SparseXMIL: Leveraging sparse convolutions for context-aware and memory-efficient classification of whole slide...

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Imaging Analysis

Background:

• Whole Slide Images (WSIs) are crucial for pathology diagnosis but present analysis challenges due to large data volumes.
• Existing Multiple Instance Learning (MIL) methods struggle to capture essential spatial context between image tiles.
• Convolutional Neural Network (CNN) adaptations for high-resolution WSIs are often GPU memory-intensive, limiting scalability.

Purpose of the Study:

• To develop a novel framework, SparseXceptionMIL (SparseXMIL), for efficient spatial interaction modeling in WSI analysis.
• To improve GPU efficiency for pathology diagnostics using sparse image representations and novel pooling operators.
• To enhance the scalability and accuracy of WSI analysis in digital pathology.

Main Methods:

• Introduction of a multidimensional sparse image representation for WSI data.
• Development of a novel pooling operator integrating sparse convolutions within the Xception architecture.
• Efficient spatial information modeling at both local and global scales within WSIs.

Main Results:

• SparseXMIL outperforms state-of-the-art MIL methods in WSI classification tasks requiring spatial context.
• The framework offers a superior trade-off in GPU memory requirements compared to traditional CNN-based methods.
• Demonstrated effectiveness in breast and lung carcinoma subtyping and DNA damage response prediction in breast cancer WSIs.

Conclusions:

• Sparse convolutional architectures hold significant potential for efficient and scalable WSI analysis.
• SparseXMIL provides an effective solution for overcoming the limitations of existing WSI analysis techniques.
• The proposed method advances digital pathology by enabling more robust and resource-efficient diagnostic tools.