Abstract
About this talk
Optical coherence tomography (OCT) is central to diagnosing and monitoring retinal diseases, including diabetic retinopathy and age-related macular degeneration. Artificial intelligence (AI), particularly U-Net-based convolutional neural networks, may automate retinal fluid and pathology segmentation, reducing the time and variability associated with manual assessment. This systematic review and meta-analysis evaluated automated AI systems for segmenting and quantifying retinal abnormalities on OCT images compared with expert manual segmentation.
PubMed, Web of Science, and Scopus were searched for eligible studies published within the preceding five years. Sixteen diagnostic-accuracy studies involving 6,847 patients were included. Study quality was assessed using QUADAS-2. Outcomes included Dice scores, intraclass correlation coefficients (ICCs), volumetric agreement, device-related performance, and processing time. Pooled analyses found no significant differences between AI and human assessment for Dice scores (standardised mean difference [SMD] −0.08; 95% CI −1.16 to 0.99; p=0.88) or ICC values (SMD −0.13; 95% CI −5.70 to 5.45; p=0.96). However, substantial heterogeneity (I² >90%) limits the certainty of these estimates. AI achieved high segmentation performance for subretinal fluid (Dice 0.88–0.96) and geographic atrophy (0.94), while intraretinal fluid remained more challenging (0.79–0.89). Device-dependent variability was evident, with kappa values of 0.37 for ZEISS and 0.73 for Spectralis systems. Processing times ranged from 100 milliseconds per B-scan to several seconds per volume.
Automated AI systems demonstrate promising, often expert-level performance for selected retinal pathologies. Clinical implementation requires standardised multi-device validation, improved cross-platform adaptation, and hybrid AI-human workflows with expert oversight.
Speaker
About the speaker

