NEURO-GUARD: Neuro-Symbolic Generalization and Unbiased Adaptive Routing for Diagnostics -- Explainable Medical AI

NEURO-GUARD: Neuro-Symbolic Generalization and Unbiased Adaptive Routing for Diagnostics - Explainable Medical AI Midhat Urooj Arizona State University Tempe, AZ, USA murooj@asu.edu Ayan Banerjee Arizona State University Tempe, AZ, USA abanerj3@asu.edu Sandeep Gupta Arizona State University Tempe, AZ, USA Sandeep.Gupta@asu.edu Abstract Accurate yet interpretable image-based diagnosis remains a central challenge in medical AI, particularly in settings with limited data, subtle visual patterns, and high-stakes clinical decisions. However, most current vision models produce black-box predictions with limited generalizabil- ity and poor real-world usability. We present NEURO- GUARD, a novel framework that combines Vision Trans- formers (ViTs) with knowledge-guided reasoning to en- hance performance, transparency, and cross-domain gen- eralization. NEURO-GUARD incorporates a retrieval- augmented generation (RAG) mechanism for language- driven self-verification, in which a large language model (LLM) iteratively generates, evaluates, and refines feature extraction code for medical images. By leveraging clini- cal guidelines and expert knowledge, this LLM-guided mod- ule progressively improves feature detection and classifica- tion, outperforming purely data-driven baselines. Extensive evaluations on diabetic retinopathy classification across four benchmark datasets (APTOS, EyePACS, Messidor-1, Messidor-2) show that NEURO-GUARD improves accuracy by 6.2% over a ViT-only model (84.69% vs. 78.4% [3]) and achieves a 5% gain in domain generalization. Fur- ther experiments on MRI-based seizure detection confirm its cross-domain robustness, consistently surpassing existing baselines. Notably, NEURO-GUARD bridges the gap be- tween symbolic medical reasoning and subsymbolic feature learning, demonstrating robust generalization across multi- ple datasets while achieving state-of-the-art performance. 1. Introduction Medical imaging plays a crucial role in disease diagnosis and treatment planning, particularly in conditions such as diabetic retinopathy (DR), tumor detection, and neurode- generative disorders. Recent advances in deep learning, Figure 1. Performance comparison of existing models versus the NEURO-GUARD framework for 5-stage Diabetic Retinopa- thy classification. particularly Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs), have significantly improved di- agnostic accuracy [3, 14]. However, their black-box na- ture limits clinical adoption due to a lack of interpretability, making it challenging for clinicians to validate AI-driven decisions. Additionally, these models suffer from domain shift vulnerabilities, struggling to generalize across imag- ing datasets with diverse acquisition protocols and patient demographics [22, 26]. Given these challenges, an ideal medical AI framework should not only provide high accu- racy but also generate clinically interpretable decisions by integrating structured domain knowledge into its reasoning process. Existing explainability techniques, such as Gradient- weighted Class Activation Mapping (Grad-CAM) [13] and Shapley Additive Explanations (SHAP) [8], provide post- hoc feature attribution but remain static, heuristic-based, and disconnected from the model’s decision logic. Hy- brid approaches incorporating attention mechanisms and uncertainty estimation attempt to improve interpretabil- arXiv:2512.18177v1 [cs.AI] 20 Dec 2025 Figure 2. Overview of the NEURO-GUARD framework. The system integrates medical knowledge with multimodal imaging to enhance disease classification and provide clinically aligned, interpretable explanations with spatial localization. ity [18, 20], but they fail to integrate structured medi- cal knowledge, limiting their ability to generalize across datasets. Reinforcement learning (RL) and meta-learning frameworks [10] enable adaptive learning, yet they lack mechanisms to ground AI decisions in clinical reasoning, reducing their reliability in real-world medical applications. To address these limitations, we propose NEURO- GUARD, a novel framework that fuses language-grounded reasoning with state-of-the-art visual recognition to enable intrinsically interpretable medical image diagnosis. In con- trast to prior systems that only add interpretability after the fact, NEURO-GUARD tightly integrates a clinical knowl- edge base and reasoning module into the model’s infer- ence pipeline. This is achieved through a modular archi- tecture combining a self-supervised ViT-based image en- coder with a knowledge-guided language model that jointly analyzes images and textual information. Crucially, our approach leverages retrieval-augmented generation to dy- namically draw on external biomedical sources (e.g., liter- ature, guidelines) for case-specific knowledge, and uses an LLM-based code synthesis engine to translate this knowl- edge into executable image analysis steps. A prompt- driven self-verification loop, optimized via reinforcement learni

…(Full text truncated)…

This content is AI-processed based on ArXiv data.