CT image segmentation for inflamed and fibrotic lungs using a multi-resolution convolutional neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Accurate quantitative analysis of lung abnormalities in computed tomography (CT) images is crucial for diagnosis and treatment monitoring.
• Existing lung segmentation algorithms often struggle with diverse pathological conditions, including diffuse opacification and consolidation.

Purpose of the Study:

• To develop a fully-automated lung segmentation algorithm robust to various density-enhancing lung abnormalities.
• To enable rapid quantitative analysis of CT images for lung diseases.

Main Methods:

• A polymorphic training approach was employed, incorporating labeled human COPD and animal acute lung injury data to train a single neural network.
• The algorithm was evaluated on CT scans from patients with COPD, COVID-19, lung cancer, and idiopathic pulmonary fibrosis (IPF).
• Lobar segmentation and regional analysis using hierarchical clustering were performed to identify COVID-19 radiographic subtypes.

Main Results:

• The algorithm achieved high performance in segmenting lungs across various diseases, including COVID-19 with diffuse consolidation, without requiring specific training data for these conditions.
• Quantitative evaluation on 87 COVID-19 CT images yielded an average symmetric surface distance of [Formula: see text] mm and a Dice coefficient of [Formula: see text].
• Hierarchical clustering identified four distinct radiographical phenotypes of COVID-19, with lower lobes frequently affected by consolidation and poor aeration.

Conclusions:

• The proposed polymorphic training approach enables robust and accurate automated lung segmentation, significantly aiding quantitative analysis in diverse lung pathologies.
• The algorithm's ability to segment unseen diseases like COVID-19 highlights its generalizability and clinical utility.
• Radiographic subtyping of COVID-19 revealed distinct patterns of lobar involvement, offering potential insights into disease severity and progression.