Automated three-dimensional vessel reconstruction based on deep segmentation and bi-plane angiographic projections

Automated three-dimensional (3D) blood vessel reconstruction to improve vascular diagnosis and therapeutics is a challenging task in which the real-time implementation of automatic segmentation and specific vessel tracking for matching artery sequences is essential. Recently, a deep learning-based segmentation technique has been proposed; however, existing state-of-the-art deep architectures exhibit reduced performance when they are employed using real in-vivo imaging because of serious issues such as low contrast and noise contamination of the X-ray images. To overcome these limitations, we propose a novel methodology composed of the de-haze image enhancement technique as pre-processing and multi-level thresholding as post-processing to be applied to the lightweight multi-resolution U-shaped architecture. Specifically, (1) bi-plane two-dimensional (2D) vessel images were extracted simultaneously using the deep architecture, (2) skeletons of the vessels were computed via a morphology operation, (3) the corresponding skeleton structure between image sequences was matched using the shape-context technique, and (4) the 3D centerline was reconstructed using stereo geometry. The method was validated using both in-vivo and in-vitro models. The results show that the proposed technique could improve the segmentation quality, reduce computation time, and reconstruct the 3D skeleton automatically. The algorithm accurately reconstructed the phantom model and the real mouse vessel in 3D in 2 s. Our proposed technique has the potential to allow therapeutic micro-agent navigation in clinical practice, thereby providing the 3D position and orientation of the vessel.