BVU-Net: A U-Net Modification by VGG-Batch Normalization for Retinal Blood Vessel Segmentation

Abstract

The study proposes a BVU-Net architecture that combines the advantages of VGG and U-Net. The VGG architecture has a smaller kernel size which speeds up the training process. VGG bears some resemblance to the Encoder portion of U-Net. In this study, the BVU-Net encoder uses the VGG architecture with the addition of batch normalization. The addition of batch normalization aims to help simplify weight initialization so that the training process is faster and reduces the risk of overfitting, while the decoder section still uses the U-Net architecture. BVU-Net is expected to be able to overcome the weakness of U-Net architecture in retinal image blood vessel segmentation. The performance results produced by BVU-Net on the DRIVE dataset are 96.36% accuracy, 78.71 sensitivity, 98.1% specificity, F1-score 78.9, and IOU 0.65. The results of BVU-Net performance on the STARE dataset are accuracy of 96.39%, sensitivity of 79.35%, specificity of 98.52%, F1-score of 77.16%, and IoU of 0.63. Based on these results indicate that BVU-Net has a better performance on the DRIVE dataset in detecting retinal blood vessels than STARE. This can be seen from the sensitivity value of DRIVE which is higher than STARE. The BVU-Net IoU results on DRIVE and STARE are greater than 0.5 and the F1-score above 70%, indicating that BVU-Net is capable and balanced in detecting the intersection area between blood vessels and the background on retinal images.

Retinal Vasculature Segmentation Based on Morphology and Pixel Level Classification

“…A. Desiani et al [25] proposed BVU-Net architecture that combines the advantages of VGG and U-Net for vessels segmentation from retinal images. Even though they used advanced deep learning methods, minor vessels determination is tough because of similar nature with background.…”

Blood Vessel Segmentation and Classification for Diabetic Retinopathy Grading Using Dandelion Optimization Algorithm with Deep Learning Model

“…This also has inspired the requirement to build an automatic diagnostic structure to help in initial DR diagnosing. Many trials are taken towards this track, and many methods founded on hand-crafted aspects have been suggested [5], which have demonstrated hopeful efficaciousness in perceiving DR areas in the images of the retinal fundus. Hand-crafted factors are generally implemented with conventional ML International Journal of Intelligent Engineering and Systems, Vol.…”