Valuing vicinity: Memory attention framework for context-based semantic segmentation in histopathology.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Accurate segmentation of tumorous and non-tumourous tissue in whole slide images is critical for personalized cancer therapy.
• Existing semantic segmentation methods struggle with high-resolution histopathological images due to limitations in processing global spatial context.
• Identifying tumor subtypes requires understanding subtle spatial relationships, which current methods often overlook.

Purpose of the Study:

• To develop an advanced framework for improved spatial context comprehension in whole slide image segmentation.
• To enhance the precision of tumorous region classification by integrating neighboring tissue information.
• To create a method that mimics a pathologist's approach to analyzing surrounding tissue for accurate diagnosis.

Main Methods:

• Proposed a patch neighbour attention mechanism to query and infuse neighboring tissue context into feature maps.
• Developed a Memory Attention Framework (MAF) that integrates into existing encoder-decoder segmentation models.
• Evaluated the MAF on breast, liver, and kidney cancer datasets using U-Net and DeeplabV3 architectures.

Main Results:

• The MAF demonstrated superior performance compared to other context-integrating algorithms.
• Achieved substantial improvements in Dice score, up to 17%, on public and internal cancer datasets.
• Successfully integrated into established segmentation models, enhancing their contextual understanding.

Conclusions:

• The Memory Attention Framework significantly improves the accuracy of tumor segmentation in histopathological images.
• Incorporating spatial context through the MAF is crucial for precise identification of tumorous regions and subtypes.
• This approach offers a promising advancement for digital pathology and personalized cancer treatment strategies.