Human visual attention-inspired knowledge distillation underlying interpretable computational pathology.

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in healthcare

Background:

• Computational pathology utilizes deep learning for high-resolution medical image analysis.
• A key challenge is balancing model efficiency, interpretability, and performance.
• Knowledge distillation (KD) compresses models but often lacks interpretability, leading to inaccurate image attention.

Purpose of the Study:

• To develop a novel knowledge distillation strategy inspired by human visual processing.
• To improve the interpretability and performance of lightweight deep learning models in computational pathology.
• To address the trade-off between model efficiency and accuracy in pathological image analysis.

Main Methods:

• Developed a human vision attention-inspired knowledge distillation (HVisKD) strategy.
• HVisKD captures local and global patch relationships for differentiated feature construction.
• Applied HVisKD to pathological analysis, focusing on segmentation tasks and intraoperative diagnosis.

Main Results:

• HVisKD enhanced performance across various lightweight models in segmentation tasks.
• Attention maps generated by HVisKD demonstrated improved consistency with human expert segmentations.
• HVisKD provided interpretable and rapid analysis in a real-world intraoperative pathological diagnosis setting.

Conclusions:

• HVisKD offers a lightweight and interpretable strategy for computational pathology.
• The approach aligns deep learning with brain-like information processing for reliable outputs.
• This method effectively balances model efficiency, interpretability, and task performance.