A Magnification Alignment Framework Enables Computation- and Communication-Efficient Computational Pathology.

Area of Science:

• Digital Pathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Current computational pathology (CPath) models struggle with diagnostic efficiency due to high-magnification whole-slide image analysis.
• This reliance on high-magnification data limits clinical utility, especially in time-sensitive scenarios and for data transfer/storage.

Purpose of the Study:

• To develop a computation- and communication-efficient framework for CPath that overcomes the limitations of high-magnification analysis.
• To enable effective CPath analysis using low-magnification inputs, reducing computational time and data requirements.

Main Methods:

• Developed Magnification-Aligned Global-Local Transformer (MAG-GLTrans) framework.
• Employed a magnification alignment (MAG) mechanism using self-supervised learning to align low- and high-magnification feature representations.
• Evaluated MAG-GLTrans on fundamental CPath tasks and a real-world telepathology application.

Main Results:

• MAG-GLTrans achieved state-of-the-art classification performance with significant efficiency gains.
• Demonstrated up to 10.7x reduction in computational time and over 20x reduction in file transfer/storage requirements.
• Successfully applied to intraoperative frozen section diagnosis of non-small cell lung carcinoma, recognizing tumor morphologies and localizing relevant regions.

Conclusions:

• MAG-GLTrans offers a promising solution for time-sensitive CPath applications, particularly intraoperative frozen section diagnosis.
• The framework enhances efficiency without compromising accuracy, making advanced CPath more accessible.
• MAG-GLTrans is versatile, functioning as a feature extractor and enhancing existing foundation models for low-magnification analysis.