Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiology research

Background:

• Automated lung segmentation in computed tomography (CT) is vital for image analysis.
• Existing segmentation methods have limited clinical applicability across diverse diseases.
• Sophisticated pipelines often require extensive training and validation on specific datasets.

Purpose of the Study:

• To compare the performance of generic deep learning approaches and established lung segmentation algorithms.
• To evaluate the impact of training data diversity on segmentation accuracy across multiple lung diseases.
• To identify the most effective strategies for robust lung segmentation in clinical settings.

Main Methods:

• Four deep learning models and two existing lung segmentation algorithms were evaluated.
• Performance was assessed on routine CT imaging data encompassing over six disease patterns.
• Comparison included training on diverse routine datasets versus public datasets (Lung Tissue Research Consortium, Anatomy 3).

Main Results:

• Deep learning models showed minimal variation in Dice Similarity Coefficients (DSCs) (≤0.02) when trained on different datasets.
• A U-net model trained on a diverse routine dataset (n=36) achieved a higher DSC (0.97 ± 0.05) than when trained on public datasets.
• U-net trained on extensive routine data (n=231) yielded superior DSC (0.98 ± 0.03) compared to reference methods (0.94 ± 0.12).

Conclusions:

• Training data diversity is more critical than model choice for accurate lung segmentation in challenging cases.
• Developing new, diverse datasets and sharing trained models are essential for advancing lung disease research.
• A publicly released, trained segmentation model (General Public License 3.0) can facilitate research on pathological lungs.