Refining weak supervision for robust lung cavity segmentation: A graph-affinity method with boundary constraints.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Pixel-level annotation of lung cavities (LCs) in CT images is difficult due to complex morphology.
• Weakly supervised semantic segmentation (WSSS) uses sparse annotations but often yields imprecise segmentation.
• Existing WSSS methods struggle with under- or over-segmentation of irregular lung lesions.

Purpose of the Study:

• To develop an advanced WSSS method for accurate lung cavity segmentation in CT scans.
• To address limitations of current WSSS approaches in handling complex lesion morphology and spatial supervision.
• To improve the precision and reliability of segmenting irregular lung lesions.

Main Methods:

• Proposed a novel Graph-based Affinity Network (GA-Net) using superpixel graphs and edge inference for structure-aware pseudo-label refinement.
• Implemented region-wise affinity propagation for precise segmentation control within 3D regions.
• Incorporated Exponential Moving Average (EMA) ensembling and a scribble-based module for boundary supervision.

Main Results:

• The proposed GA-Net method significantly outperforms existing state-of-the-art medical WSSS techniques.
• Achieved precise and reliable segmentation of complex lung cavities in CT scans across three benchmark datasets.
• Demonstrated superior performance in refining pseudo-labels and providing effective spatial supervision.

Conclusions:

• The developed GA-Net method offers a robust solution for challenging lung cavity segmentation in medical imaging.
• This approach enhances WSSS by effectively modeling long-range dependencies and improving spatial accuracy.
• The findings suggest a significant advancement in automated analysis of lung pathologies from CT data.