GCUNet: a graph neural network-based contextual learning network for tertiary lymphoid structure semantic...

Area of Science:

• Computational pathology
• Digital pathology
• Immunohistochemistry analysis

Background:

• Tertiary lymphoid structures (TLS) are crucial in immune responses and cancer prognosis.
• Accurate TLS segmentation in whole slide images (WSIs) is challenging due to image scale and the need for contextual information.
• Current patch-based segmentation methods for TLS limit performance by excluding external contextual data.

Purpose of the Study:

• To develop an advanced method for TLS semantic segmentation in WSIs.
• To address the limitations of patch-based approaches by incorporating long-range contextual information.
• To improve the accuracy and detail of TLS boundary and maturity identification.

Main Methods:

• Proposed GCUNet, a graph neural network-based contextual learning network for TLS semantic segmentation.
• Developed a detail and context fusion block (DCFusion) to integrate target patch details with aggregated external contexts.
• Created and utilized four TLS semantic segmentation datasets (TCGA-COAD, TCGA-LUSC, TCGA-BLCA, PUMCH-PAAD), with three datasets to be publicly released.

Main Results:

• GCUNet demonstrated superior performance in TLS semantic segmentation compared to state-of-the-art methods.
• Achieved a minimum mean F1-score (mF1) improvement of 7.41% over the best baseline.
• Experiments confirmed GCUNet's effectiveness across multiple datasets.

Conclusions:

• GCUNet significantly enhances TLS semantic segmentation accuracy in WSIs by effectively utilizing contextual information.
• The proposed method shows strong potential for accurate TLS assessment in computational pathology.
• This work facilitates advancements in analyzing the immune microenvironment through computational pathology.