所属单位：计算机科学与工程学院

发表刊物：2023 4th International Conference on Computer Vision, Image and Deep Learning

项目来源：自选课题

关键字：-component; neural network; image segmentation; semantic segmentation

摘要：Medical image segmentation technology is an impo rtant foundation for the development of modern medical technolo gy and has significant implications for the diagnosis and treatmen t of diseases. For many years, convolutional neural networks (CN N) have dominated computer vision, especially innovative neural networks with higher architectures, such as U-shaped models and pyramid pooling techniques, which can be applied to various me dical image segmentation tasks. However, due to the limitation of receptive field in convolution, CNN can not capture long distance context information and lacks spatial interactivity, which is beyo nd doubt. On the other hand, although Transformer has the glob al Receptive field capture feature, it does not capture local featur es as well as CNN. In this article, we propose MUNformer, a mult i-scale global semantic feature transformer-based extractor and Unet-based architecture, a type of Transformer that focuses on m ulti-scale network structures for semantic segmentation. MUNfor mer has two attractive features: 1) MUNformer extracts features at different levels from two feature extractors in the encoder part to enrich Semantic information. 2) The decoder interacts with th e features extracted by two different feature extractors in the enc oder (features at the corresponding level), combines local attentio n and global attention to present a powerful representation, and f inally performs feature stitching to obtain more abundant Seman tic information. We believe that the combination of CNN and Tra nsformer can extract powerful feature representations through m ulti-level interaction, utilizing the advantages of both. MUNforme r performs better than various competitive methods in different medical applications, including multi organ segmentation and he art segmentation.

合写作者：徐子奇,鲍学良,仲昭昊

第一作者：刘冰

论文类型：论文集

页面范围：2

是否译文：否

发表时间：2023-05-14