Semi-supervised motion and anatomy dual aware fusion network for lung 4D CT ventilation imaging

Four-dimensional computed tomography(4DCT)-derived lung ventilation imaging is emerging as a potentially crucial tool for assessing regional lung function and guiding functional avoidance radiotherapy. However, acquiring accurate lung ventilation images from 4D CT scans remains challenging, due to difficulties in capturing both lung volume changes and anatomical density variations across the 4D CT sequence, as well as the difficulty of effectively integrating these heterogeneous types of information. In this work, we propose a semi-supervised motion and anatomy dual aware fusion network for lung 4D CT ventilation imaging. Specifically, the deformation field obtained from registering adjacent phases of the 4D CT serves as the lung motion information, from which regional changes in lung volume are quantified by calculating the Jacobian determinant of deformation field. Then, a motion-guided anatomical density enhancement strategy is designed, aiming to capture anatomical density variations through the lung motion information. Subsequently, a motion-anatomy fusion module is designed to address the discrepancies between motion-derived volume changes and anatomical density variations, resulting in a robust fused ventilation representation. Furthermore, we employ a semi-supervised mean teacher framework for training, to mitigate scarcity of labeled data. Experimental results on three public datasets indicate that the mean dice similarity coefficient (DSC) for high-functional lung regions is 0.71±0.06, and for low-functional lung regions, the DSC is 0.75±0.08. The mean Spearman correlation r_s between predicted and corresponding reference ventilation images is 0.76±0.09, achieving higher imaging accuracy than state-of-the-art methods. Moreover, additional held-out validation experiments across different cohorts and functional imaging modalities further demonstrate the robustness and generalization ability of the proposed method to new data distributions.