Abstract
The advancement of deep learning theory and infrastructure is crucial in the progress of automatic segmentation techniques. Compared with traditional segmentation methods, automatic segmentation methods have considerable strengths such as convenience, accuracy, and so on. However, the drawbacks cannot be neglected. In the laboratory environment, most of the segmentation frameworks are based on deep learning at the cost of sacrificing the lightweight network architecture, adding a lot of parameters in the network to trade for excellent segmentation accuracy. In practical clinical applications, the lack of high computing performance (GPU) machines to maintain operational efficiency poses a huge challenge for the migration from laboratory to clinic. Recently, an alternative to the CNN and Transformer frameworks has been enthusiastically touted, with MLP-based network parameters being significantly decreased as all parameters are learned in the linear layer of the MLP and generate striking outcomes similar to both. Inspired by the MLP-based framework, we recommend leveraging the MS-DC-UNeXt as an alternative solution for medical image segmentation, which is mainly composed of Tokenized MLP block, Dual Channel block(DC-block), and Bottleneck (Res-ASPP). Please refer to the paper for the complete abstract
Abstract (translated)
URL
https://arxiv.org/abs/2210.12361
