Optimized KiU-Net: Lightweight Convolutional Neural Network for Retinal Vessel Segmentation in Medical Images

Medical image segmentation helps with computer-assisted disease analysis, operations, and therapy. Blood vessel segmentation is very important for the diagnosis and treatment of different diseases. Lately, the U-Net and KiU-Net based vessel segmentation techniques have demonstrated reasonable achievements. The U-Net architecture belongs to the group of undercomplete autoencoders which ignores the semantic features of the thin and low contrast vessels. On the other hand, the KiU-Net uses a combination of undercomplete and overcomplete architectures to segment the small structure and fine edges better than U-Net. However, this solution is still not accurate enough and computationally complex. We propose an Optimized KiU-Net model to increase the segmentation accuracy of thin and low-contrast blood vessels and improve the computational efficiency of this lightweight network. The proposed model selects the ideal length of the encoder and the number of convolutional channels. Moreover, our proposed model has better convergence and uses a smaller number of parameters by combining the feature map at the final layer instead at each block. Our proposed network outperforms the KiU-Net on vessel segmentation in the RITE dataset. It obtained an overall enhancement of about 4% in terms of F1 score and 6% in terms of IoU compared to KiU-Net. Evaluation and comparison were also conducted on the GLASS dataset, and the results show that the proposed model is effective.