Title: Pathway-Structured Privileged Distillation for Deployable Computational Pathology

Abstract:

Although combining transcriptomics with histopathology holds significant potential for enhancing cancer risk assessment, its clinical application is often hindered by the scarcity of RNA profiling data in standard practice. To address this, we present Mixture of Pathway Experts (MoPE), a novel framework based on knowledge distillation that transforms multimodal learning into privileged distillation, enabling inference using histology data alone.

The MoPE approach is grounded in the concept of partial observability between whole-slide images (WSIs) and RNA profiles. While histological images can reflect the morphological consequences of specific molecular programs, they are insufficient for reconstructing the complete transcriptomic landscape. MoPE addresses this by encoding pathways derived from RNA data and transferring molecular supervision to pathology experts indexed by pathways, utilizing memory-usage alignment.

Evaluated across various public benchmarks and two distinct breast cancer cohorts, MoPE demonstrated consistent improvements in WSI-only inference performance compared to baseline techniques. Furthermore, analyses of pathway usage and visual inspections conducted by human experts offer a bounded understanding of the model’s behavior and identify potential morphology-linked readouts. These findings suggest that pathway-structured privileged distillation offers a viable strategy for leveraging molecular information during the training phase while maintaining the ability to perform inference without RNA data.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC