Monitoring three-dimensional (3D) structural information of cardiac vessels is crucial for diagnosis and treatment of cardiovascular diseases. The current gold standard imaging modality in clinical practice is two-dimensional (2D) coronary angiography, where surgeons empirically infer 3D cardiovascular topology from multiple X-ray views. This process is time-consuming and requires rich diagnostic expertise, due to the continuous heartbeats and extremely sparse observations in surgical routine. Despite extensive research efforts on cardiovascular reconstruction algorithms, existing methods face the dilemma between the need of manual annotations and reconstruction accuracy, limiting their utility in providing clinical assistance. In this study, we introduce AutoCAR, a deep-learning-based solution for fully-automated 3D coronary angiography reconstruction. AutoCAR integrates expert-guided domain adaptation and sparse backward projection. The former imitates imaging preferences of experts for generalizability, while the latter utilizes vascular sparsity, as 1D manifolds in 3D space, to preserve high-resolution details by on-manifold computation.AutoCAR significantly outperforms state-of-the-art methods in both qualitative and quantitative evaluations on real-world data. AutoCAR achieves competitive performance with a time reduction of two orders of magnitude (7s versus ~200s per 3D volume), compared to experienced interventionalists. In addition, we provide open software for public users to upload their own images for 3D reconstruction, which can be integrated into standard X-ray imaging platforms. We envision that AutoCAR will significantly facilitate current diagnostic and intervention procedures and pave the way for real-time visual guidance and autonomous catheter navigation in cardiac intervention.
* If the 3D structure is not in deformation, please refresh webpage or view it in
desktop
browser.
* Click Controls to switch examples and adjust settings. More cases
* Move mouse over 2D images for before-after comparision.
* The 2D-2D/2D-3D correspondences is color encoded.
In current intravascular intervention practice, rotational X-ray angiography is commonly used for image guidance. From C-arm scanner(0:00) with protocol(0:03), weakly synchronized multi-view video(0:12) can be obtained. In several clinical procedure, e.g. Stenosis Severity Diagnosis, Intravascular Navigation, a fundamental task is finding the correspondence across multiple view and time(0:23). Instead of relying on manual interpretation, we propose AutoCAR, a learning-based dynamic reconstruction method, as a fully automated solution to assist junior operator and reduce procedure time which further reduce radiation exposure and contrast medium usage(0:36).
*click timestamp(0:00) for jumping to corresponding video
section.
* Selected cases are shown above. For more cases, click here
* Move mouse over 2D images for before-after comparision.
* The 2D-2D/2D-3D correspondences is color encoded.
Due to browser
limitation
, please check out in a separated page.
* Move mouse over 2D images for before-after comparision.
* The 2D-2D/2D-3D correspondences is color encoded.
* the style file was borrowed from MipNerf