Foundation model based prostate cancer segmentation on whole mount digitized H&E radical prostatectomy section

Prostate cancer (PCa) diagnosis typically relies on needle core biopsy, which samples only a small portion of the tumor. After biopsy confirmation, radical prostatectomy (RP) may be performed, with whole mount processing of RP specimens allowing for full-gland assessment of PCa extent and distribution. In this study, we refine a pathology foundation model (UNI) for PCa segmentation using only prostate biopsy whole-slide images (WSIs) and evaluate its performance on RP specimens. We use UNI, a state-of-the-art self-supervised model, to extract feature embeddings WSIs. We used the publicly available PANDA dataset as the training set (D1), comprising over 10,000 WSIs from 2,113 patients. Image patches were extracted from WSIs and labeled based on Gleason pattern-annotated regions. Patches for clinically significant PCa (Gleason score ≥3+4) were merged into a single class. A logistic regression classifier was trained on these patch-level embeddings to predict PCa presence. For evaluation (D2), we used an independent test set consisting of expert-annotated whole mount H&E-stained RP WSIs from 48 patients. Our PCa classification model achieved a mean sensitivity of 0.82 and specificity of 0.87 on the training set (D1), and 0.76 and 0.93 on the validation set (D2). Following classification, the identified patches were further processed to generate the predicted segmentation masks.