Automated Retinal Dysplasia Segmentation in Mouse Optical Coherence Tomography Scans Using a UNet-Based model

PurposeOptical coherence tomography (OCT) is the state-of-the-art non-invasive imaging technique for preclinical retinopathy studies. However, manual pathology annotations in mouse OCT scans are labour-intensive and susceptible to inter-rater variability. To alleviate these issues, we developed a neural network-based model for automated annotation of retinal dysplasia in mouse OCT scans. MethodsOur model was trained on 205 expert-annotated OCT stacks and validated on 40 unseen stacks with additional expert annotation comparisons in a subset of them. ResultsThe model detects pathologies with high accuracy (F1-score > 0.95) and consistency with experts (median Dice score > 0.8). We integrated the model into a cross-platform app ("OCTOPUS") that includes batch processing, automated annotation, manual annotation editing, and quantitative longitudinal tracking with pathological area estimation on the fundus image, and exportable results in CSV and SVG formats. ConclusionsOur open-source tool aims to facilitate and standardise OCT assessments between laboratories for efficient preclinical screening and high-throughput phenotyping in mouse models of retinal disease.