Continuous Eye Disease Severity Evaluation System using Siamese Neural Networks

Abstract

Evaluating the severity of eye diseases using medical images is a very essential and routine task performed in medical diagnosis and treatment. Current grading systems which are largely based on discrete classification are unreliable and do reflect not the entire spectrum of eye disease severity. The unreliability of discrete classification systems for eye diseases is clear, as classification is subjective and done based on the personal opinion of various medical experts, which may vary. In a bid to solve these issues, this study proposes a system for determining the severity of eye diseases on a continuous range using a twin-convoluted neural network approach known as Siamese Neural Networks. This system is demonstrated in the domain of diabetic retinopathy. Samples of retinal fundus images from an eye clinic in India are taken as test cases to evaluate the performance of a Siamese Triplet network which attempts to find the distance between their image embedding. The outputs of the Siamese network when a reference image is juxtaposed with a collection of images with distant severity categories (negative images), as well as when two reference images are compared to each other, are found to have a positive correlation (95\%) with originally assigned severity classes. Hence, these outputs indicate a continuous range of the severity and change in eye diseases.

DFU-Helper: An Innovative Framework for Longitudinal Diabetic Foot Ulcer Diseases Evaluation Using Deep Learning

“…As far as the architecture of our SNN is concerned, as opposed to all the work detailed in the related works sections, including those for severity evaluation over a time period [24,26,27], none uses an ensemble of CNN and ViT for the sub-networks. For CNN, we use EfficientNet, which is known to achieve state-of-the-art results while dealing with images, the novel and widely used ViT, and more specifically, the BEiTV2 [59].…”

“…In the selection of anchors for image comparison over time in DFU-Helper, we took a different approach compared to AbdulRaheem et al [26] and Akbar et al [27]. Instead of using 5 or 16 images per class, we utilized the maximum number of available images per class in our dataset to generate the class anchors.…”

“…In a recent study by AbdulRaheem et al [26], the authors propose the use of Siamese Neural Networks (SNNs) for the continuous evaluation of eye disease severity. Similar to the work of Li et al [24] mentioned earlier, they employ a twin-CNN architecture.…”