A multi-stage 3D convolutional neural network algorithm for CT-based lung segment parcellation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Traditional lung parcellation relies on fissures for lobar volume estimation.
• Deep learning offers enhanced assessment of regional ventilation and perfusion heterogeneity.

Purpose of the Study:

• To validate and demonstrate the clinical applicability of deep learning-based CT lung segment parcellation.
• To assess the technique in a clinical cohort with mixed airway diseases.

Main Methods:

• A 3D convolutional neural network was used for airway centerline determination and tertiary bronchi identification.
• End-to-end lung segment parcellation was trained directly from chest CT images.
• Performance was evaluated using Dice score and inclusion rate on training data and qualitatively by radiologists on external validation data.

Main Results:

• Quantitative analysis showed a mean Dice score of 86.81 and inclusion rate of 0.75.
• Qualitative evaluation demonstrated high intra-reader agreement (99.2%).
• Patients with COPD exhibited greater segmentation mismatch compared to healthy individuals.

Conclusions:

• A deep learning algorithm successfully generates lung parcellation masks from CT scans.
• Encouraging quantitative and qualitative results support potential clinical use for pulmonary segment-level lung analysis.
• The method is particularly relevant for patients with structural airway disease.