Computationally efficient adaptive decompression for whole slide image processing.

Area of Science:

• Digital Pathology
• Computational Imaging
• Machine Learning

Background:

• Whole slide image (WSI) analysis is crucial in modern pathology.
• Deep learning methods excel at WSI tasks but demand substantial computational resources.
• Existing WSI analysis often requires full image decompression, limiting practical application.

Purpose of the Study:

• To develop computation-efficient analysis workflows for WSI classification.
• To adapt state-of-the-art WSI classification models for practical, resource-constrained environments.
• To leverage compression domain processing for faster and less memory-intensive WSI analysis.

Main Methods:

• Utilized pyramidal magnification structure and compression domain features of WSI files.
• Assigned variable decompression depths to WSI patches based on compression features.
• Employed attention-based clustering for low-magnification screening and feature-based selection for high-magnification patches.

Main Results:

• Achieved a 7.2× overall speedup in WSI analysis.
• Reduced memory costs by 1.1 orders of magnitude.
• Maintained model accuracy comparable to the original, non-optimized workflow.

Conclusions:

• Compression domain processing offers a viable solution for efficient WSI analysis.
• The proposed methods significantly reduce computational demands without sacrificing classification accuracy.
• This approach enhances the practical usability of deep learning for WSI classification tasks.