Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FireGNN: Neuro-Symbolic Graph Neural Networks with Trainable Fuzzy Rules for Interpretable Medical Image Classification

Prajit Sengupta, Islem Rekik

TL;DR

FireGNN tackles the interpretability bottleneck in medical image classification by integrating trainable fuzzy rules over local graph topology into GNNs. The method formulates topological features $f_u=[d(u), C(u), L(u)]$ and learns thresholds $\theta_i$ and sharpness $\alpha_i$ for three fuzzy rules, whose activations $r_i(u)=\sigma(\alpha_i(f_u[i]-\theta_i))$ are fused with neural embeddings through a gating mechanism to yield interpretable predictions. Empirically, FireGNN achieves strong performance across six datasets (five MedMNIST variants and MorphoMNIST) and provides intrinsic explanations via rule activations, outperforming auxiliary self-supervised baselines. While demonstrating clear benefits for trustworthy medical AI, the approach incurs higher computational costs and currently uses a compact rule set, pointing to future work on richer rule vocabularies and efficiency improvements.

Abstract

Medical image classification requires not only high predictive performance but also interpretability to ensure clinical trust and adoption. Graph Neural Networks (GNNs) offer a powerful framework for modeling relational structures within datasets; however, standard GNNs often operate as black boxes, limiting transparency and usability, particularly in clinical settings. In this work, we present an interpretable graph-based learning framework named FireGNN that integrates trainable fuzzy rules into GNNs for medical image classification. These rules embed topological descriptors - node degree, clustering coefficient, and label agreement - using learnable thresholds and sharpness parameters to enable intrinsic symbolic reasoning. Additionally, we explore auxiliary self-supervised tasks (e.g., homophily prediction, similarity entropy) as a benchmark to evaluate the contribution of topological learning. Our fuzzy-rule-enhanced model achieves strong performance across five MedMNIST benchmarks and the synthetic dataset MorphoMNIST, while also generating interpretable rule-based explanations. To our knowledge, this is the first integration of trainable fuzzy rules within a GNN. Source Code: https://github.com/basiralab/FireGNN

FireGNN: Neuro-Symbolic Graph Neural Networks with Trainable Fuzzy Rules for Interpretable Medical Image Classification

TL;DR

FireGNN tackles the interpretability bottleneck in medical image classification by integrating trainable fuzzy rules over local graph topology into GNNs. The method formulates topological features and learns thresholds and sharpness for three fuzzy rules, whose activations are fused with neural embeddings through a gating mechanism to yield interpretable predictions. Empirically, FireGNN achieves strong performance across six datasets (five MedMNIST variants and MorphoMNIST) and provides intrinsic explanations via rule activations, outperforming auxiliary self-supervised baselines. While demonstrating clear benefits for trustworthy medical AI, the approach incurs higher computational costs and currently uses a compact rule set, pointing to future work on richer rule vocabularies and efficiency improvements.

Abstract

Medical image classification requires not only high predictive performance but also interpretability to ensure clinical trust and adoption. Graph Neural Networks (GNNs) offer a powerful framework for modeling relational structures within datasets; however, standard GNNs often operate as black boxes, limiting transparency and usability, particularly in clinical settings. In this work, we present an interpretable graph-based learning framework named FireGNN that integrates trainable fuzzy rules into GNNs for medical image classification. These rules embed topological descriptors - node degree, clustering coefficient, and label agreement - using learnable thresholds and sharpness parameters to enable intrinsic symbolic reasoning. Additionally, we explore auxiliary self-supervised tasks (e.g., homophily prediction, similarity entropy) as a benchmark to evaluate the contribution of topological learning. Our fuzzy-rule-enhanced model achieves strong performance across five MedMNIST benchmarks and the synthetic dataset MorphoMNIST, while also generating interpretable rule-based explanations. To our knowledge, this is the first integration of trainable fuzzy rules within a GNN. Source Code: https://github.com/basiralab/FireGNN

Paper Structure

This paper contains 18 sections, 11 equations, 1 figure, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Results and Discussion
  5. Technical Appendices and Supplementary Material
  6. Dataset Overview
  7. GNN with Auxiliary Tasks Mathematical Formalization
  8. Computational efficiency & learning dynamics
  9. Hyperparameters for FireGNN Model & Baselines
  10. Proposed FireGNN vs Limitations
  11. Extended Showcase of Learned Fuzzy Rules
  12. Learned Rules for BloodMNIST
  13. Learned Rules for MorphoMNIST
  14. Alternative Graph-Theoretic Rules for FireGNN
  15. 1. Rules from Node Centrality Measures
  16. ...and 3 more sections

Figures (1)

  • Figure 1: FireGNN architecture with four main components: (A) Graph construction, (B) Extraction of topological features, (C) GNN processing, and (D) Trainable fuzzy rules.