Retinal blood vessels segmentation using CNN algorithm

Abstract

The precise identification of blood vessels in fundus is crucial for diagnosing fundus diseases. In order to address the issues of inaccurate segmentation and low precision in conventional retinal image analysis for segmentation methods, a new approach was developed.The suggested method merges the U-Net and Dense-Net approaches and aims to enhance vascular feature information. To achieve this, the method employs several techniques such asHistogram equalization with limited contrast enhancement, median filtering, normalization of data, and morphological transformation. Furthermore, to correct artifacts, the method utilizes adaptive gamma correction. Next, randomly selected image blocks are utilized as training data to expand the data and enhance the generalization capability. The Dice loss function was optimized using stochastic gradient descent to improve the accuracy of segmentation, and ultimately, the Dense-U-net model was used for performing the segmentation. The algorithm achieved specificity, accuracy, sensitivity, and AUC of 0.9896, 0.9698, 0.7931, and 0.8946 respectively, indicating significant improvement in vessel segmentation accuracy, particularly in identifying small vessels.