Abstract
Purpose :
Deep learning (DL) has been showing great promise in ophthalmology, especially in diagnosis of referable diabetic retinopathy (DR). To date, these advances have always been done using clinically collected datasets. However, the real power of DL is in moving the experience of highly qualified ophthalmologists to community settings, especially in rural and developing areas. The images collected in non-clinical settings by field workers without clinical training present worse quality and high degree of image variance due to acquisition set-ups and illumination. DL systems can contribute to create cost-effective diagnostic models in non-clinical settings. In this study we developed a DL based system for the classification of referable DR using hand held non-mydriatic retinal photographs.
Methods :
Non-mydriatic retinal photographs were collected as part of the ORNATE India Project, an ongoing community-based study to test novel diabetic screening services. Retinal photographs were acquired by field workers as part of door-door and point of care non-clinical screening using Zeiss Visuscout® 100 mobile fundus camera and graded by trained ophthalmologists. A total of 6,655 images (11.3% referral samples) were used to train and validate the DL system. Due to the highly unbalanced distribution of the data, the testing set was formed by 40% of the referral samples and the same number of non-referral samples. An Inception-ResNet V2 model was fined tuned on augmented and oversampled training data to balance class distribution.
Results :
In the validation set, precision, recall and AUC (Area Under the Curve) for referable DR achieved 75.74%, 82.79% and 86.15%. Due to the highly imbalanced dataset, results benefit from imbalanced data augmentation and oversampling to train the model under an apparently balanced distribution. Highly regularized training set-ups, using techniques such as L2 regularization and dropout, improved performance by reducing the model tendency to overfit training data.
Conclusions :
The DL system demonstrated promising results in detecting referral DR suspects using non-mydriatic retinal images. We showed that artificial intelligence can be used to create grading algorithms using data collected at non-clinical settings by field workers. These support future work on the development of automated screening programs for diagnosis of DR in special populations or rural areas.
This is a 2020 ARVO Annual Meeting abstract.