Seeing the Invisible: On Aortic Valve Reconstruction in Non-contrast CT

Accurate segmentation of the aortic valve (AV) in computed tomography (CT) scans is crucial for assessing AV disease severity and identifying patients who may benefit from interventional treatments, such as surgical and percutaneous procedures. Evaluation of AV calcium score on non-contrast CT scans emphasizes the importance of identifying AV from these scans. However, it is not a trivial task due to the extremely low visibility of AV in this type of medical images. In this paper, we propose a method for semi-automatic generation of ground truth (GT) data for this problem based on image registration. In a weakly-supervised learning process, we train neural network models capable of accurate segmentation of AV based exclusively on non-contrast CT scans. We also present a novel approach for the evaluation of segmentation accuracy, based on per-patient, rigid registration of masks segmented in contrast and non-contrast images. Evaluation on an open-source dataset demonstrates that our model can identify AV with a mean error of less than 1 mm, suggesting significant potential for clinical application. In particular, the model can be used to enhance end-to-end deep learning approaches for AV calcium scoring by offering substantial accuracy improvements and increasing the explainability. Furthermore, it contributes to lowering the rate of false positives in coronary artery calcium scoring through the meticulous exclusion of aortic root calcifications.