Lesion-Aware Binary Diabetic Retinopathy Detection in an Imbalanced Fundus Image Dataset

Abstract

Diabetic retinopathy (DR) is one of the leading causes of vision impairment in the world. Early diagnosis is needed to prevent permanent vision loss. Despite being shown to perform well on benchmark DR datasets, class imbalance and loss of smaller lesion features in preprocessing often lead to performance degradation of convolutional neural networks (CNNs) on real-world, imbalanced, multi-disease datasets. Precise lesion identification is particularly important is particularly important in multi-disease datasets, in which DR-related lesions can be small and easily overlooked, since it has a direct impact on sensitivity and clinical reliability. This work uses the imbalanced multi-disease Retinal Fundus Multi-disease Image Dataset (RFMiD) to evaluate a baseline CNN pipeline, originally designed on the balanced Messidor dataset. We propose two targeted augmentation techniques: rotation and photometric transformations (contrast adjustment, sharpening, color shifting), along with an improved preprocessing pipeline that incorporates Difference of Gaussian (DoG) and Dilated DoG (DDoG) to address decreased sensitivity and improve lesion visibility. The proposed lesion-aware preprocessing approach was tested using six different pretrained CNN architectures, such as AlexNet, ResNet-18, VGGNet-S, VGGNet-16, VGGNet-19, and GoogleNet. The application of rotation augmentation along with the proposed preprocessing pipeline resulted in considerable improvements in the performance metrics for each network architecture, where sensitivity increased to 95.16% (VGGNet-19), showing an increase of 13.7%. At the same time, the increase in accuracy (91.72%) and AUC (0.9605) values also increased. Moreover, the results obtained through photometric augmentation and proposed preprocessing showed significant improvement in the robustness of the models, with maximum sensitivity recorded as 91.94% (VGGNet-16), increased by 8.9%, and AUC scores increasing to 0.9646 (VGGNet-19). The PSNR value of 31.26dB and SSIM of 0.88 further confirm the efficacy of the proposed preprocessing pipeline. In summary, lesion-aware image preprocessing alongside rotations and photometric data augmentation methods has significantly increased the accuracy of various pretrained CNN models in diagnosing diabetic retinopathy.