STAMP-Net: Oral Cancer AI

Early cancer detection in histopathology slides suffers from high misdiagnosis risk, and slide staining varies enough across labs to break models trained on a single source.

Whole-slide pathology images are too large for a CNN to see at once, and pathologists need tools that hold up across labs, not just the lab that produced the training data.

Implemented in PyTorch. Used attention-based MIL so the model learns which regions of a whole-slide image matter without needing patch-level labels, and adversarial training to disentangle stain color from diagnostic signal.

Chose attention-based MIL specifically because it gives interpretability — which region of the slide drove the classification — which matters for medical AI in a way raw accuracy doesn't.

Model interpretability (attention maps) mattered as much to reviewers as raw classification accuracy for a medical-imaging paper.

Extend the model to other carcinoma types and package it as a web-based tool for pathologists.