Cerebral Artery Classification: Integrating Synthetic Models with MRA Data.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• The Circle of Willis (CoW) is vital for cerebral blood flow but exhibits significant anatomical variation.
• Accurate segmentation and classification of CoW arteries are challenging due to this variability and limited annotated datasets.
• Magnetic Resonance Angiography (MRA) is a key imaging modality, but data scarcity hampers machine learning model performance.

Purpose of the Study:

• To improve the segmentation and classification of Circle of Willis arteries using deep learning.
• To address the challenge of limited annotated data in medical imaging datasets for cerebrovascular research.
• To evaluate the effectiveness of using synthetic vascular models to augment real MRA data.

Main Methods:

• Leveraged VaMos, a vascular modeling tool, to generate anatomically plausible synthetic Circle of Willis models.
• Augmented the existing TopCoW dataset with these synthetic models.
• Employed the nnUNet deep learning framework, optimized for medical image segmentation, to train models on the combined dataset.
• Evaluated performance improvements, particularly for underrepresented arteries.

Main Results:

• The integration of synthetic data significantly improved artery classification accuracy.
• A notable improvement of approximately 16% was observed for smaller and thinner arteries, including the posterior and anterior communicating arteries.
• The study demonstrated the efficacy of synthetic data in enhancing deep learning model performance for challenging anatomical structures.

Conclusions:

• Synthetic vascular models offer a viable solution to data scarcity in medical imaging.
• Augmenting real MRA data with anatomically plausible synthetic models enhances the performance of deep learning segmentation models.
• This approach holds promise for improving clinical applications requiring accurate cerebrovascular segmentation.