GRAM: Graph Regularizable Assessment Metric
Keywords: Predicted brain graphs · Quality metrics · Customized metrics
TL;DR: We propose a universal graph quality metric for evaluating the quality of predicted graphs using generative models.
Abstract: Here, we propose the Graph Regularizable Assessment Metric ($GRAM$), a customizable tool for evaluating the quality of generated brain graphs. Current geometric deep learning methods often lack robust quantification techniques for assessing the synthetic brain graphs integrity. $GRAM$ addresses this gap by proportionally combining a set of existing graph metrics to establish a linear correlation between distortions' levels and metric values of ground-truth graphs. To evaluate the performance of our model, we generated a synthetic dataset of structural brain connectomes which was derived from an existing dataset and used to simulate a set of predicted connectomes from a generative model with controlled levels of distortions. Our results show that $GRAM$ outperforms single metrics in quantifying the distortion between generated and original graphs. This approach is a significant step towards establishing a universal graph quality index for graph-based predictive studies.
Submission Number: 6
Loading