Disease Classification of Pulmonary Xenon Ventilation MRI Using Artificial Intelligence

Area of Science:

• Pulmonary imaging
• Artificial intelligence in medicine
• Medical diagnostics

Background:

• Hyperpolarized 129Xenon MRI quantifies lung ventilation via ventilation defect percent (VDP).
• VDP alone cannot differentiate between various lung diseases.
• Prior studies noted anecdotal disease-specific patterns in xenon MRI, but these were not systematically analyzed.

Purpose of the Study:

• To investigate if artificial intelligence (AI) using convolutional neural networks (CNNs) can classify lung diseases based solely on spatial patterns in xenon MRI.
• To determine if AI can identify disease-specific patterns in xenon ventilation images.

Main Methods:

• Xenon MRI data from 262 participants across six conditions (control, asthma, bronchiolitis obliterans syndrome, bronchopulmonary dysplasia, cystic fibrosis, LAM) were used.
• Convolutional neural networks, including VGG-16, were trained to classify diseases based on image patterns.
• Network performance was evaluated using accuracy, recall, precision, and AUC; Grad-CAM visualized classification drivers.

Main Results:

• The top-performing VGG-16 network achieved 56% top-1 and 78% top-2 accuracy.
• The AI model demonstrated higher accuracy on larger patient cohorts (e.g., control, cystic fibrosis, LAM).
• The AI network outperformed human observers, achieving 61% top-1 accuracy compared to 40% for humans on a specific test set.

Conclusions:

• An AI tool was developed capable of classifying lung diseases using only xenon ventilation MRI images.
• The AI tool demonstrated superior performance compared to human interpretation, suggesting inherent disease-specific information in the images.
• Xenon MRI, analyzed by AI, holds potential for challenging clinical cases and disease phenotyping.