Abstract
Purpose :
Glaucoma, diabetic retinopathy (DR) and age-related macular degeneration (AMD) are leading causes of blindness. Computer-assisted automated screening of colour fundus photographs could be a method by which early diagnosis and treatment to prevent blindness can be achieved on a population level. While previous studies have trained convolutional neural networks to detect individual eye diseases, a fit-for-purpose classifier must be capable of screening for multiple diseases. To address this, we trained a network to classify colour fundus photographs between healthy, glaucoma, DR and AMD by leveraging several existing public datasets.
Methods :
We amalgamated 10 public datasets (DiaretDB, Drishti-GS, DRIVE, HRF, IDRiD, 39Kaggle, MESSIDOR, ORIGA-light, REFUGE, STARE) to create a multi-disease fundus dataset containing 4205 images, with approximately 58%, 9%, 32%, 1% images labeled as healthy, glaucoma, DR and AMD respectively. To train our classifier, we fine-tuned all layers of an ‘Inception v3’ neural network model, pretrained on ImageNet, on 80% of the dataset using data augmentation (horizontal mirroring, random scaling and cropping) and weighted cross-entropy loss to tackle the class imbalance.
Results :
Validating on the remaining 20% of the images, we found a mean area under the curve (AUC) of 95.6% (healthy: 92.7%, glaucoma: 93.4%, DR: 96.5%, AMD: 99.5%). Our model stabilized within 20 epochs, and required approximately 30 minutes to train on a single graphics processing unit (GPU).
Conclusions :
We designed an accurate eye disease classification model for healthy, glaucoma, DR and AMD fundus images, despite the modest dataset size. To our knowledge, this is the first reported multi-disease classifier for colour fundus images. We also created a user-friendly pilot interface for clinicians to access the classifier, and test their own images for research purposes (link provided on poster).
This is a 2020 ARVO Annual Meeting abstract.