Spatially Aware GCNs for efficient, high-accuracy cancer grading: Mitigating oversmoothing via frequency analysis

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Graph-based machine learning

Background:

• Cancer grading is crucial for treatment but relies on complex spatial relationships difficult for standard GCNs.
• Deeper GCNs often suffer from oversmoothing, hindering the capture of intricate cellular structures in histopathology.

Purpose of the Study:

• To develop a Spatially Aware Graph Convolutional Network (SA-GCN) for improved classification of colorectal and non-small cell lung cancer grades.
• To address the oversmoothing problem in GCNs for better preservation of high-resolution spatial and frequency information in histopathological data.

Main Methods:

• Introduced SA-GCN with a dilated layer and quantile-based aggregation to balance low- and high-frequency graph data.
• Optimized graph construction to reduce computational complexity from O(N^2) to O(N).
• Conducted experiments on colorectal and non-small cell lung cancer datasets.

Main Results:

• Achieved accuracy improvements of 0.87% and 0.81% for colorectal and non-small cell lung cancer datasets, respectively.
• Demonstrated 98.05% ± 0.99% accuracy with a seven-layer dilated SA-GCN on the colorectal dataset.
• Showcased significant computational efficiency and guaranteed preserved graph signal diversity.

Conclusions:

• SA-GCN effectively classifies cancer grades by preserving spatial and frequency information, outperforming state-of-the-art methods.
• The proposed architecture enables deeper GCNs and scales to large datasets, addressing limitations of existing techniques.
• SA-GCN offers a robust solution for computational pathology, enhancing diagnostic accuracy and efficiency.