Research on 3D/2D Vascular Registration Based on Monte Carlo Tree Search and Manifold Regularization

In interventional vascular surgery, enhanced intraoperative real-time imaging technology can compensate for the shortcomings of DSA navigation, such as the lack of depth information and excessive use of toxic contrast agents, by projecting preoperative computed tomography angiography (CTA) images onto intraoperative digital subtraction angiography (DSA) images. Among these technologies, 3D/2D vascular registration is a critical step. This study proposes a 3D/2D registration method based on vascular image matching.

I. Background and Research Motivation

Digital Subtraction Angiography (DSA) is the main imaging method used in minimally invasive interventional vascular surgery. By injecting a contrast agent into the vascular lumen of interest, a 2D image is obtained. Although DSA has high spatial and temporal resolution, it lacks depth information, and the overuse of contrast agents increases the patient’s burden. Therefore, combining the 3D blood vessels extracted from preoperative CTA images with DSA images is a practical application of 3D/2D registration technology, which can provide interventional radiologists with vascular depth information to support the manipulation of guide wires or catheters.

This paper is authored by Jianjun Zhu and colleagues from Hanglok-Tech and Zhongda Hospital, Southeast University, and was published in IEEE Transactions on Medical Imaging in May 2024.

II. Research Methods

Data Source and Preprocessing

The study used clinical data from three hospitals as well as public data to train a deep learning-based segmentation model and conducted experiments using simulated data. Vascular models were obtained through 3D CTA vessel segmentation and 2D DSA vessel segmentation.

3D/2D Registration Based on Vascular Image Matching

This study uses the Monte Carlo Tree Search (MCTS) method to achieve 3D/2D registration, decomposing vascular matching into a series of related states and constructing it into a tree structure. Each search tree node records the current vascular matching state and registration result. By iteratively expanding part of the search tree and evaluating node scores, the node with the highest overlap degree is found to achieve registration.

Non-Rigid Registration with Manifold Regularization

In non-rigid registration, dense 3D and 2D vascular correspondences are necessary. This paper introduces manifold regularization into the vascular deformation model, constructs the manifold regularization term of the objective function, and simplifies the gradient calculation formula to improve the efficiency of non-rigid registration.

III. Experimental Results

The study was validated on clinical hepatic artery, coronary artery, and aorta data and compared with seven rigid and three non-rigid registration methods.

Rigid Registration Experiments

Rigid registration experiments used simulated data to test the performance of various methods under different rotation angles and noise conditions. The experimental results showed that the proposed method outperformed existing methods in registration accuracy and computational efficiency. In particular, for situations with larger rotation angles, the proposed method demonstrated strong pose independence.

Non-Rigid Registration Experiments

Non-rigid registration was tested on simulated data with noise and deformation. The results indicated that the non-rigid registration method based on manifold regularization outperformed existing methods in terms of registration error and computation time. Especially for situations with higher noise levels, the proposed method demonstrated better robustness.

Generated Coronary Artery Vascular Image Registration

Using coronary artery CTA and DSA data, 24 pairs of vascular images were generated to test the performance of various methods under noise and deformation interference. The results showed that the proposed method had significant advantages in registration error and computation time.

Manually Annotated Vascular Registration

The study also used manually annotated hepatic artery, aorta, and coronary artery data to eliminate excess noise and verify the effectiveness of the proposed method in clinical applications. Analysis results showed that the proposed method performed excellently in registration accuracy and efficiency.

IV. Research Conclusion

This paper proposes a 3D/2D vascular registration method based on vascular topology information, combining Monte Carlo Tree Search and manifold regularization. In rigid registration, the method constructs a vascular matching search tree and performs iterative searches, significantly improving search efficiency and registration accuracy. In non-rigid registration, the deformation model based on manifold regularization effectively solves the deformation matching problem, further enhancing registration performance.

Through validation on hepatic artery, coronary artery, and aorta data, experimental results showed that the proposed method outperformed existing mainstream methods in terms of registration accuracy and efficiency. The paper also explored methods for constructing and matching vascular images and detailed the method for calculating the maximum match and its application in constructing deformation models.

Highlights and Significance

The research method covers vascular registration from rigid to non-rigid, utilizing new methods for vascular image matching and manifold regularization to improve registration accuracy and computational efficiency. This research has important scientific value and provides practical application support for clinical surgical navigation. Particularly, the proposed method for manifold regularization can be applied to other cross-dimensional registration tasks based on topological structures.

Limitations and Outlook

Despite the excellent performance of the proposed method in experiments, its dependence on vascular centerline topology remains a limitation, as errors in generating the topological structure may affect registration results. Future directions include using more efficient computational strategies to further enhance real-time performance and exploring the possibility of directly using CT and DSA images for rapid and accurate registration without preprocessing steps.

Through experimental validation on typical vascular anatomical structures and systematic comparisons with various existing methods, the proposed 3D/2D vascular registration method demonstrated good robustness and high efficiency, showing potential for greater application in clinical practice.