Eye diseases such as age-related macular degeneration (AMD) are major causes of irreversible vision loss. Early and accurate detection of these diseases is essential for effective management. Optical coherence tomography (OCT) imaging provides clinicians with in vivo, cross-sectional views of the retina, enabling the identification of key pathological features. However, manual interpretation of OCT scans is labor-intensive and prone to variability, often leading to diagnostic inconsistencies. To address this, we leveraged the RETFound model, a foundation model pretrained on 1.6 million unlabeled retinal OCT images, to automate the classification of key disease signatures on OCT. We finetuned RETFound and compared its performance with the widely used ResNet-50 model, using single-task and multitask modes. The dataset included 1770 labeled B-scans with various disease features, including subretinal fluid (SRF), intraretinal fluid (IRF), drusen, and pigment epithelial detachment (PED). The performance was evaluated using accuracy and AUC-ROC values, which ranged across models from 0.75 to 0.77 and 0.75 to 0.80, respectively. RETFound models display comparable specificity and sensitivity to ResNet-50 models overall, making it also a promising tool for retinal disease diagnosis. These findings suggest that RETFound may offer improved diagnostic accuracy and interpretability for specific tasks, potentially aiding clinicians in more efficient and reliable OCT image analysis.

Purpose: To validate a deep learning algorithm for automated segmentation and quantitative assessment of the ellipsoid zone (EZ) and RPE–Bruch’s membrane (BM) complex in healthy and geographic atrophy (GA) eyes. Methods: In this retrospective study, SD-OCT volume scans from 30 healthy and 30 eyes with GA were analysed. NMI-Outer Retina Analyzer was used to segment the inner EZ, inner RPE, and outer BM. Average thicknesses of EZ-RPE, EZ-BM, and RPE-BM were calculated from volumes and across nine ETDRS sectors. Manual segmentations were corrected by two masked expert graders and were compared using ICC. Dice coefficients (DC), Pearson correlation, and absolute thickness differences were used to assess agreement between automated and manual segmentation. Heat maps were generated to visualize thicknesses. Results: Thirty healthy eyes and thirty GA eyes were included in the analysis. Mean EZ–RPE, EZ–BM, and RPE–BM thicknesses were 47.55 ± 6.75 µm, 69.49 ± 6.92 µm, and 21.94 ± 3.46 µm, in the healthy eyes and 15.65 ± 11.09 µm, 39.18 ± 23.28 µm, and 23.52 ± 16.21 µm in GA eyes respectively. The model demonstrated high segmentation accuracy, with mean DC of 0.998 in healthy eyes and 0.995–0.998 in GA eyes. In healthy eyes, differences between automated and manual measurements were minimal (1.42 ± 3.39 μm (2.98%) for EZ–RPE, 1.31 ± 3.18 μm (1.88%) for EZ–BM, and 0.67 ± 1.71 μm (3.05%) for RPE–BM) which is within 1.88-3.05% from the gold standard (manual corrections), whereas GA eyes showed greater variability (mean differences of 3.61 ± 8.62 μm (23.06%) for EZ–RPE, 4.28 ± 11.34 μm (10.92%) for EZ–BM, and 4.4 ± 10.45 μm (18.71%) for RPE–BM). Heat maps revealed increased variability at the junctional zone surrounding atrophy. Automated and manual measurements showed strong correlations across all sectors in GA eyes (r = 0.97 for EZ–BM, 0.96 for EZ–RPE, and 0.89 for RPE–BM). Conclusions: The NMI-ORA enables accurate, automated segmentation and quantification of outer retinal layers, with performance comparable to expert graders.

Purpose: This study aimed to validate a deep learning algorithm for automated segmentation and quantitative assessment of the ellipsoid zone (EZ) and the retinal pigment epithelium (RPE)–Bruch’s membrane (BM) complex in healthy eyes and geographic atrophy (GA) eyes. Methods: In this retrospective study, spectral-domain optical coherence tomography (SD-OCT) volume scans from 30 healthy eyes and 30 eyes with GA were analyzed. An NMI-Outer Retina Analyzer was used to segment the inner EZ, inner RPE, and outer BM. Average thicknesses of the EZ-RPE, EZ-BM, and RPE-BM were calculated from volumes and across nine ETDRS sectors. Manual segmentations were corrected by two masked expert graders and compared using intraclass correlation coefficients (ICCs). Dice coefficients (DCs), Pearson correlation coefficients, and absolute thickness differences were used to assess agreement between automated and manual segmentation. Heat maps were generated to visualize thickness. Results: Thirty healthy eyes and thirty GA eyes were included in the analysis. Mean EZ-RPE, EZ-BM, and RPE-BM thicknesses were 47.55 ± 6.75 µm, 69.49 ± 6.92 µm, and 21.94 ± 3.46 µm, respectively, in the healthy eyes and 15.65 ± 11.09 µm, 39.18 ± 23.28 µm, and 23.52 ± 16.21 µm, respectively, in GA eyes. The model demonstrated high segmentation accuracy, with a mean DC of 0.998 in healthy eyes and 0.995–0.998 in GA eyes. In healthy eyes, differences between automated and manual measurements were minimal (1.42 ± 3.39 μm (2.98%) for EZ-RPE, 1.31 ± 3.18 μm (1.88%) for EZ-BM, and 0.67 ± 1.71 μm (3.05%) for RPE-BM), all within 1.88–3.05% of the gold standard (manual corrections). In contrast, GA eyes showed greater variability (mean differences of 3.61 ± 8.62 μm (23.06%) for EZ-RPE, 4.28 ± 11.34 μm (10.92%) for EZ-BM, and 4.4 ± 10.45 μm (18.71%) for RPE-BM). Heat maps revealed increased variability in the junctional zone surrounding the atrophy. Automated and manual measurements showed strong correlations across all sectors in GA eyes (r = 0.97 for EZ-BM, 0.96 for EZ-RPE, and 0.89 for RPE-BM). Conclusions: The NMI-ORA enables accurate, automated segmentation and quantification of outer retinal layers, with performance comparable to that of expert graders.

Background/Objectives: To evaluate the challenges and limitations of a deep learning model for automated choroidal boundary detection in eyes with geographic atrophy (GA) using spectral-domain OCT (SD-OCT), and to assess the workflow efficiency of an AI-assisted manual verification approach. Methods: In this retrospective study, total 5723 scans (Heidelberg Spectralis) with GA were analyzed. A previously validated tool (NMI ChoroidAI) was used to segment the choroidal inner (CIB) and outer (COB) boundaries. We compared the “AI-assisted” workflow (automated segmentation followed by manual verification) against “manual segmentation only” in terms of accuracy and time consumption. Slice-wise boundary errors were graded as 0 (accurate), 1 (≤33% deviation), 2 (33–66% deviation), or 3 (>66% deviation). Outcomes included error rates and weighted F₁ score (and precision where applicable). Total time for manual-only segmentation versus AI-assisted verification was recorded. -Interreader variability was assessed between the two readers using intraclass correlation coefficient. Results: For CIB, only 5.2% of B-scans showed any deviation (strictly accurate in 94.8%), with weighted F₁ score 0.97 and precision 1.00. COB was more error-prone: 19.0% of B-scans showed deviation; however, when minor deviations were considered acceptable, COB acceptability increased to 94.2% (i.e., 5.8% remained >33% deviated). Only 13.2% of B-scans required minor manual correction. For a 97-scan volume, processing time decreased from an average of 7 h (manual only) to 45 min (AI + human verification), an approximate 90% reduction in manual effort. Inter-reader agreement was high (ICC 0.923 for CIB and 0.938 for COB). Conclusions: Although the deep learning model exhibits limitations in COB detection due to artifacts, it serves as a valuable assistive tool. Our model substantially reduces human effort, but mandatory human verification is required to correct boundary errors caused by hyper-transmission before use in clinical trials.

Manufacturer-defined signal-strength indices are frequently employed as quality benchmarks for automated optical coherence tomography analysis, yet their empirical relationship with deep learning segmentation accuracy remains unclear. Because these metrics were originally developed for conventional image-processing pipelines, their ability to predict modern model-based segmentation accuracy has not been empirically validated. To address this gap, we evaluated the Heidelberg Spectralis Q-score against U-Net segmentation performance across 5,047 B-scans from 103 eyes for three anatomical boundaries of the posterior segment of the eye: the Ellipsoid Zone (EZ), Bruch’s Membrane (BM), and Choroid Outer Boundary (COB). Alongside standard boundary agreement metrics (MAE, MSE, Dice Similarity Coefficient), we adapted the Earth Mover’s Distance (EMD) from optimal transport theory as a boundary evaluation metric. Unlike column-wise averages, EMD quantifies boundary agreement as a 2-D geometric displacement, directly measuring residual spatial displacement between the model segmented boundary and the ground-truth boundary. Our results demonstrate that the Q-score — originally designed to gate image-processing-based automated analysis — is a poor predictor of deep learning boundary segmentation accuracy, with explained variance ( R ² ) failing to exceed 1.4% across all three boundaries. We further observed a monotonically increasing error hierarchy with anatomical depth (EZ < BM < COB), consistent across metrics, which is unexplained by the signal strength. At the COB, correlations were paradoxically positive, explained by a B-scan-level mediation chain: higher Q-scores correspond to greater choroidal thickness ( r = 0.113, ρ = 0.158), which in turn predicts higher COB segmentation error ( r = 0.165, ρ = 0.191) — a localization difficulty that global signal strength cannot capture. Collectively, these findings challenge the implicit assumption that signal-strength-based quality thresholds are a reliable proxy for deep learning model performance, and motivate a shift toward task-specific acquisition quality criteria calibrated to model performance rather than signal interpretability.