Pan-Cancer Histopathology Foundation Models Can Diagnose Oral Lesions Robustly Without Fine-Tuning

Oral Squamous Cell Carcinoma (OSCC) is a common form of cancer worldwide, yet it remains underrepresented in public imaging datasets, as it is more common in the global south. On the other hand, there exist extensive repositories for skin cancers that are prevalent in the global north. This bias bleeds into the development of cancer AI models (like large Vision Transformer (ViT)–based backbones) since they are data hungry. In this work, we demonstrate that a pretrained ViT based digital pathology foundation model, trained on pan-cancer tissue types, can serve as a powerful frozen feature embedder for OSCC histopathological images without any further fine-tuning. Training only a shallow multilayer perceptron (MLP) as a downstream classifier on the ViT embeddings, our approach achieves an accuracy of 95.24% and an F1 score of 95.29% on a publicly available OSCC histopathological images dataset, comparable to fully trained convolutional neural network (CNN) models which are much slower to train from scratch. This establishes the blueprint for a path for foundation model based oral cancer detection without the need for large domain-specific datasets or expensive training pipelines, especially in data and computation constrained clinical environments in S/S-E Asia where oral cancer is more prevalent.