This narrative review on diabetic optic neuropathy (DON) argues AI—especially radiomics and multimodal machine learning—could surface candidate imaging biomarkers and quantify subtle optic nerve neurodegeneration earlier than current tools, helping shift care from reactive to proactive, precision-based management in Diabetes mellitus (DM).

Why It Matters To Your Practice

• DON is an underdiagnosed, vision-threatening neurodegenerative complication of DM, in part because today’s clinical tools and criteria struggle with heterogeneity and lack validated biomarkers.

• AI approaches that succeeded in diabetic retinopathy (DR) screening may be translatable to DON—but the target changes from microvascular lesions to neurodegenerative optic nerve changes.

Clinical Implications

• Radiomics could enable earlier detection by extracting high-dimensional features from imaging that are not reliably visible or quantifiable at the slit lamp or on routine reads.

• Multimodal models (imaging + clinical data) may improve differential diagnosis (e.g., separating DON from other optic neuropathies) and support risk stratification and progression prediction.

• Implementation will likely require explainable AI outputs (feature attribution, uncertainty reporting) to build clinician trust and support accountable decisions.

Insights

• The review frames DON as multifactorial (metabolic, vascular, inflammatory, neurodegenerative), implying single-modality biomarkers may be insufficient and favoring integrated models.

• Key AI use cases highlighted include optic nerve head segmentation, classification, predictive analytics, and biomarker discovery—radiomics being the bridge from “pattern recognition” to “candidate mechanism-linked markers.”

• Longitudinal datasets are positioned as the unlock for moving from detection to forecasting clinically meaningful decline and treatment windows.

The Bottom Line

• For clinicians, the near-term value proposition is decision support: earlier suspicion of DON, better phenotyping, and clearer trajectories—if models are explainable, validated, and integrated into workflow.