Understanding Synergistic Interactions among Pathology Foundation Models via Adaptive Fusion

2026-08-02

AI Summary

This work addresses representational biases arising from differences in training data, architectures, and objectives among existing histopathology foundation models, which compromise robustness and obscure their individual strengths. To overcome this limitation, the authors propose AdaFusion—a lightweight, adaptive fusion framework that dynamically weights the contributions of multiple frozen foundation models through low-dimensional feature compression and a sample-conditioned gating mechanism. AdaFusion further incorporates channel-wise weight adjustment and contribution-driven visualization to enable interpretable model integration. This approach constitutes the first method to achieve explainable, synergistic fusion of multiple pathology foundation models, revealing associations between model-specific preferences and tissue phenotypes. Evaluated on three public histopathology benchmark tasks, AdaFusion consistently outperforms individual models and alternative fusion strategies while producing interpretable outputs aligned with tissue morphology.

Abstract

Pathology foundation models (PFMs) provide strong tile-level representations via self-supervised pre-training on large-scale pathology images. Yet, PFMs are developed under diverse and often opaque data, architecture, and objective choices, inducing latent representational biases that limit robustness and obscure what each model specialises in. We present AdaFusion, a lightweight adaptive fusion framework that integrates complementary signals from multiple frozen PFMs through (1) low-dimensional feature compression and (2) a sample-conditioned gating module that reweights model-wise (and optionally channel-wise) contributions. Beyond improving predictive accuracy, AdaFusion provides contribution-driven interpretation that offers evidence consistent with model-specific preferences and synergistic interactions across tissue phenotypes. We evaluate AdaFusion on three public benchmarks spanning treatment response prediction, prostate cancer grading, and spatial gene expression inference. AdaFusion consistently outperforms individual PFMs and other fusion baselines, while providing interpretable tissue visualisation which aligns model preferences with morphological patterns. Code is available at: https://github.com/xyx-98/PathoOracle.

Problem

• Pathology foundation models
• representational bias
• model robustness
• model specialization
• synergistic interactions

Innovation

• adaptive fusion
• pathology foundation models
• feature compression
• sample-conditioned gating
• interpretable AI