A Lightweight Multi-Cancer Tumor Localization Framework for Deployable Digital Pathology

A team of researchers, including Brian Isett and Riyue Bao, has published a new paper on arXiv detailing MuCTaL (Multi-Cancer Tumor Localization), a lightweight AI framework designed to identify tumor regions in whole-slide pathology images. The core challenge addressed is that AI models trained on specific cancers often fail to generalize to other tumor types. To solve this, the team used balanced training across four different cancers—melanoma, hepatocellular carcinoma, colorectal cancer, and non-small cell lung cancer—leveraging a dataset of 79,984 non-overlapping image tiles. Using transfer learning with the DenseNet169 architecture, the model achieved a high tile-level ROC-AUC (Receiver Operating Characteristic - Area Under the Curve) score of 0.97 on validation data from those four cancers.

Crucially, the model demonstrated its ability to generalize by achieving a score of 0.71 on an independent cohort of pancreatic ductal adenocarcinoma, a cancer type it was not explicitly trained on. The researchers built a scalable inference workflow that generates spatial tumor probability heatmaps, which are compatible with existing digital pathology tools for further analysis. The code and models are publicly available, promoting reproducibility and deployment in research settings where computational resources may be limited. This work represents a significant step toward more robust and versatile AI assistants for pathologists.