Purpose: To evaluate the performance of state-of-the-art semantic segmentation methods on OCT data for age-related macular degeneration (AMD). We measured variability between annotators to quantify differences in ground truth arising from personal bias.

Methods: From 94 patients suffering from exudative neovascular AMD (nAMD), 24 volume scans (49 slices each) were selected. Trained members of a reading center for AMD created pixel-wise masks for 12 retinal layers and two pathological labels (fluid, hyperreflective material) to benchmark two-dimensional (2D) and three-dimensional (3D) segmentation models on clinical data. Models were evaluated using fivefold cross-validation, and the best model was used to quantify errors between ground truth and predictions.

Results: The nnU-Net (3D) achieves the best segmentation performance (mean Dice similarity coefficient [DSC] of 0.907), leaving a theoretical gap of 0.036 DSC to the mean interrater agreement, which is the upper bound of model performances. Comparing the volumes calculated for each structure using the model masks with the ground truth produced an average error of 0.065 mm3.

Conclusions: Models like nnU-Net can produce high-quality 3D masks, challenging the conventional reliance on 2D slices for optimal performance. Both DSC and low average errors indicate that such a model is fit for the large-scale analysis of cohorts.

Translational relevance: The presented approach can streamline clinical workflows by reducing the time and effort required for manual annotations, ultimately supporting more efficient and accurate monitoring of AMD progression and treatment response. We provide open-source access to the model weights, annotation instructions and sample data.