Abstract
Automatic ophthalmic screening systems for a small number of common eye diseases have seen an impressive improvement in performance in recent years, largely due to the wide spread adoption of deep neural network (DNN) based image classification models in medical domains. In contrast, robustness and generalizability often lacks in applications relating to rare eye disease. A major challenge and cause of such disparity in performance is the scarcity of patient data due to patient privacy concerns, disease rarity, and healthcare accessibility. Common ocular diseases, particularly diabetic retinopathy, cataracts, and glaucoma, have been thoroughly studied on account of having a sufficient amount of publicly available data for large-scale analysis from observed cases in clinical settings. However, the majority of other ocular diseases that are studied in statistical and deep learning classification applications lack abundant data samples. This issue can lead to severe overfitting, especially in traditional DNN models. In this paper, we propose a few-shot image classification method to classify rare eye disease consisting of an ensemble of ImageNet pretrained feature extractors and a simple classifier head (support vector machine). In our most challenging 5-way classification problem, the experimental results show that the ensemble model exceeds the best performing single feature extractor model by over 4% and the worst by over 15% in terms of accuracy. Additionally, the ensemble method obtains a performance lower bound in all of our experiments and can easily be incorporated into concurrent state of the art few-shot classification approaches.