SEA-NET: medical image segmentation network based on spiral squeeze-and-excitation and attention modules

SEA-NET: medical image segmentation network based on spiral squeeze-and-excitation and attention modules

2024 | Liangli Xiong1, Chen Yi1, Qiliang Xiong1 and Shaofeng Jiang1*
The paper introduces SEA-NET, a novel medical image segmentation model that combines the U-Net architecture with Squeeze-and-Excitation (SE) and attention modules to enhance the performance of small target segmentation. The SE-Res module and attention module are designed to integrate deep semantic information into shallow feature maps, addressing the issue of feature loss during multiple convolutions. The model uses a hybrid loss function, combining cross entropy and Tversky loss, to handle imbalanced datasets and ensure fast convergence. Experimental results on the LPBA40 brain MRI dataset and a peripheral blood smear image dataset demonstrate that SEA-NET outperforms U-Net and its variants in terms of Dice coefficient, sensitivity, and specificity, particularly in small target segmentation tasks. The proposed model effectively retains both shallow and deep features, improving the accuracy and robustness of medical image segmentation.The paper introduces SEA-NET, a novel medical image segmentation model that combines the U-Net architecture with Squeeze-and-Excitation (SE) and attention modules to enhance the performance of small target segmentation. The SE-Res module and attention module are designed to integrate deep semantic information into shallow feature maps, addressing the issue of feature loss during multiple convolutions. The model uses a hybrid loss function, combining cross entropy and Tversky loss, to handle imbalanced datasets and ensure fast convergence. Experimental results on the LPBA40 brain MRI dataset and a peripheral blood smear image dataset demonstrate that SEA-NET outperforms U-Net and its variants in terms of Dice coefficient, sensitivity, and specificity, particularly in small target segmentation tasks. The proposed model effectively retains both shallow and deep features, improving the accuracy and robustness of medical image segmentation.
Reach us at info@study.space