Attention-based multi-residual network for lung segmentation in diseased lungs with custom data augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models struggle with lung segmentation in chest X-rays (CXRs) due to variations from disease and imaging settings.
• Performance degrades with complex abnormalities like pulmonary opacifications, limiting diagnostic accuracy.

Purpose of the Study:

• To develop a robust deep learning framework for accurate lung segmentation in CXRs, addressing challenges posed by severe lung abnormalities.
• To enhance segmentation performance using attention mechanisms, multi-residual blocks, and a novel data augmentation technique.

Main Methods:

• Proposed AMRU++ (attention-based multi-residual UNet++) network integrating attention modules and multi-residual blocks.
• Introduced a data augmentation strategy simulating CXR pathology features to overcome limited annotated data.
• Validated the framework on diverse datasets including pneumoconiosis, COVID-19, and tuberculosis (350 cases).

Main Results:

• AMRU++ demonstrated robust and accurate lung segmentation capabilities across normal and severely abnormal CXR images.
• The proposed data augmentation technique effectively addressed limitations of scarce annotated data.
• Experimental validation confirmed the framework's effectiveness on public and private datasets.

Conclusions:

• The AMRU++ network offers a significant advancement in automated lung segmentation for complex CXR abnormalities.
• The combined approach of advanced deep learning architecture and tailored data augmentation enhances diagnostic potential in thoracic imaging.