FireGNN is an interpretable graph-based learning framework developed for medical image classification. FireGNN integrates trainable fuzzy rules into graph neural networks (GNNs) to enhance clinical reliability and usability. This framework embeds topological features, such as node degree, cluster coefficients, and label concordance, using learnable thresholds and sharpness parameters to enable inherent symbolic inference. Furthermore, we explore auxiliary self-supervised learning tasks, such as homology prediction and similarity entropy, to evaluate the contribution of topological learning. FireGNN achieves robust performance on five MedMNIST benchmarks and the synthetic dataset MorphoMNIST, generating interpretable rule-based explanations.
