An improved nnUNet with Axial Attention Mechanism and Attention Gated Units for Efficient and Accurate Kidney Segmentation in CT Images.
L’essentiel
Accurate segmentation of computed tomography (CT) images plays a vital role in surgical planning and therapeutic evaluation for kidney disease diagnosis and treatment. Although nnUNet is widely used in medical image segmentation, its consecutive convolutions suffer from limitations in feature extraction, leading to weight bias and difficulty in effectively modeling long-range dependencies. Additionally, there exists a large semantic gap between encoder and decoder feature maps. Direct feature fusion via skip connections will degrade segmentation accuracy. This study develops the AAAG-UNet model based on the nnUNet framework by integrating the axial attention mechanism (AAM) and attention gated unit (AGU). Optimized by these two modules, the model overcomes existing technical bottlenecks and effectively improves segmentation performance, offering strong support for the clinical diagnosis of kidney diseases. The proposed model integrates axial attention mechanisms into the encoder and attention gated units into the decoder to enhance the capture of long-range spatial dependencies and improve kidney segmentation performance. Specifically, AAM computes self-attention weights along different axes to effectively fuse global information and preserve more spatial details. Additionally, since it performs only local axial attention computation, it maintains stable segmentation performance with low computational cost. Furthermore, AGU is utilized to suppress irrelevant regions, enhance important local features, and eliminate redundant information in CT images, thereby further improving segmentation accuracy. Experimental results show that our method has remarkable performance in kidney segmentation. On the KiTS19, Kid184 and Kid438 datasets, the model obtains mean Dice values of 0.9639, 0.9712, and 0.9770 for the left kidney, and 0.9660, 0.9742, and 0.9785 for the right kidney, respectively. The results show that incorporating axial attention mechanisms into the encoder and attention gated units into the decoder of the proposed model can effectively improve kidney segmentation performance.
Synthèse détaillée
Résumé original
Accurate segmentation of computed tomography (CT) images plays a vital role in surgical planning and therapeutic evaluation for kidney disease diagnosis and treatment. Although nnUNet is widely used in medical image segmentation, its consecutive convolutions suffer from limitations in feature extraction, leading to weight bias and difficulty in effectively modeling long-range dependencies. Additionally, there exists a large semantic gap between encoder and decoder feature maps. Direct feature fusion via skip connections will degrade segmentation accuracy. This study develops the AAAG-UNet model based on the nnUNet framework by integrating the axial attention mechanism (AAM) and attention gated unit (AGU). Optimized by these two modules, the model overcomes existing technical bottlenecks and effectively improves segmentation performance, offering strong support for the clinical diagnosis of kidney diseases. The proposed model integrates axial attention mechanisms into the encoder and attention gated units into the decoder to enhance the capture of long-range spatial dependencies and improve kidney segmentation performance. Specifically, AAM computes self-attention weights along different axes to effectively fuse global information and preserve more spatial details. Additionally, since it performs only local axial attention computation, it maintains stable segmentation performance with low computational cost. Furthermore, AGU is utilized to suppress irrelevant regions, enhance important local features, and eliminate redundant information in CT images, thereby further improving segmentation accuracy. Experimental results show that our method has remarkable performance in kidney segmentation. On the KiTS19, Kid184 and Kid438 datasets, the model obtains mean Dice values of 0.9639, 0.9712, and 0.9770 for the left kidney, and 0.9660, 0.9742, and 0.9785 for the right kidney, respectively. The results show that incorporating axial attention mechanisms into the encoder and attention gated units into the decoder of the proposed model can effectively improve kidney segmentation performance.