Dual-channel grouped cross-dimension attention V-Net for pulmonary nodule segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate pulmonary nodule segmentation is vital for early lung cancer diagnosis and treatment planning.
• Challenges include nodule similarity to surrounding tissues and unpredictable variations in location, size, and quantity.
• Existing methods struggle with the nuances of nodule segmentation.

Purpose of the Study:

• To develop an advanced deep learning network for precise pulmonary nodule segmentation.
• To address the limitations of current segmentation techniques.
• To improve early lung cancer detection and patient outcomes.

Main Methods:

• Proposed a dual-channel grouped cross-dimension attention V-Net (DGCA V-Net) model.
• Incorporated global grouped coordinate attention (GGCA) for comprehensive feature capture during downsampling.
• Utilized grouped split attention (GSA) and dual-input guided feature aggregation (DGA) modules to preserve detailed information and refine localization.

Main Results:

• The DGCA V-Net model demonstrated superior performance on the LUNA16 dataset compared to the baseline V-Net.
• Achieved significant improvements in Dice Similarity Coefficient (DSC) (+4.72% to 0.7921), Intersection over Union (IoU) (+6.92% to 0.6662), precision (+2.90% to 0.8102), and recall (+4.80% to 0.7993).

Conclusions:

• The DGCA V-Net model significantly enhances lung nodule segmentation accuracy.
• Ablation tests confirm the model's robust segmentation and generalization capabilities.
• The model shows potential for broader applications in medical image segmentation and is open-source.