Automatic cerebrovascular segmentation and 3D-2D registration for image-guided endovascular interventions

Le résumé fourni par la source

Image-guided endovascular interventions (IGEVI) is widely used to treat cerebrovascular diseases, in which catheters and guidewires are navigated using image guidance. Typically, pre-operative 3D images are acquired to visualize the entire vascular structure and lesions for treatment planning, while intra-operative 2D images provide real-time guidance. Fusing these imaging modalities enables precise navigation and accurate targeting during the IGEVI. To improve procedural efficiency, we propose a fully automated framework that integrates cerebrovascular segmentation with 3D-2D registration. The segmentation employs a deep learning model, featuring a dual-path encoder composed of a convolutional path for capturing spatial features and a multi-head self-attention transformer path for extracting contextual information. The registration estimates the 6 degrees of freedom (3 rotations and 3 translations) source pose using a differentiable operation that enables gradient-based optimization with respect to segmented vessel structures. The proposed framework was trained and tested using publicly available COSTA dataset comprising six sub-datasets collected from multiple centers. The segmentation model was trained and tested on 294 and 61 TOF-MRA images, respectively. The proposed method achieved Dice similarity coefficient (DSC), sensitivity and precision of 0.901±0.020, 0.912±0.026, and 0.891±0.035 (mean ± standard deviation), respectively, outperforming 3D Unet and SwinUnetr on all test cases. The low variances indicate consistent segmentation performance across sub-datasets. The registration was performed on simulated 2D digitally reconstructed radiographs generated by forward projecting the 3D vessel volume of 61 test images along anterior-posterior direction. The initial poses for registration were randomly offset from ground truth source poses, and pose estimation was carried out using a combined normalized cross correlation and centerline DSC loss function. The registration yielded rotation and translation errors of 1.731±3.385° and 1.028±1.827 mm, respectively. The experimental results demonstrate promising performance for the segmentation and challenging single-view 3D-2D registration, showing the strong potential for use in real IGEVI procedures.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.