Shape-aware inference scheme for selective extraction of head-neck arteries on computer tomography angiography images

Background and Objective:

The extraction of head-neck blood vessels from 3D medical images plays a vital role in diagnosing vascular diseases. While many existing methods rely on convolutional neural networks (CNNs), they can encounter challenges in maintaining the continuity of extracted vessels, particularly when segmenting these slender tubular structures within 3D images. This study aims to address these challenges by introducing a novel inference method that preserves vessel continuity taking into consideration the global geometry of vascular structures, while also reducing the number of high resolution patches being processed.

Methods:

The proposed approach employs a two-stage CNN process complimented by a centerline-aware thresholding step. First, a CNN performs preliminary localization of arteries within a highly downsampled volume, identifying seed points. These seed points serve as input for a second CNN, specialized in capturing local artery appearances, to perform precise segmentation. A centerline-based artery tracking algorithm is then applied to guide patch segmentation along the artery until the entire vascular structure is segmented. Physical connectivity is ensured by our novel centerline-aware thresholding strategy which is used to construct the final segmentation mask.

Results:

The proposed method effectively reduces the number of high-resolution patches processed by the neural network, thereby not only addressing the issue of class imbalance by primarily concentrating on patches containing arteries but also reducing computational complexity. The performance is comparable to state of the art methods across various metrics, while the algorithm additionally recovers missing segments of falsely interrupted arteries, thereby facilitating automatic extraction of medically relevant quantities like for example artery length. The method requires significantly less memory and performs approximately one-tenth of the computations needed to segment a patient.

Conclusion:

The proposed approach offers an effective solution for extracting blood vessels from 3D medical images, overcoming the limitations of traditional CNN-based methods by preserving vessel continuity and addressing the challenge of class imbalance. This approach enables the selective segmentation of specific arteries in a resource-efficient manner, thereby enhancing the diagnosis and treatment of vascular diseases by providing more accurate vascular segmentations.