AMMU-Net: Automatic Multi-Scale Feature Selection Network Based on Mixture of Experts for Lung Nodule Segmentation

In the field of medical image processing, accurate segmentation of lung nodules is crucial for the early diagnosis of lung cancer. Precise segmentation of lung nodules typically requires the network to have an efficient multi-scale feature extraction capability. Furthermore, optimizing the multi-scale feature extraction process and achieving automatic selection of scale features is a key direction in current research. Some existing studies have not focused on the selection of scale features, and there is still a lack of research on lung nodule segmentation tasks. Considering this, we have designed a novel Automatic Multi-Scale Feature Selection Network Based on Mixture of Experts (AMMU-Net). This method integrates MoE based on Convolution and Transformer, respectively implementing multiple pooling and multi-scale Swin Transformer block, to mine multi-scale features and select better optimized parts. Extensive experimental results on the Luna16 dataset show that our method surpasses existing advanced models in the performance of lung nodule segmentation tasks, achieving a Dice score of 87.67%, while maintaining optimized parameter quantity.