Prostate cancer is the second most frequently diagnosed cancer in men worldwide, with over 1.4 million new cases recorded in 2020. The standard pathway for confirming a prostate cancer diagnosis is a 12-core needle biopsy, in which 12 tissue samples are taken from different regions of the prostate and submitted for microscopic examination by a pathologist.
Each biopsy case therefore generates 12 separate tissue cores for review. Pathologists must examine each core under the microscope, a time-consuming process that demands sustained concentration. The volume of prostate biopsies performed each year -- driven by widespread PSA screening -- creates a substantial diagnostic workload, increasing the risk of errors due to fatigue and limited specialist availability.
Diagnosing prostate adenocarcinoma is not always straightforward. While the majority of cases are clear-cut, separating well-differentiated adenocarcinoma from benign mimics such as adenosis, atrophy, and atypical hyperplasia, and distinguishing poorly differentiated adenocarcinoma from inflammatory infiltrates, requires considerable expertise. Small cancer foci in a core can be easily overlooked during rapid review.
Computational pathology -- the application of deep learning to digitized tissue slides, known as whole-slide images (WSIs) -- has the potential to screen large volumes of biopsy specimens quickly, flag suspicious areas, and serve as a second check to reduce missed diagnoses. This study developed and evaluated a deep learning model specifically for prostate adenocarcinoma classification in needle biopsy WSIs.
The study used a retrospective dataset of 2,926 hematoxylin and eosin (H&E)-stained WSIs from five hospitals in the Kamachi Group (Fukuoka, Japan): 1,682 needle biopsy and 1,244 transurethral resection of the prostate (TUR-P) specimens. Cases were randomly selected to reflect real clinical variety. At least two pathologists reviewed each case, with final verification by a senior pathologist. All slides were scanned at 20x magnification using the same Leica Aperio AT2 scanner.
The training and validation sets used 1,122 needle biopsy WSIs (438 adenocarcinoma, 684 benign) from five hospitals. Importantly, training used only weakly supervised learning -- the model received only slide-level labels (cancer or benign) without any annotation of specific cancer regions. This reflects a realistic clinical data availability scenario where expert annotation of every cancer focus within a slide would be impractical at scale.
Three deep learning models were compared. The best-performing model used transfer learning from an existing colon poorly differentiated adenocarcinoma model rather than a general image classification network. The premise was that adenocarcinomas from different organs share underlying morphological features -- irregular gland formation, nuclear atypia -- that a model trained on one organ type may partially recognize in another.
During training, the CNN was applied across each WSI using a sliding window tiling approach, generating predictions for each tile. A hard mining strategy was used: once initial training stabilized, the model iteratively identified the tiles it was most uncertain about and prioritized those for subsequent training, forcing improvement on the most challenging examples. The final WSI-level prediction used the maximum probability tile across the whole slide.
The best model -- the transfer learning model derived from the colon adenocarcinoma classifier -- achieved ROC-AUC values of 0.967 to 0.978 across three independent hospital needle biopsy test sets, and an AUC of 0.987 on the public TCGA prostate cancer dataset. These performance levels represent near-expert diagnostic accuracy for distinguishing adenocarcinoma from benign tissue on biopsy slides.
On needle biopsy test sets, the model achieved accuracy of 91.8 to 94.5%, sensitivity of 89.0 to 93.0%, and specificity of 92.4 to 98.0% across the three test hospitals. The F1-scores ranged from 0.918 to 0.944. Performance was consistent across independent hospitals, suggesting the model generalizes across different clinical populations and staining practices within the same scanner type.
Performance was notably lower on TUR-P specimens (AUC 0.737 to 0.909), with sensitivity dropping to 41.5 to 79.8% depending on the hospital. TUR-P samples the inner transition zone of the prostate, which contains more benign prostatic hyperplasia than peripheral zone needle biopsy tissue. The model had been trained exclusively on needle biopsy slides and was applied to TUR-P without any additional fine-tuning, explaining the performance gap.
The model also outperformed all ten pre-existing cross-organ models tested directly on prostate biopsy slides without fine-tuning. The colon poorly differentiated adenocarcinoma model achieved the highest AUC (0.817) among these unadapted models, suggesting meaningful morphological similarity between poorly differentiated colon cancer and prostate adenocarcinoma -- which informed the transfer learning strategy.
True positive predictions were visualized using heat maps overlaid on the WSI, with red indicating high cancer probability and blue indicating low probability. The model successfully highlighted the correct adenocarcinoma-infiltrating tissue cores in multi-core biopsy slides while generating true negative predictions for clearly benign cores -- demonstrating that the model localized cancer regions spatially within the slide despite never being trained with spatial annotations.
The most common cause of false positive predictions was chronic inflammatory cell infiltration (histiocytes, lymphocytes, and plasma cells), which can morphologically resemble adenocarcinoma cells in their density and tissue infiltration pattern. Prostatic hyperplasia with irregularly shaped tubular structures was the second major cause of false positives -- a recognized mimicker of prostate cancer for human pathologists as well.
The most notable cause of false negative predictions in the illustrated case was small or atrophic adenocarcinoma foci that lacked the dense cellularity the model learned to associate with cancer. Prostate adenocarcinoma with atrophic features is a known diagnostic pitfall for both human pathologists and AI systems, as its architectural subtlety mimics benign acinar atrophy.
On TUR-P specimens, an additional source of false positives was unique to the specimen type: thermal cauterization artifacts at tissue margins from the surgical electrocautery instrument. These degenerated marginal zones produced tissue appearances the model misclassified as adenocarcinoma -- a confound that does not exist in needle biopsy specimens and that would need to be addressed with TUR-P-specific training data.
This model was trained on only 1,122 WSIs without any manual region-level annotations -- substantially fewer than prior studies that required approximately 8,400 annotated slides to achieve comparable performance. The weakly supervised transfer learning approach thus dramatically lowers the barrier to developing high-performance diagnostic AI, as it does not require the costly and time-consuming expert annotation process.
The model's intended clinical role is as a screening and double-check tool, not a replacement for pathologist review. It could process digitized biopsy slides automatically and flag likely-positive cores for priority review, effectively directing a pathologist's attention to the most suspicious tissue regions. This workflow integration could reduce the risk of missed cancers in high-volume settings.
The authors note that the model would need to be trained on TUR-P-specific examples before it could be reliably applied to that specimen type. More broadly, models trained on slides from one scanner or staining laboratory may require re-validation or fine-tuning when deployed at institutions using different equipment. The consistency of performance across three independent hospitals using the same scanner in this study is encouraging but does not guarantee generalization to different imaging platforms.
Future development should include adding challenging cases -- atrophic adenocarcinoma, mimickers, and rare subtypes -- to the training set to address remaining error modes. Integration with Gleason grading pipelines (which assess cancer aggressiveness) would be the natural next step beyond binary cancer detection, moving toward a fully automated preliminary pathology workflow for prostate biopsy specimens.
This study demonstrates that transfer learning from a colon cancer classifier can produce a high-performance prostate adenocarcinoma detection model with limited prostate-specific training data. The best model achieved AUCs above 0.96 on needle biopsy slides across three independent hospitals and 0.987 on the public TCGA dataset, establishing strong generalizability for this specimen type.
The major practical contribution of this work is the demonstration that weakly supervised learning -- using only slide-level labels without detailed region annotations -- can achieve near-expert detection accuracy. This significantly reduces the expert time required to build training datasets, making the approach scalable to larger and more diverse datasets in future work.
The authors conclude that while further clinical validation is required before deployment in real diagnostic practice, these deep learning models show strong promise as screening aids that could help manage the growing workload of prostate pathology, reduce the risk of overlooked cancer foci due to fatigue, and ultimately support more consistent and efficient cancer diagnosis.