Multi-scale deformable attention fusion network with global context modeling for chest X-ray lesion segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate chest X-ray segmentation is vital for diagnosing thoracic diseases like pneumonia and COVID-19.
• Challenges include blurry lesion boundaries and irregular contours, hindering precise segmentation.
• Existing methods struggle with diverse pathological tissue shapes and sizes.

Purpose of the Study:

• To introduce the Multi-Scale Deformable Attention Fusion Network (MSDAFNet) for improved chest X-ray lesion segmentation.
• To enhance the model's ability to adaptively perceive and segment lesions of varying shapes and sizes.
• To improve robustness in segmenting diverse pathological tissues in chest X-rays.

Main Methods:

• Developed a Multi-Scale Deformable Convolution Attention module using channel grouping and dilated convolutions for multi-scale feature extraction.
• Incorporated learnable offsets in convolution kernels for adaptive sampling position adjustment.
• Introduced a memory module with learnable key-value pairs and attention mechanisms to model long-range dependencies for small lesion regions.

Main Results:

• MSDAFNet demonstrated superior accuracy and robustness in anatomical structure segmentation compared to existing methods.
• The model effectively handles variations in lesion shapes and sizes, improving segmentation performance.
• Experiments on QaTa-COV19-v1 and QaTa-COV19-v2 datasets confirmed the network's competitive performance.

Conclusions:

• MSDAFNet offers a significant advancement in chest X-ray segmentation, particularly for challenging cases with irregular lesion boundaries.
• The proposed network architecture enhances the segmentation of diverse pathological tissues, aiding clinical diagnosis.
• MSDAFNet provides a robust and accurate tool for quantitative analysis in thoracic disease screening.