Enhanced medical image segmentation via dynamic and static attention aggregation

Enhanced medical image segmentation via dynamic and static attention aggregation

연구 분야: Strategies

학회: The Visual Computer

In numerous organ or diseased tissue segmentation tasks, normal tissues and diseased regions frequently appear concurrently, which can cause models to be confounded by the co-occurrence phenomenon. Many models based on the Transformer architecture struggle to capture contextual information effectively, and their overly intricate decoder structures may lead the model to become overly reliant on data distribution. This ultimately diminishes the model's generalization and robustness. To address these challenges, we propose a novel medical image segmentation method based on a dynamic and static attention aggregation mechanism, termed DSAEA-Net. Our approach integrates a dynamic and static attention aggregation module within the encoder, which extracts both dynamic and static feature information and constructs a convolutional aggregation branch to obtain comprehensive contextual information. An efficient feature aggregation layer further refines these features by employing multi-scale convolutions. In the decoder, we introduce an efficient upsampling block and a large-kernel convolution spatial activation layer to optimize the upsampling process and enhance the overall segmentation performance. Extensive experiments conducted on two public medical datasets, ISIC-2018 and TN3K, demonstrate the effectiveness of our proposed method. DSAEA-Net achieves improvements of 1.53% in mDice and 1.85% in mIoU on the ISIC-2018 dataset and 0.76% in mDice on the TN3K dataset, compared to state-of-the-art methods. These results underscore the critical significance of DSAEA-Net in the domain of image segmentation and offer novel insights for this category of tasks. Our codes can be found at https://github.com/MoSenTeam/DSAEA-Net.