BLOOD VESSEL SEGMENTATION IN RETINAL IMAGES USING RESVNET ARCHITECTURE

The U-Net architecture is often used in medical blood vessel segmentation due to its ability to produce good segmentation. However, U-Net has high complexity due to the presence of the bridge part, which increases the parameters and training time. To overcome this, this research modifies U-Net by removing the bridge part, resulting in V-Net architecture. V-Net architecture faces challenges in capturing deep and complex features. This research proposes modifying V-Net with ResNet architecture in the encoder part, resulting in ResVNet architecture. ResNet, with residual connections, enables the training of very deep networks with more stability and effectiveness in capturing complex features. At the encoder, ResNet is used for more effective training of deep networks and capturing complex features. While at the decoder, U-Net is used to preserve the high resolution and spatial information of the image in segmentation. This study aims to determine the performance evaluation results of the ResVNet architecture. The evaluation measures used are accuracy, sensitivity, precision and Jaccard score. Tests were conducted on the DRIVE and STARE datasets. The measurement results of blood vessel segmentation using ResVNet on the DRIVE dataset resulted in accuracy 96.57%, sensitivity 82.28%, precision 79.57%, and Jaccard score 67.61%. On the STARE dataset, the accuracy results are 96.71%, sensitivity 79.44%, precission 79.44%, and Jaccard score 65.05%. The sensitivity results on the STARE dataset as well as the precision and Jaccard score values on the two datasets produced are still low, in the future this research will make improvements to the ResVNet architecture used.