Many cerebrovascular diseases are related to morphological variations in the Circle of Willis (CoW), an arterial network located at the base of the brain. Early detection of these structural abnormalities can result in effective treatments and helps to prevent the progression of the diseases to more advanced stages. This necessitates the need for developing a computer-aided model capable of automatically identifying anatomical variants of the CoW using a standardized classification framework, such as the Lippert and Pabst system. However, there are no reported studies that have applied the Lippert and Pabst classification in the context of computer-assisted analysis of CoW variants. Due to the small size and high class imbalance often present in medical datasets, it becomes challenging for standard CNNs to effectively capture CoW variants in classification tasks. To address this, we developed a novel graph-based method that incorporates spectral analysis with a hybrid Convolutional Neural Network and Graph Neural Network architecture to capture the complex morphological structures of the CoW. We conducted a detailed study comparing the performance of the proposed method in classifying anterior and posterior CoW variants across various configurations of VGG and ResNet networks. The proposed method attains a balanced accuracy of 0.69 for anterior and 0.71 for posterior CoW classification, indicating that the proposed framework significantly improved CoW classification performance across both anterior and posterior classes.
Preena et al. (Mon,) studied this question.