Deep Feature‐Based Normality Modeling for Automated Out‐Of‐Distribution Detection in Sheep Retinal Fundus Images
Büşra Kibar, Sıtkıcan Okur, Büşra Baykal, Taner Arslan, Çağlar Özkalipçi, Berfin PetekABSTRACT
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.