Grading diabetic retinopathy using multiresolution based CNN

Diabetic Retinopathy (DR) refers to a medical condition that affects the eye; it occurs due to diabetes, and, if not detected early on, results in a reduction of visual capacity and may even result in blindness. The process of diagnosing DR manually by ophthalmologists can be both time-consuming and expensive, and there is a risk of misdiagnosis. On the other hand, computer-aided diagnostic systems can provide a more accurate and efficient diagnosis and can help ophthalmic specialists by offering a second opinion for effective treatments. This paper presents a method in which the feature extraction is based on the multiresolution-based decomposition of Discrete Wavelet Transform (DWT), and classification is performed using Convolutional Neural Network (CNN) for grading DR images. The suggested approach begins by utilizing Contrast Limited Adaptive Histogram Equalization (CLAHE) to refine the contrast level of the fundus images as a pre-processing technique. The images in the datasets (IDRiD, DDR, and EyePACS) are not balanced, so oversampling is applied to ensure that an equal number of images from each of the grade categories are present during the training process. The results of the experiments demonstrate that the proposed model achieves a classification accuracy of 90.07%, 96.20%, and 93.53% for all DR stages, outperforming existing models on all the three datasets. Therefore, the proposed method offers a superior alternative to current approaches.