MSCPT: Few-Shot Whole Slide Image Classification With Multi-Scale and Context-Focused Prompt Tuning.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Multiple instance learning (MIL) is standard for whole slide image (WSI) classification but requires extensive labeled data.
• Few-shot Weakly Supervised WSI Classification (FSWC) faces challenges due to limited data and rare diseases.
• Existing prompt tuning methods for natural images are suboptimal for WSIs, failing to utilize VLM text priors and WSI multi-scale/contextual information.

Purpose of the Study:

• To develop an advanced prompt tuning method for Few-shot Weakly Supervised WSI Classification (FSWC).
• To enhance the utilization of Vision-Language Model (VLM) prior knowledge and WSI-specific features.
• To overcome limitations in current prompt tuning approaches for whole slide images.

Main Methods:

• Proposed Multi-Scale and Context-focused Prompt Tuning (MSCPT) method for FSWC.
• Utilized frozen large language models to generate multi-scale pathological visual language prior knowledge for hierarchical prompt tuning.
• Incorporated a graph prompt tuning module for contextual information and a non-parametric cross-guided instance aggregation module for WSI-level feature extraction.

Main Results:

• MSCPT demonstrated strong performance across five datasets and three downstream tasks using three different VLMs.
• The method effectively leverages VLM text modality and captures multi-scale/contextual information in WSIs.
• Visualizations and interpretability analyses confirmed the method's efficacy.

Conclusions:

• MSCPT offers a robust solution for Few-shot Weakly Supervised WSI Classification, particularly beneficial for rare diseases.
• The proposed approach enhances prompt tuning by integrating multi-scale, contextual, and instance aggregation strategies.
• The public availability of the code facilitates further research and application in computational pathology.