Parallel Network - A Deep Learning Approach for Blood Vessel Segmentation in Retinal fundus Images

In this modern era, computerized Retinal Blood Vessel (RVS) segmentation plays major role in diagnosis of various diseases like Diabetic Retinopathy (DR), Neovascularization, Hemorrhage. Early detection of retinal diseases can aid in the preservation of the patient’s vision. Deep learning based a new modified Convolution Neural Network (CNN) architecture is proposed for RVS. The Proposed architecture has two layers in which the layer one is for detecting the blood vessels of size thick and layer two for small vessel detection. Then the output of two layers has been combined to get the desired output. The proposed method is tested on the generally accepted public databases for this research field, DRIVE and STARE. In addition, various pre-processed methods are studied to investigate network performance enhancement. The pre-processing of the raw input image is much needed for better segmented output of blood vessels. In this work CLAHE, normalization and morphological operation of opening are done in the pre-processing stage. The proposed parallel architecture is then used to segment the retinal vessels. Accuracy, specificity, and sensitivity achieved with this proposed network are 98.02, 98.02, 88.04 respectively. Regardless of vessel thickness, the developed system performs better in terms of vessel extraction. The architecture can also be used to identify blood vessels that are frequently obstructed by factors such as lesions and hemorrhages, regardless of vessel thickness.