SIB-UNet: A dual encoder medical image segmentation model with selective fusion and information bottleneck fusion

Medical image segmentation aims to accurately mark the lesion area in the image to assist doctors in disease diagnosis and guidance of surgical operations. However, the shape and size of lesions in medical images often have large uncertainties, which directly affects the accuracy of segmentation results. Currently, many models try to solve this problem by fusing global and local features. However, these methods generally face the problems of large number of model parameters and too simple feature fusion methods. In order to effectively solve the above problems, we propose the SIB-UNet model. The model dynamically fuses global and local features through a Selective fusion module. In addition, we also added an Information Bottleneck module to the skip connection of the model to eliminate the interference of irrelevant features. Experimental verification on three public datasets shows that compared with the existing state-of-the-art models, SIB-UNet not only reduces the number of parameters, but also achieves better segmentation results. •Dual encoder structure, including selective fusion module, can dynamically fuse global and local features.•The skip connection structure based on information bottleneck theory can narrow the semantic gap.•Obtain better segmentation results with less resource consumption.