Integrating Human Expertise With Artificial Intelligence (AI) Models for Optical Coherence Tomography (OCT) Retinal Fluid and Pathology Quantification: A Systematic Review.

Abstract

Artificial intelligence (AI) and intensive learning show promise in ophthalmology, using optical coherence tomography (OCT) to diagnose conditions such as diabetic retinopathy. However, precise segmentation of retinal fluid remains challenging, especially in atypical cases and low-quality scans. In hospital settings, manual segmentation is time-consuming; however, integrating human expertise with AI could improve efficiency and accuracy in quantifying retinal pathologies. This review assesses the efficacy of human-AI collaborative workflows for enhancing the accuracy, efficiency, and clinical utility of retinal pathology quantification in OCT. We conducted a systematic review of the literature from 2021 to 2025 across three databases: PubMed, Web of Science, and Scopus. This review included nine studies that quantified retinal pathology using human-AI collaborative workflows in OCT. Two independent reviewers screened the records, extracted relevant data, including study design, AI architecture, and performance metrics, and assessed the quality of the studies using the Quality Assessment of Diagnostic Accuracy Studies - 2nd version (QUADAS-2 ) and Risk Of Bias In Non-randomized Studies of Interventions (ROBINS-I) tools.  This systematic review of nine AI-OCT studies (2021-2025) found that hybrid AI-clinician workflows achieved expert-level reliability for 11 of 13 retinal biomarkers across cohorts of 16-1,097 individuals. AI architectures, including U-Net variants, polypoidal choroidal vasculopathy (PCV)-Net, and custom convolutional neural networks (CNNs), performed well for well-defined features such as retinal layers (Dice 0.94), intraretinal/subretinal fluid (Dice 0.61-0.67), and atrophic areas (F1 0.78-0.89), but struggled with complex biomarkers like sub-retinal pigment epithelium (sub-RPE) lesions (Dice 0.11). Clinician agreement on fluid volumes was strong (Pearson r > 0.85), though volumetric errors increased in atrophic regions. These workflows reduced processing time by over 50% compared with manual grading while improving monitoring precision for neovascular age-related macular degeneration (AMD) and retinitis pigmentosa complications in both single-center and international trial settings. Combining AI quantification with clinician expertise enhances both the accuracy and efficiency of retinal pathology assessments. This integration supports personalized treatment planning and facilitates large-scale research. Hybrid approaches address AI's limitations, highlighting their practicality for clinical use in ophthalmology. We suggest incorporating hybrid AI-human workflows in clinical practice to improve the efficiency and accuracy of OCT analysis. Future developments in AI should focus on standardized training for complex biomarkers, such as sub-RPE lesions.