Deep learning-based optic disc classification is affected by optic-disc tilt

We aimed to determine the effect of optic disc tilt on deep learning-based optic disc classification. Image annotation was performed to label pathologic changes of the optic disc (normal, glaucomatous optic disc changes, disc swelling, and disc pallor) and note the appearance of a tilted optic disc (non-tilted versus tilted). Deep learning-based classification modeling was implemented to develop an optic-disc appearance classification. We acquired 2,507 fundus photographs from 2,236 subjects. Of the 2,507 data, 1,010 (40.3%) had tilted optic discs. The AUC of the models trained and tested using the non-tilted disc dataset was 0.988 ± 0.002, 0.991 ± 0.003, and 0.986 ± 0.003 for VGG16, VGG19, and DenseNet121, respectively. The AUC of the models trained and tested using the tilted disc dataset was 0.924 ± 0.046, 0.928 ± 0.017, and 0.935 ± 0.008. The model performance indicated by the AUC was better for non-tilted discs, regardless of the dataset used for training. In each pathologic change, non-tilted disc models showed better sensitivity than the tilted disc model. In the groups of glaucoma, disc pallor, and disc swelling, non-tilted disc models showed better specificity than the tilted disc model. We developed deep learning-based optic disc appearance classification systems using the fundus photographs of patients with and without tilted optic discs. The classification accuracy was lower in patients with the appearance of tilted discs compared to non-tilted discs, suggesting the need for identifying and adjusting for the effect of optic disc tilt on the optic disc classification algorithm in future development.