Prostate Cancer Histopathology Classification Using Multi-Instance Learning and Vision Transformers

Prostate cancer is the most frequently diagnosed malignancy among men worldwide and remains a major public health concern. Although histopathological evaluation of prostate biopsies is the diagnostic gold standard, it faces limitations such as sampling errors, inter-observer variability, and limited access to specialized pathologists, underscoring the need for computational support. In this study, we propose a computer vision framework for automated classification of prostate histopathology using a multiple instance learning (MIL) approach built upon DINOv2-based foundation model embeddings. The system was trained and validated on a large dataset comprising prostate tissue whole-slide images from TCGA, GTEx, and a private Brazilian laboratory (ARGOS Patologia). Data were divided into training, testing, and an independent validation set, with patientlevel separation to prevent data leakage across subsets. The proposed model achieved an accuracy and sensitivity of 94% and an area under the receiver operating characteristic curve (AUC) of 98% on the independent validation set, demonstrating robust generalization across heterogeneous real-world samples. These results highlight the potential of MIL combined with foundation model representations to support reliable and scalable prostate cancer diagnosis.