PatHME: Hierarchical Multi-Expert Knowledge Distillation for Whole Slide Image Analysis

Foundation models have shown strong generalization across a variety of pathology whole slide image (WSI) analysis tasks, including classification, segmentation and report generation. However, most existing models are trained exclusively on single-scale, high-magnification patches, often overlooking the rich contextual information available at lower magnifications. Applying such models directly to low-resolution inputs can lead to degraded or misleading feature representations, as crucial spatial context may be lost. Moreover, fine-tuning foundation models for each task is both computationally intensive and inefficient. To address these challenges, we propose Hierarchical Multi-Expert Knowledge Distillation, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PatHME</i>, a unified framework that integrates multi-scale information and leverages multiple expert foundation models for more effective and efficient WSI analysis. First, the Prompt-guided Multi-Scale Distillation module bridges the gap between low- and high-resolution representations by distilling features across magnifications using scale-aware prompts, enabling the model to capture both global tissue architecture and fine-grained cellular detail. We further introduce a Multi-Expert Knowledge Distillation strategy that enables feature-level knowledge transfer from a teacher expert to a student expert without retraining, enhancing adaptability while reducing computational cost. Finally, we incorporate a Hierarchical Multi-Scale Attention mechanism to dynamically fuse multi-resolution features based on tissue context, enabling the model to effectively capture both fine-grained details and global structural patterns. Experimental results on multiple TCGA benchmarks demonstrate that our proposed method significantly enhances model performance and efficiency, providing a scalable and generalizable solution for advanced digital pathology applications.