MANet: Multi-branch attention auxiliary learning for lung nodule detection and segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Pulmonary nodule detection and segmentation are crucial for early lung cancer diagnosis via chest CT.
• Distinguishing nodules from background is challenging due to similar biological characteristics and varied sizes.
• Existing methods struggle with false positives and precise segmentation.

Purpose of the Study:

• To propose an automatic method for pulmonary nodule detection and segmentation in chest CT.
• To enhance feature information of pulmonary nodules for improved analysis.
• To reduce false positives and increase segmentation accuracy.

Main Methods:

• Developed MANet, a UNet-based network with a multi-branch attention auxiliary learning mechanism.
• Incorporated novel Projection, Fast Cascading Context, and Boundary Enhancement modules.
• Implemented a Proposal Refinement step to reduce false positives and improve segmentation quality.

Main Results:

• MANet outperformed state-of-the-art methods on LUNA16 and LIDC-IDRI benchmarks.
• Achieved an 88.11% FROC score for nodule detection.
• Reached 71.29% IoU and 82.74% DSC scores for nodule segmentation.
• Ablation studies confirmed the effectiveness of the novel modules.

Conclusions:

• The proposed MANet with multi-branch attention demonstrates a promising approach for pulmonary nodule detection and segmentation.
• The method offers improved performance compared to the original UNet design.
• The novel modules can be integrated into other UNet-based models.