Deep Feature-Based Normality Modeling for Automated Out-Of-Distribution Detection in Sheep Retinal Fundus Images

Objective: To develop and evaluate a deep feature-based normality modeling approach for automated out-of-distribution (OOD) detection in sheep retinal fundus images. Animals Studied: Retinal fundus images from 75 adult sheep (n = 271 images) and additional OOD images from non-target species (cattle, dogs, and cats; n = 346 images). Procedures: Deep feature embeddings were extracted using a ResNet50 convolutional neural network pretrained on ImageNet. Normal retinal appearance was modeled in feature space using healthy images. Anomaly scores were calculated using a k-nearest neighbor (k = 5) distance-based approach. OOD detection was evaluated using receiver operating characteristic (ROC) analysis. The anomaly threshold was defined as the 95th percentile of validation scores. Results: The proposed framework demonstrated a clear separation between in-distribution sheep retinal images and OOD samples. The model achieved an area under the ROC curve of 1.00 (95% CI: 0.99–1.00). At the predefined threshold, all OOD images were correctly identified (100% detection rate), with a false alarm rate of 11.9% in the sheep test set. Conclusions: Deep feature-based normality modeling can characterize normal sheep retinal morphology and identify out-of-distribution samples. The proposed approach should be interpreted as a proof-of-concept screening and quality-control tool, not a disease-specific diagnostic system. Further validation using animal-level partitioning and intra-species pathological datasets is required to establish clinical utility. Normality modeling may hold promise for future screening of retinal disease within sheep; however, this has not yet been evaluated. Importantly, this study evaluates only cross-species OOD detection and does not assess the detection of retinal abnormalities within sheep.