LIAUNet: Rethinking the Application of Loss Information in Medical Image Segmentation

In recent years, with the development of medical imaging and deep learning technologies, medical image segmentation has played a crucial role in assisting physicians with diagnosis and treatment. Networks based on encoder-decoder architectures have been widely used in the design of medical image segmentation models. However, a large amount of information, including critical segmentation detail information, is lost in the process of encoder down-sampling. In order to rationally utilize this loss information, this paper proposes a novel network: Loss Information Aggregation Network (LIAUNet), which utilizes loss information and rationally aggregates it. Specifically, a MultiScale Hybrid Pooling Loss Extraction Module (MHPLE) is firstly designed to extract the loss information of four scales in the down-sampling process and to supplement and adaptively adjust it. In addition, we design the Information-Preserving multiscale Attention Aggregator (IMAA) to aggregate the extracted loss information through a multiscale approach, which enhances the features and reduces further loss or sparsity of the loss information. Finally, we add a Spectral-Gated Convolution Block (SGCB) after encoder down-sampling to enhance the network's ability to jointly model frequency and spatial domain features. Experiments on multiple medical image datasets show that, compared with other methods, our method has higher segmentation accuracy. This method provides effective support for the advancement of medical imaging and clinical diagnosis.