Automated diagnostic classification of diabetic retinopathy with microvascular structure of fundus images using deep learning method

Abstract

Multi-class classification is a major concern in the research field, especially in medical image analysis. This work proposes a novel method for automatically segmenting the blood vessels and classifying Diabetic Retinopathy (DR) using fundus images. This system helps ophthalmologists with the early detection and grading of DR diseases. The main aim here is to detect any pathological changes happening in the retinal vascular structure for the development of DR, through which the patients can avoid undergoing expensive scans. The proposed system involves three different stages, including pre-processing, vessel segmentation, and classification. The input images are processed first to eliminate the noise, followed by green channel extraction and enhancement with Contrast Limited Adaptive Histogram Equalization (CLAHE) and gamma correction. Retinal Vascular Structure (RVS) segmentation is a major concern in this work as it is responsible for detecting the different stages of DR by detecting the presence of microaneurysms, haemorrhages, and exudates. The U-Net is used as a base architecture to develop the segmentation model. The contracting path in the U-Net contains four consecutive downsampling and upsampling layers with skip connections. After performing downsampling four times, information on the tiny blood vessels may be missed. Therefore, ResEAD2Net is introduced in this work, where the number of downsampling and upsampling layers is reduced to two instead of four, and two such contracting paths and expansion paths are added to the network. Thus, detailed semantic information can be retained with this structure. Residual blocks are included instead of convolution blocks to increase the computational speed. To include structural connectivity, the segmented output is passed through the proposed Regularized Random Walker (RRW) algorithm, which focuses on the broken blood vessels. Finally, the features are extracted from the vessel structure and passed through the Machine Learning (ML) classifier to predict the DR grading. The proposed method achieves better performance in segmentation with accuracy, sensitivity, specificity, and area under the curve values of 98.07%, 90.24%, 99.01%, and 97.51%, respectively, for the STARE dataset and 97.55, 90.07, 98.01, and 97.64, respectively, for the DRIVE dataset. It is also found that the features obtained from the segmentation and ML classifier outperforms the existing methods in multiclass classification by achieving accuracy, sensitivity, and specificity of 98.88%, 98.91%, and 98.29% with the Messidor-2 dataset.