eICAB: A novel deep learning pipeline for Circle of Willis multiclass segmentation and analysis

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Accurate segmentation, labeling, and quantification of cerebral blood vessels in MR imaging are crucial for research.
• Current methods often lack generalizability and require significant user intervention, necessitating automated solutions.

Purpose of the Study:

• To develop an automated method for segmenting, labeling, and quantifying Circle of Willis (CW) arteries using deep convolutional neural networks (CNNs) on MRA images.

Main Methods:

• A CNN ensemble model (eICAB) was trained on 101 MRA images and tested on 15, with 14 CW arterial segments manually annotated.
• Performance was evaluated using quantitative (Dice score) and qualitative analyses on internal and external datasets.
• Reliability was assessed through test-retest analysis of vessel diameters and volumes.

Main Results:

• eICAB achieved high accuracy in predicting large (99%), medium (97%), and small (88%) arteries.
• Average Dice scores were 0.76 for large, 0.76 for medium, and 0.41 for small vessels.
• The method demonstrated high reliability (ICC=0.99) and outperformed inter-expert variability in some cases.

Conclusions:

• An open-source, CNN-based method (eICAB) was developed for accurate and reliable segmentation and labeling of the CW in MRA images.
• The method is largely independent of image quality and offers a significant step towards automated MRA database analysis.
• This approach holds promise for advancing basic and clinical research in cerebrovascular diseases.