Abstract: Objective To evaluate the accuracy and feasibility of using YOLOv8 model to assist in diagnosing oblique muscle function. Methods Fundus photographs of abnormal oblique muscle function and normal subjects were collected from Shenzhen Eye Hospital. The YOLOv8 model was used for training, and the macular areas and optic disc area in fundus photographs were marked and identified, and the disc-fovea angle(DFA)of the macular fovea was quantitatively detected to evaluate the oblique muscle function. Results There were no significant differences in gender(χ2=0.478, P=0.489)and age(U=358 91.5, P=0.770)between the abnormal oblique muscle function group and the normal group. The results of the probability density curve showed that there were differences in DFA distribution between the abnormal oblique muscle function group and the normal one. The YOLOv8 model demonstrated a 100% recognition accuracy for the optic disc category and a 95% recognition accuracy for the macular category. The model exhibited superior recognition performance for the optic disc category in comparison to the macular category across diverse confidence thresholds(for the overall category, the F1 score of the model was the highest, at 0.92, when the confidence was 0.307). Independent sample t test, Mann-Whitney U test, Kolmogorov-Smirnov test and Levene's test all showed that there were significant differences in DFA between abnormal and normal oblique muscle function groups based on YOLOv8 model(all P<0.001). When DFA<-17.286 or DFA>6.278, the probability of DFA distribution results belonging to the category of oblique muscle dysfunction is greater than 80%. Conclusion The application of YOLOv8 model can assist clinical evaluation of oblique muscle function. When DFA is less than-17.286° or greater than 6.278°, the probability of abnormal oblique muscle function is greater than 80%.

Key words: Oblique muscle function, Fundus photography, Disc-fovea angle, YOLOv8 model, Machine learning