Cross-domain visual prompting with spatial proximity knowledge distillation for histological image classification.

Area of Science:

• Computational pathology
• Digital histopathology
• Machine learning in medicine

Background:

• Histological classification is complex due to tissue variations and blurry edges.
• Large deep learning models achieve good performance but require significant resources and data.
• Knowledge Distillation (KD) can enable smaller models to mimic larger ones, but struggles with high-dimensional features and edge relationships.

Purpose of the Study:

• To address the limitations of current knowledge distillation methods in histological classification.
• To develop a computationally efficient approach for accurate histological analysis.
• To improve the practical application of AI in histopathology.

Main Methods:

• Proposed a novel cross-domain visual prompting distillation approach.
• Utilized a teacher network to extract high-dimensional features into low-dimensional maps for the student network.
• Introduced a dynamic learnable temperature module with vector-based spatial proximity for enhanced student imitation.

Main Results:

• Demonstrated effectiveness on histological datasets (NCT-CRC-HE-100K, LC25000) and a dermoscopic dataset (ISIC-2019).
• Achieved superior performance and robustness compared to state-of-the-art knowledge distillation methods.
• Showcased optimal domain adaptation capabilities.

Conclusions:

• Introduced VPSP, a novel distillation architecture specifically for histological classification.
• VPSP achieves superior performance and optimal domain adaptation, enhancing clinical utility.
• The source code will be publicly released to facilitate further research and application.