00 for cancer detection and 0.86 to 0.97 for aggressive cancer detection. This level of agreement demonstrates that the automated Gleason grading algorithm captures essentially the same cancer extent as a trained human pathologist, while being applicable to all tissue slices without requiring prohibitive manual annotation time.
The pixel-level digital pathologist labels (DPath-Pixel) showed a uniquely important capability: they selectively labeled aggressive (Gleason 4 and above) and indolent (Gleason 3) cancer components separately within mixed lesions. No other label type can achieve this level of intra-lesion spatial resolution, since both radiologists and pathologists classify the entire lesion by its predominant cancer type rather than mapping each component independently.
Models trained with digital pathologist labels (DPath-Lesion) consistently outperformed or matched models trained with radiologist labels (Rad) across all four deep learning architectures and both patient cohorts. In the radical prostatectomy test cohort (C1), SPCNet trained with DPath-Lesion achieved an aggressive lesion ROC-AUC of 0.92 versus 0.88 for SPCNet trained with Rad, and outperformed Rad-trained models in both lesion detection rate and cancer extent localization.
Models trained with pathologist labels (Path) and digital pathologist labels (DPath-Lesion) showed closely matched performance, confirming that automated digital pathologist Gleason grading can serve as an effective substitute for manual pathologist annotation -- eliminating the hours of expert labor required to annotate every tissue slice while maintaining equivalent model training quality.
In the independent targeted biopsy cohort (C2), models trained with digital pathologist labels demonstrated generalizable performance comparable to radiologist label-trained models, with aggressive lesion ROC-AUCs of 0.87 to 0.88 across architectures for DPath-trained versus 0.86 to 0.89 for Rad-trained models. This generalizability is critical for clinical deployment because biopsy cohorts have a different distribution of cancer size, grade, and location than surgical cohorts.
Only models trained with pixel-level digital pathologist labels (DPath-Pixel) could selectively identify indolent cancer components in mixed lesions on MRI, achieving an indolent cancer ROC-AUC of 0.64 and a Dice of 0.12 in the biopsy cohort -- compared to effectively zero for all other label types, which could not distinguish indolent from aggressive regions within the same lesion. This capability, though still imperfect, represents a clinically important first step toward using MRI to guide focal treatment decisions.
Among the four deep learning architectures tested, SPCNet (Stanford Prostate Cancer Network) consistently achieved the best lesion detection performance, outperforming U-Net, branched U-Net, and DeepLabv3+ across most metrics and most label types in both patient cohorts. This finding confirms that architectural choices matter independently of label quality.
Importantly, the superiority of digital pathologist labels over radiologist labels was consistent across all four architectures. This architecture-independence is a key finding: it demonstrates that the label quality effect is not an artifact of one particular model design but is a generalizable property of the training data. Any deep learning approach for prostate cancer detection will benefit from switching from radiologist labels to digital pathologist labels.
The best models across architectures achieved highest Dice overlaps when evaluated against radiologist labels rather than pathologist labels, even though models were trained with pathology-derived labels. This counterintuitive finding reflects that MRI-visible, radiologist-detected lesions are by definition the most prominent tumors, making them easier for any model to detect regardless of training labels. Harder-to-see lesions captured by pathology labels but not radiologist labels are inherently lower-contrast on MRI.
Five-fold cross-validation across the 75-patient training cohort confirmed consistent performance, with digital pathologist label-trained models showing stable results across all folds, supporting the reliability of the findings despite the relatively modest single-institution training dataset size.
This study demonstrates that AI-generated Gleason pattern labels derived from whole-mount histopathology images, called digital pathologist labels, can bridge the gap between prostate radiology and pathology by providing training data for MRI-based cancer detection that is simultaneously more accurate than radiologist annotations, as accurate as expert pathologist annotations, and far more practical to generate at scale than either human labeling approach.
The clinical value of reliable prostate cancer detection on MRI is direct: accurate digital radiologist models can help standardize currently variable radiologist interpretations, detect aggressive cancers that are currently missed before biopsy, reduce unnecessary biopsies in men with indolent or no cancer, and localize aggressive cancer components within mixed lesions to guide precision treatment decisions like focal therapy versus whole-gland treatment.
The practical advantage of digital pathologist labels goes beyond performance: they eliminate the need for expert pathologists to spend hours manually annotating every tissue slice with Gleason patterns -- a task that is both impractical and subject to significant inter- and intra-pathologist variability. Automated Gleason grading algorithms have already demonstrated kappa values exceeding 0.9 in external validation, making them reliable enough to serve as the annotation source for AI training pipelines.
Acknowledged limitations include the single-institution dataset, the retrospective design, and the use of a digital pathologist trained on biopsy specimens to annotate whole-mount prostatectomy specimens. Future work will focus on prospective multi-institution validation, extending the approach to larger and more diverse patient populations, and developing end-to-end workflows that directly incorporate digital pathologist labeling into clinical MRI interpretation pipelines.