Background: MRI is a valuable tool for the intervention and treatment of cerebrovascular diseases. However, quantitative segmentation of blood vessels from raw MRI data, to study the disease and perform clinical measurements, is difficult and time-consuming. Automated methods provide a solution to this; however, such tools are scarce.
Aims: Due to the absence of open-access tools, we aim to develop an open-access segmentation method that generates cerebral vessel segmentations from Magnetic Resonance Angiography (MRA).
Methods: The novel method combines vessel size-specific Hessian filters and hysteresis thresholding for segmentation and connected component correction (CCC) for noise removal. The optimal choice of processing steps was evaluated with a blinded scoring by a clinician using 24 image samples from the IXI Dataset. The code is available at https://github.com/georgiakenyon/Segmentation-method-for-cerebral-blood-vessels-from-MRA-using-hysteresis.
Results: The ablation study confirmed that all steps of the method, in the specific sequence established, were required to optimize the segmentation output and produce the highest quality score (14.2/15). Omitting the CCC caused the largest quality loss (11.0/15), as CCC removed clusters of noise without affecting the structural integrity of the vessels, unlike existing morphological operations (e.g. erosion). Hysteresis thresholding reduced the loss of vessel edges during segmentation. Optimisation of Hessian filters for vessels of different radii improved segmentation, with complete loss of small vessel detail shown when a low Hessian parameter filter was omitted (12.8/15), and under-prediction of larger vessels when a high value Hessian filter was removed (12.4/15).
Conclusion: We provide and validate a method to efficiently segment vessels from brain MRA images for quantitative studies, improved diagnostics and intervention planning. The method, which is available on GitHub, can further be used to produce training data for solving the vessel segmentation problem from non-MRA modalities (e.g. T2-weighted images) using deep learning.