SLDPC: Slide-Level Dual-Prompt Collaboration for few-shot whole slide image classification.

Area of Science:

• Digital pathology
• Computer-aided diagnosis
• Machine learning in medicine

Background:

• Digital pathology standardizes workflows via slide digitization and algorithmic analysis.
• Few-shot Weakly Supervised Whole Slide Image (FSWC) Classification is a key challenge.
• Conventional methods like Multiple Instance Learning (MIL) require extensive data and struggle with distribution shifts.

Purpose of the Study:

• To address limitations of patch-level feature aggregation in Vision-Language Model (VLM)-based methods for FSWC.
• To propose a novel framework that effectively models crucial slide-level diagnostic information.
• To enhance the performance of FSWC by leveraging slide-level multimodal representations.

Main Methods:

• Introduction of the Slide-Level Dual-Prompt Collaboration (SLDPC) framework.
• Utilizing a slide-level VLM for direct prompt tuning at the slide level.
• Employing bidirectional InfoNCE loss for feature alignment and a weighted fusion mechanism for prompt combination.

Main Results:

• SLDPC demonstrates superior performance across four experimental datasets.
• Slide-level prompt learning effectively tackles FSWC challenges.
• The framework shows significant improvements in model performance for digital pathology tasks.

Conclusions:

• The proposed SLDPC framework offers an effective solution for FSWC in digital pathology.
• Leveraging slide-level information and dual-prompt collaboration enhances diagnostic capabilities.
• This approach advances parameter-efficient few-shot learning for whole slide image analysis.