Generative-based airway and vessel morphology quantification on chest CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Accurate characterization of small pulmonary structures (airways, vessels) in CT images is crucial for diagnosing lung diseases.
• Traditional methods for measuring these structures are often limited by image resolution and artifacts.
• Airway morphology changes are linked to airflow resistance in Chronic Obstructive Pulmonary Disease (COPD), while vessel size is important for identifying vascular changes.

Purpose of the Study:

• To develop and validate a Convolutional Neural Regressor (CNR) for precise cross-sectional measurements of airway lumen, airway wall thickness, and vessel radius from CT images.
• To address limitations of traditional methods in characterizing small pulmonary structures.

Main Methods:

• A Convolutional Neural Regressor (CNR) was developed.
• CNR was trained using synthetic airway and vessel data generated by a combination of a generative model and Simulated and Unsupervised Generative Adversarial Network (SimGAN).
• Validation involved comparing CNR measurements against traditional methods using synthetic data and assessing in-vivo correlations with physiological lung function parameters (FEV1%, Pi10 for airways; small-vessel blood volume, DLCO for vessels).

Main Results:

• The proposed generative model and SimGAN successfully created synthetic airways and vessels with ground-truth for training and validation.
• CNR demonstrated accuracy in measuring airway and vessel dimensions compared to traditional methods.
• In-vivo validation showed significant physiological correlates: CNR-derived airway measurements correlated with lung function parameters (FEV1%, Pi10), and vessel measurements correlated with diffusing capacity (DLCO).

Conclusions:

• Convolutional Neural Networks (CNNs), specifically the developed CNR, offer a promising approach for accurate quantification of airways and vessels in chest CT images.
• The method provides measurements with significant physiological relevance, supporting its potential clinical utility in diagnosing and monitoring pulmonary diseases.