ResMU-Net: Residual Multi-kernel U-Net for blood vessel segmentation in retinal fundus images

Accurate and early diagnosis of eye disease can save human vision. Blood vessel segmentation plays a vital role in the detection of vision-threatening vascular diseases. It is challenging to segment blood vessels in the presence of bright and dark lesions. A modified UNet-based model is proposed to improve the system’s overall performance. The Residual Multi-kernel UNet (ResMUNet) model is combination of the residual network at the encoder and Multi-kernel Dilation Convolution (MKDC). The ResMUNet model enables the construction deeper network without affecting gradient descent issues. Using multiple dilated convolutions, the proposed network captures more information, improves basic features, and increases its field of view. Also, necessary features are obtained by data augmentation technique. The data augmentation technique maintains the feature requirement by providing sufficient data for training and validation. The benchmark datasets DRIVE, STARE, CHASE_DB and HRF were utilized in this experiment for training and assessing the model. The results clearly indicate that, in comparison to the previous approaches, our method provides comparable performances. Moreover, small blood vessels in a retinal fundus image are challenging to segment, but our model performs better as a result. However, our model achieved AUC of 0.9842/0.9863/0.9714, accuracy of 0.9685/0.9756/0.9704 and F1 score of 0.8149/0.8310/0.7794 on DRIVE, CHASE and HRF respectively. The proposed model performed well when evaluated on different datasets, demonstrating its robustness. •Effective segmentation of blood vessels using Residual MultiKernel U-Net Model.•Utilizing residual blocks enabled the construction of deeper networks.•Extending rich information using Multi-Kernel Dilation.•The data augmentation reduced size of input images while maintaining the features.•Experiment performed on various modalities datasets like DRIVE, STARE, CHASE and HRF.