<![CDATA[Background: Diabetic Retinopathy (DR) is a progressive microvascular complication of diabetes and a leading cause of preventable vision impairment worldwide. Early and accurate identification of DR severity is essential for timely intervention and effective disease management. However, conventional retinal screening relies heavily on manual examination of fundus images, which is labor-intensive, subjective, and dependent on specialist expertise. Although recent deep-learning models have demonstrated promising diagnostic performance, many existing approaches inadequately capture lesion-level relationships, underutilize clinical information, and provide limited interpretability for clinical decision-making. Aims: this study proposes a novel Neuro-Spectral Causal Fusion Network (NSCF-Net) for explainable diabetic retinopathy detection and severity grading. Methods: The proposed framework integrates an adaptive Neuro-Spectral Enhancement Module to improve lesion visibility, a Lesion Dependency Reasoning Graph to model pathological interactions among retinal abnormalities, and a Dual-Path Hybrid Transformer to jointly learn local lesion characteristics and global retinal context. Furthermore, a Causal Multimodal Fusion Unit combines retinal image representations with clinically relevant parameters, including HbA1c level, blood pressure, age, and duration of diabetes. To improve transparency and clinical trustworthiness, an Explainable Lesion Attribution Module generates counterfactual attention maps and hierarchical saliency visualizations highlighting diagnostically significant retinal regions. The proposed framework was evaluated on the APTOS 2019 diabetic retinopathy dataset, consisting of 3,662 retinal fundus images distributed across five DR severity classes. Results: Experimental results demonstrate the effectiveness of NSCF-Net, achieving 99.1% accuracy, 98.9% precision, 99.0% recall, and 98.8% F1-score. Comparative analyses and ablation studies further confirm the contribution of neuro-spectral enhancement, lesion dependency reasoning, multimodal fusion, and explainability components to overall performance improvement. Conclusion: These findings indicate that NSCF-Net provides an accurate, interpretable, and clinically reliable framework for automated diabetic retinopathy screening, severity assessment, and computer-aided ophthalmic decision support.]]>
