Coronary artery disease stands as a leading cause of mortality worldwide. Myocardium at Risk is an important clinical metric for assessing the significance of a coronary artery lesion and determining treatment options. To compute MAR, each vessel needs to be mapped with the associated myocardial region it sustains, thereby determining the portion of myocardium affected by the disease of a particular vessel. This mapping of the coronary arteries is also important for patient-specific blood flow simulations. One prevalent approach reported in the literature for computing MAR from Coronary Computed Tomography Angiography (CCTA) involves Voronoi-based subdivision. However, in clinical CCTA imaging, image resolution and noise are limiting factors when segmenting the coronary tree, compromising the accuracy of Voronoi-based parcellation to compute MAR. Our study aims to introduce a novel approach based on Graph Convolutional Networks, to leverage the results obtained from Voronoi-based simulations on myocardial meshes with detailed trees and enhance outcomes in scenarios with less detailed trees. The model has been trained on a dataset of detailed trees extracted from high quality CCTA images (600 patients) that have been synthetically trimmed to reduce detail, and tested on a dataset (208 patients) with human augmented annotations with three levels of detail. Our findings reveal a reduction in variance between detailed and less detailed tree parcellations using our proposed model compared to the baseline Voronoi method, while still achieving a high accuracy rate of 0.9 on both synthetically trimmed and real datasets.