Ventilation Prediction from IE-CT Using Self-Supervised Learning and a Physics-Informed Swin Transformer Architecture

Introduction: : Pulmonary ventilation imaging enables functional avoidance radiotherapy treatment plans by quantifying regional lung function. However, current clinically available imaging, such as 99𝑚Tc-based single-photon emission computed tomography (SPECT), relies on radioactive gases, which can introduce artifacts and expose patients to additional radiation. Deep learning (DL) models have been proposed to generate ventilation images from non-contrast Computed Tomography (CT). While they achieve accuracy higher than conventional physics-based methods, these models require input from multiple respiratory phases, making them computationally and data expensive. This study presents a DL approach that predicts ventilation images using only inhale and exhale CT (IE-CT) pairs, exploring the feasibility of using breath-hold CTs as input.

Materials and

Methods: : We developed a SwinUNETR-based architecture tailored for Siamese IE-CT inputs, integrating physics-informed modeling with self-supervised learning for lung function prediction. The model was first pre-trained using self-supervised learning on 523 IE-CT scans from nine 4DCT datasets, where it learned a low-dimensional feature space by reconstructing image volumes from randomly augmented data. Supervised training was then performed using transfer learning on a cohort of 44 lung cancer patients, each with paired IE-CT and SPECT ventilation scans. Model performance was evaluated using voxel-wise Spearman correlation and Dice Similarity Coefficient (DSC) and compared against existing state-of-the-art methods.

Results, Conclusions, and Discussions:: Using random bootstrapping, we estimated the mean voxel-wise Spearman correlation between our predictions and the SPECT ground truth to be 0.728 ± 0.028. The mean DSC for the lowest 20% ventilation regions (DSClow) was 0.617 ± 0.028. For the highest 20% ventilation regions, the mean value was 0.596 ± 0.025. Despite the relatively small dataset, our model’s performance was comparable to existing state-of-the-art studies, demonstrating a promising new approach to predict ventilation images from IE-CT. By leveraging low-dimensional feature representations from IE-CT and integrating insights from a physics-informed model, our approach achieves high spatial correlation with SPECT-based ground truth. These results suggest the feasibility of using breath-hold CTs for predicting lung function, enabling faster, more accurate, and contrast-free lung function assessment within clinical workflows.

Acknowledgements and/or References (Optional)::