MR-Unet: Modified Recurrent Unet for Medical Image Segmentation

In recent years, there has been significant interest in medical image segmentation. Traditional methods are being surpassed by deep learning, which has demonstrated their superiority. However, one drawback of using standard convolutional neural networks (CNNs) in deep learning is their large model size, which leads to excessive memory consumption and longer computation time. To address this issue, a novel model called MR-Unet was introduced in this study. MR-Unet combines the advantages of Recurrent Convolutional Neural Networks (RCNN) with the multiple layers Unet (Un-Net) architecture. By incorporating RCNN and leveraging all output feature maps in the convolution units of the network nodes, the overall network size was reduced. To assess the effectiveness of our proposed model, we conducted experiments on liver segmentation using the LiTS 2017 dataset, spleen segmentation using the Medical Segmentation Decathlon Challenge 2018 datasets, and skin lesion segmentation on dataset is supplied by the ISIC-2018 Challenge. The experimental results clearly demonstrate that our model not only achieves smaller size but also improves performance compared to existing models.