Siamese network based fine grained classification for Diabetic Retinopathy grading. (September 2022)

• Record Type: Journal Article
• Authors: Nirthika, Rajendran; Manivannan, Siyamalan; Ramanan, Amirthalingam
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: Diabetic Retinopathy grading -- Deep learning for retinal image classification -- Loss function for ordinal classification; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103874
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Diabetic Retinopathy (DR) is a complication of diabetes which affects the retina. Early and the correct identification of the DR is necessary to determine the appropriate treatment. Automated DR grading is a fine-grained classification problem as some of the lesions are very small in size and are difficult to distinguish from non-DR regions. In this work, we propose a Siamese network based Convolutional Neural Network architecture for DR grading, which aims to improve the single eye based DR grading performance by incorporating the single eye based features with the patient-level DR features (i.e., features extracted from both eyes of the patient). To capture the patient-level DR information we propose two approaches: one is based on Bilinear pooling which aims to capture higher order statistical information by considering the correlation between the eyes of a patient, and the other one is based on simply averaging the features from both eyes. We experimentally show that both of the proposed approaches perform better than other approaches proposed in the literature for DR grading. In addition, as DR grading is an ordinal classification problem, we investigated the effect of different loss functions including the widely used Cross Entropy loss, Quadratic Weighted Kappa loss, Mean Squared Error Loss, and Ordinal Regression based loss, and show that Mean Squared Error loss gives the better system–annotator agreement. On the challenging, large scale, public Kaggle EyePACS dataset (consists of 88, 702 images) our proposed approach achieves a Kappa score of 0.86 and an Accuracy value of 84.6%, indicating that the proposed approach is the new state-of-the-art. Highlights: A novel Siamese network based CNN architecture is proposed for DR grading. Bilinear-pooling and averaging-based approaches were proposed to capture DR features. Different loss functions were investigated for DR grading.