Automatic Kidney Parenchyma Segmentation Based on Improved UNeXt Model

In recent years, image segmentation techniques based on deep learning have achieved good results in medical image processing, while there are still challenges in segmenting kidney in nephrohic grapimages. Due to the problems of poor image quality, blurred boundaries, redundant information in kidney map images, existing networks are difficult to achieve accurate segmentation. Thus, this paper proposes a medical image segmentation network based on the improved UNeXt model for automatic segmentation of renogram images. Based on this, this paper takes UNeXt network as the benchmark model by embedding multiscale information extraction module together with feature attention module unit to improve the shortcomings of the benchmark network model. we choose to introduce a multiscale information extraction module to better obtain the multiscale information of the image, embedding the feature attention module unit to make the network better focus on the segmentation region. To verify the effectiveness of the network, comparative experiments are conducted on the kidney map dataset provided by the hospital and the skin lesion segmentation dataset of the open challenge ISIC (2018), according to the experimental results, the segmentation accuracy (IoU) of the proposed network in this paper is improved on both types of datasets compared to the benchmark network UNeXt. Compared with other classical semantic segmentation networks, the network in this paper also achieves better The results are also better than other classical semantic segmentation networks. The effectiveness of the introduced module is also verified by conducting corresponding ablation experiments in this paper.