XAI-MeD: Explainable Knowledge Guided Neuro-Symbolic Framework for Domain Generalization and Rare Class Detection in Medical Imaging

TL;DR

XAI-MeD introduces a principled neuro-symbolic framework that integrates structured clinical knowledge with deep neural representations to address interpretability, domain generalization, and rare-class reliability in medical imaging. The EKSAII algorithm, driven by Entropy Imbalance Gain and Rare-Class Gini, adaptively routes inputs to class-specialized neural or symbolic experts, with outputs fused into a clinically aligned diagnosis and explanations via an LLM (e.g., GPT-4). Across two clinically important tasks—Seizure Onset Zone localization and Diabetic Retinopathy grading—XAI-MeD achieves robust cross- and multi-domain generalization, improves rare-class detection (e.g., up to a 10% absolute F1 gain), and reduces clinician workload through explainable reasoning that aligns with pathology. The framework is validated on ten multicenter datasets, with ablations showing symbolic components act as regularizers, and outputs that promote human-centered AI with transparent diagnostic rationales for end users.

Abstract

Explainability domain generalization and rare class reliability are critical challenges in medical AI where deep models often fail under real world distribution shifts and exhibit bias against infrequent clinical conditions This paper introduces XAIMeD an explainable medical AI framework that integrates clinically accurate expert knowledge into deep learning through a unified neuro symbolic architecture XAIMeD is designed to improve robustness under distribution shift enhance rare class sensitivity and deliver transparent clinically aligned interpretations The framework encodes clinical expertise as logical connectives over atomic medical propositions transforming them into machine checkable class specific rules Their diagnostic utility is quantified through weighted feature satisfaction scores enabling a symbolic reasoning branch that complements neural predictions A confidence weighted fusion integrates symbolic and deep outputs while a Hunt inspired adaptive routing mechanism guided by Entropy Imbalance Gain EIG and Rare Class Gini mitigates class imbalance high intra class variability and uncertainty We evaluate XAIMeD across diverse modalities on four challenging tasks i Seizure Onset Zone SOZ localization from rs fMRI ii Diabetic Retinopathy grading across 6 multicenter datasets demonstrate substantial performance improvements including 6 percent gains in cross domain generalization and a 10 percent improved rare class F1 score far outperforming state of the art deep learning baselines Ablation studies confirm that the clinically grounded symbolic components act as effective regularizers ensuring robustness to distribution shifts XAIMeD thus provides a principled clinically faithful and interpretable approach to multimodal medical AI.

Figures (4)

• Figure 1: Conceptual overview of the XAI-MeD framework.
• Figure 2: Overview of the XAI-MeD framework. The system integrates medical knowledge with multimodal imaging to enhance disease classification and provide clinically aligned, interpretable explanations with spatial localization.
• Figure 3: DeepXSOZ: A Hybrid Knowledge-AI Architecture for Seizure Onset Zone (SOZ) Localization. The framework employs a bipartite training architecture to classify Independent Components (ICs) derived from resting-state fMRI (rs-fMRI). The Deep Learning Machine ($\text{M}_{\text{D}}$) as Classifier 1 is trained on rs-fMRI ICs for an initial Noise/Non-noise component discrimination. Concurrently, the Expert Knowledge Integrator and Explainer Machine ($\text{M}_{\text{k}}$) as Classifier 2 computes a set of expert-derived knowledge components and learns the optimal weight configurations necessary for robust SOZ/RSN (Resting State Network) distinction and the generation of localized classification explanations. During inference, the final SOZ classification is determined by integrating the labels from both $\text{M}_{\text{D}}$ and $\text{M}_{\text{K}}$ via Algorithm 1, yielding a final, integrated, and explainable diagnostic result.
• Figure 4: The system integrates a Deep Learning Machine ($M_d^c$, ViT backbone for each class c) and an Expert Knowledge Machine ($M_k^c$, clinical features/guidelines for each class c) within a decision tree. The EKSAII Algorithm iteratively selects the optimal binary classifier (maximum Entropy Imbalance Gain, EIG) for node splitting, achieving $84\%$ accuracy across 5 DR classes through an orchestrated, sequential classification path.