Local-to-global spatial learning for whole-slide image representation and classification.

Area of Science:

• Digital Pathology
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Whole-slide images (WSIs) are crucial for clinical diagnosis.
• Multi-instance learning (MIL) is used for WSI classification, but struggles with instance correlation.
• Existing MIL methods often have moderate performance due to their instance-level classification strategies.

Purpose of the Study:

• To propose a novel method for improved WSI classification by enhancing WSI-level representation.
• To develop a local-to-global spatial learning strategy for mining both positional and morphological information within WSIs.
• To introduce Global-Local Attentional Multi-Instance Learning (GLAMIL) for better correlation mining between image instances.

Main Methods:

• Introduced Global-Local Attentional Multi-Instance Learning (GLAMIL), a novel MIL-based WSI classification strategy.
• Implemented a local-to-global approach to aggregate region correlations and model positional relationships between regions.
• Utilized Transformer layers to effectively model global and local spatial information, improving upon standard feature extraction.

Main Results:

• GLAMIL demonstrated superior performance in learning WSI-level representations by focusing on regional relationships.
• The method successfully aggregated patch relationships within local pools to capture tissue type correlations.
• Evaluated on three benchmarks, GLAMIL achieved satisfactory results, outperforming state-of-the-art methods and baselines by approximately 1% and 10%, respectively.

Conclusions:

• GLAMIL offers a more effective strategy for WSI classification by capturing complex spatial relationships.
• The proposed method significantly improves upon existing MIL-based approaches for WSI analysis.
• GLAMIL's ability to mine global and local spatial information enhances diagnostic accuracy in digital pathology.