Yingwei Yang

and 5 more

Recently, UNet and its variants have been widely used in medical image segmentation for diagnosis. However, these algorithms usually have a large number of parameters and are computationally complex, which limits their applications on medical devices. In order to solve the above problems, we propose a lightweight medical image segmentation model DHR-Net. DHR-Net is mainly composed of three key components: DSDC, HA, and ResECA. DSDC extracts contextual features from data images in a staged, multi-layered manner. HA is able to better capture relevant features without additional resource consumption by combining Channel Attention, which recognizes the relevance of different features, and Spatial Attention, which locks the location of relevant features. ResECA acts as a bottleneck layer to further process the up-sampled features, and achieves more compact feature extraction through the cross-channel interaction strategy without dimensionality reduction. We performed comprehensive experiments on four publicly available medical datasets (Kvasir-Seg, DDTI, ISIC2018, and Synapse). For example, DHR-Net has the same level of parameter number and computational complexity as EGE-UNet, but has better performance on different datasets. Compared with UNext, DHR-Net not only has a lower number of parameters and computational complexity, but also better other metrics. Finally, it should be emphasized that DHR-Net is the first model that reduces the computational complexity to 0.06. Code at https://github.com/Phil-y/DHR-Net.