DIM-UNet: Boosting medical image segmentation via diffusion models and information bottleneck theory mixed with MLP

In recent years, UNet and its latest extensions, such as UNeXt and TransUNet, have become the leading medical image segmentation methods. However, medical image usually contain a high amount of noise and a small number of samples, making it difficult for the models to learn features accurately, and potentially leading to overfitting. To address these challenges, we propose a DIM-UNet model based on Diffusion models, Information bottleneck theory, and MLP. DIM-UNet introduces two key modules: the Diffusion-MLP module and the IB-MLP module. The Diffusion-MLP module can de-noise the feature map while capturing global features by combining the ideas of diffusion models. The IB-MLP module is located at the bottom of the model, using information bottleneck theory to compress the learned features. This module can retain the most relevant features to the target task and discard irrelevant features to improve the model’s generalization ability. We compare DIM-UNet with state-of-the-art models on three public datasets and achieve competitive segmentation results. •We propose a Diffusion-MLP module that can denoise feature maps while capturing global features.•We propose an IB-MLP module that compresses irrelevant and retains only relevant features.•We studied the impact of various parameters in Diffusion-MLP and IB-MLP on model performance.