Workshop: The 12th Computational Approaches for Cancer Workshop (CAFCW26)
Authors: Subhankar Chattoraj, Karuna Pande Joshi, and Snigdhansu Chatterjee (University of Maryland Baltimore County)
Abstract: Multi-view histopathology provides multiple histopathological images (HI) at different magnifications and tissue regions for each patient. Existing methodology includes complex fusion based learning architecture. However, existing state-of-the-art deep fusion models fails to capture the acquisition geometry of the multi-view HI. On the recently released Multi-Oral Squamous Cell Carcinoma (OSCC) benchmark, a self-attention fusion module was reported to under-perform plain feature concatenation, suggesting that unstructured attention cannot recover the scale-and-region relationships from a small cohort of 925 training patients.
In this work, we proposed dot-product based Geometry-Routed Attention fusion model (GRA), that incorporates a low-parameter routing that considers the acquisition geometry along with tumor core/edge regions prior directly into the attention logits. We study two realizations, ALiBi-style bias (GRA-P) and a learned per-relation routing (GRA-R). We validate our proposed methodology on Multi-OSCC dataset comprising of six classes with HI captured under three magnification factors. GRA-R attains a mean test AUC of 77.97%, with the highest class specific AUC for recurrence of 91.27%. Our results surpass existing benchmark Concat by 2.97% on mean multi-class performance. In particular, GRA-R surpass existing state of the art by 4.89% for lymph node metastasis and 5.04% for tumor differentiation. Additionally, the data-scaling ablation demonstrates that the GRA-R largest performance yield in low-data regimes, highlighting the utility of the proposed model under limited training data. Component-wise ablations further demonstrate that both acquisition-geometry and region-aware routing terms contribute independently to the overall performance.
Back to The 12th Computational Approaches for Cancer Workshop (CAFCW26) Archive Listing
Back to Full Workshop Archive Listing