SCEA-Net: A hybrid framework from spatial-channel-aware external attention for accurate 3D medical image segmentation

Abstract

Accurate medical image segmentation is essential for disease diagnosis and treatment. Although methods based on convolutional neural networks (CNNs) have delivered remarkable segmentation outcomes, they face challenges in capturing long-range dependencies due to the inherent limitations of convolution operators. On the other hand, transformer-based methods can establish such dependencies through self-attention, but they suffer from quadratic computational complexity, making it challenging to process 3D inputs. Additionally, the pooling operation in their encoding stage often results in feature loss, and their ability to extract multi-scale contextual information is limited. To overcome these challenges, we introduce SCEA-Net, a novel approach tailored for precise 3D medical image segmentation. The core of SCEA-Net is the spatial-channel-aware external attention model (SCEAM), which integrates parallel external attention mechanisms with convolutions to effectively capture crucial spatial and channel information. This attention mechanism leverages memory-based storage units to consider potential relationships among all samples in the dataset, while also reducing model complexity through linear computation methods. Furthermore, we have designed parallel pooling and convolutional down-sampling to minimize the loss of detailed features during the down-sampling process. Experimental results on the ACDC, Synapse, Tumor and cardiac left atrium segmentation datasets demonstrate that SCEA-Net outperforms other state-of-the-art methods, validating the effectiveness of our approach.