Abstract
Purpose :
Existing automated retinal image quality assessment (RIQA) models, commonly trained on diabetic retinopathy (DR) images, have limited applicability to neuro-ophthalmic conditions (e.g., papilledema, ischemic optic neuropathy (ION), etc.). Their inconsistent performance is due to the appearance of the abnormal optic disc (swelling, atrophy, etc). Hence, our aim was to develop a new automated RIQA system that can predict, without human intervention, if a fundus image has an acceptable image quality for subsequent evaluation of the optic nerve head (ONH).
Methods :
A total of 2,082 fundus images obtained from the EyeQ, a publicly available DR data set, and 5,208 images collected within the Brain and Optic Nerve Study with Artificial Intelligence (BONSAI) Consortium data set (total of 486 ION, 1,644 papilledema, 832 optic atrophy, 559 DR, and 3,769 normal fundoscopic images) were segmented for their ONH region and given quality-labels by a trained human classifier. The data set was divided proportionally into images used for training (80%) or testing (20%). A dedicated deep-learning system (EfficientNet-B5 CNN), was trained to automatically perform a binary quality classification of retinal images (i.e. “Acceptable” or “Rejected”). The classification performance of our RIQA was evaluated by calculating the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.
Results :
Using 5-fold cross-validation on the training data set, our model classified ONH segmented fundus images into “Acceptable” vs “Rejected”, yielding an average AUC of 0.982, accuracy of 94.2%, sensitivity of 93.3%, and specificity of 94.6%. When tested against the testing data set for external validation, our model achieved an average AUC of 0.982, accuracy of 92.6%, sensitivity of 93.9%, and of specificity 92%.
Conclusions :
An automated deep learning system trained on digital color images of normal and abnormal optic discs due to various optic neuropathies can discriminate between acceptable and poor image quality. Further developments will aim to provide instantaneous image quality evaluation coupled to fundus cameras and/or independent deep learning-based diagnostic algorithms.
This is a 2021 ARVO Annual Meeting abstract.