CT Image Segmentation for Inflamed and Fibrotic Lungs Using a Multi-Resolution Convolutional Neural Network.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Quantitative CT Analysis

Background:

• Accurate lung segmentation is crucial for quantitative analysis of computed tomography (CT) images, especially in patients with lung abnormalities.
• Existing segmentation methods often struggle with diverse pathologies like consolidation and opacification, hindering rapid analysis.
• Developing a robust, automated algorithm is essential for efficient clinical workflows and research.

Approach:

• A polymorphic training approach was developed, incorporating labeled human (COPD) and animal (acute lung injury) lung data into a single neural network.
• The neural network was trained to predict left and right lung regions, robust to various density-enhancing abnormalities.
• LobeNet algorithm and hierarchical clustering were used for lobar segmentation and radiographical subtype identification, respectively.

Key Points:

• The algorithm achieved high performance (Dice: 0.985 ± 0.011, SSD: 0.495 ± 0.309 mm) on diverse CT scans (COPD, COVID-19, lung cancer, IPF), despite limited labeled data for some conditions.
• Polymorphic training enabled accurate segmentation of COVID-19 cases with diffuse consolidation without requiring specific COVID-19 training data.
• Hierarchical clustering identified four COVID-19 radiographical phenotypes based on lobar consolidation and aeration, with lower lobes frequently affected.

Conclusions:

• The proposed automated lung segmentation algorithm is robust and facilitates rapid quantitative analysis of CT images with lung abnormalities.
• The polymorphic training strategy enhances algorithm generalizability across different lung diseases.
• This approach aids in identifying distinct radiographical phenotypes of diseases like COVID-19, potentially improving diagnostic and prognostic insights.