Explainable AI-based classification of asthma and COPD from ¹²⁹ Xe ventilation MRI

Background Diagnosis of asthma and/or COPD is challenging due to overlapping clinical and physiological characteristics; however, due to varying treatment recommendations, determining the correct diagnosis is vital to clinical care. Hyperpolarised xenon-129 ventilation ( 129 Xe)-MRI provides high spatial resolution 3D images of gas distribution within the lungs. Recently, AI-based models have successfully classified pulmonary pathologies via medical images, but due to their black-box nature, clinical application is challenging. Here, we aim to develop an explainable AI framework to classify patients with asthma and/or COPD using 129 Xe-MRI alone. Methods 129 Xe-MRI from 207 patients with asthma and/or COPD were classified using a vision transformer (ViT) and convolutional neural network (CNN). Regional visualisations of areas of importance (AOI) associated with each diagnosis were generated. Regional patterns were investigated to develop a shape-based metric which quantified asthma-like and COPD-like features. Results The ViT accurately classified asthma and/or COPD patients in 83% (n=25/30) of cases (AUC=0.86, F 1 =0.82), outperforming the CNN (AUC=0.71, F 1 =0.62). Patients with asthma had more diffuse, elongated AOIs associated with disease grouping while patients with COPD had more focal, spherical AOIs. The proposed sphericity-score weakly correlated with conventional lung function metrics in clinically understandable ways. When applied to patients with overlapping features of asthma and COPD (asthma+COPD), the classification framework informed the prevalence and regional distribution of asthma-like and COPD-like features. Conclusion Explainable AI classification of 129 Xe-MRI could facilitate diagnosis of patients with asthma and/or COPD, including determining the prevalence of asthma-like or COPD-like features in patients with asthma+COPD, potentially informing management and treatment.