Multi-scale dense selective network based on border modeling for lung nodule segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate quantification of pulmonary nodules is crucial for lung cancer diagnosis and treatment.
• Existing segmentation methods may struggle with the efficiency and accuracy of irregular nodule identification.

Purpose of the Study:

• To improve the segmentation efficiency of irregular pulmonary nodules.
• To maintain high segmentation accuracy for all nodule types.
• To enhance the clinical utility of pulmonary nodule analysis.

Main Methods:

• Proposed BorDenNet, a multi-scale dense selective network incorporating border modeling.
• Utilized a dual-branch encoder-decoder for parallel processing of image and border streams.
• Introduced dense attention and multi-scale selective attention modules for feature enhancement and correlation.
• Incorporated a border context enhancement module for fusing edge-related features.

Main Results:

• Achieved an average Dice score of 92.78% and sensitivity of 91.37% on the LIDC-IDRI dataset.
• Demonstrated excellent generalization on a private hospital dataset.
• Improved segmentation efficiency for complex nodule types like adherent and ground-glass nodules.

Conclusions:

• Accurate segmentation of irregular pulmonary nodules provides vital clinical parameters.
• The proposed method aids clinicians in diagnosis and improves overall clinical efficiency.