Self-Supervised Pathology Foundation Models based Feature Encoder and Multiple Instance Learning for Breast and Skin Cancer Diagnosis

Foundation models (FMs) are an artificial intelligence paradigm that gives transferable and stable feature extraction for hard medical image tasks. In this study, we investigate the transferability of five leading-performing pathology FMs UNI, CONCH, ExaonePath, Lunit, and Prism, pretrained on large histopathology to two domains: skin cancer classification with the HAM10000 dataset and breast cancer detection using the PatchCamelyon (PCam) dataset. We have taken the features from the images using all five models in skin cancer and compared their performances with MLP and the 1D CNN classifiers. The UNI model achieved 95.01% precision and AUC of 97.09% with MLP with no domain difference between pathology and dermoscopy. For breast cancer, UNI, CONCH, and Vision Transformer (ViT) baseline features were extracted, and common machine learning classifiers and Multiple Instance Learning (MIL) models were trained. The MIL model with UNI features had a testing accuracy of 95.70% and an AUC of 98.97%, reflecting improved discriminability. Our findings indicate that foundation pathology models can generalize enormously to new domains beyond their original training domains, with high cross-domain utility for cancer diagnosis. Our research reports the feasibility of foundation models as universal feature encoders in medical imaging to facilitate scalable, efficient, and accurate AI-based solutions to digital pathology and skin cancer.