MorphoNet: Morphological sub-region-based structure learning for WSI analysis.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Whole slide image (WSI) representation learning is crucial for computational pathology tasks like tumor subtyping and survival prediction.
• Current patch-centric methods suffer from redundancy and suboptimal spatial modeling, leading to issues in Multiple Instance Learning (MIL), graph-based, and prototype-based models.

Purpose of the Study:

• To introduce MorphoNet, a novel network for learning morphological structures and long-range spatial tissue relationships in WSIs.
• To address the limitations of existing patch-based approaches in computational pathology.

Main Methods:

• MorphoNet utilizes Morphological Sub-Region Grouping (MSRG) to cluster spatially adjacent patches into coherent sub-region embeddings, reducing redundancy.
• A lightweight Graph Neural Network (GNN) processes a sub-region graph to model contextual dependencies and generate slide-level representations.
• MSRG is designed as a plug-and-play module adaptable to existing MIL, graph-based, and prototype-based pipelines.

Main Results:

• MorphoNet demonstrated superior performance across ten public benchmarks for tumor subtyping and survival prediction.
• The MSRG module consistently improved the performance of integrated computational pathology pipelines.
• The method effectively captures long-range spatial dependencies and informative morphological patterns.

Conclusions:

• MorphoNet offers a significant advancement in WSI representation learning for computational pathology.
• The proposed MSRG module enhances the modeling of spatial coherence and tissue morphology.
• MorphoNet provides a robust and adaptable framework for improving downstream tasks in digital pathology.