The Impact of Interstitial Diseases Patterns on Lung CT Segmentation

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Image Segmentation

Background:

• Accurate lung segmentation is crucial for analyzing interstitial lung diseases (ILDs) using Computed Tomography (CT).
• Removing background noise and irrelevant data is essential for efficient and effective ILD analysis.
• Existing models face challenges segmenting lungs with severe disease manifestations.

Purpose of the Study:

• To develop a lightweight and faster U-net based architecture for lung segmentation in CT images.
• To improve the efficiency of computer-aided decision systems for ILD investigation.
• To evaluate the model's performance on heterogeneous datasets, including severe disease patterns.

Main Methods:

• Proposed a novel, lightweight, and faster 2D U-net architecture for lung segmentation.
• Utilized a combination of two publicly available databases to enhance training data heterogeneity.
• Conducted experiments comparing the proposed architecture against the original U-net.

Main Results:

• The proposed architecture achieved comparable performance to the original U-net (DSC: 0.894 ± 0.060, HD: 4.493 ± 0.633, HD-95: 4.457 ± 0.628).
• Demonstrated more efficient computational usage compared to the original U-net.
• Evaluated the model's ability to handle high-density lung patterns associated with severe ILD.

Conclusions:

• The developed lightweight architecture offers an efficient alternative for lung segmentation in ILD analysis.
• Performance is maintained while improving computational efficiency.
• The study underscores the necessity of representative and diverse datasets for robust medical image segmentation tools.