PDS-UKAN: Subdivision hopping connected to the U-KAN network for medical image segmentation.
Authors
Affiliations (6)
Affiliations (6)
- Heilongjiang Provincial Key Laboratory of Complex Intelligent System and Integration, School of Automation, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, China; School of Computer Science and Technology, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, China.
- Heilongjiang Provincial Key Laboratory of Complex Intelligent System and Integration, School of Automation, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, China.
- School of Computer Science and Technology, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, China.
- State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong 510060, China; Guangdong Esophageal Cancer Institute, Guangzhou, Guangdong 510060, China.
- State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong 510060, China; Guangdong Esophageal Cancer Institute, Guangzhou, Guangdong 510060, China. Electronic address: [email protected].
- Institute for Brain Research and Rehabilitation, South China Normal University, Guangzhou 510631, China. Electronic address: [email protected].
Abstract
Accurate and efficient segmentation of medical images plays a vital role in clinical tasks, such as diagnostic procedures and planning treatments. Traditional U-shaped encoder-decoder architectures, built on convolutional and transformer-based networks, have shown strong performance in medical image processing. However, the simple skip connections commonly used in these networks face limitations, such as insufficient nonlinear modeling capacity, weak global multiscale context modeling, and limited interpretability. To address these challenges, this study proposes the PDS-UKAN network, an innovative subdivision-based U-KAN architecture, designed to improve segmentation accuracy. The PDS-UKAN incorporates a PKAN module-comprising partial convolutions and Kolmogorov - Arnold network layers-into the encoder bottleneck, enhancing the network's nonlinear modeling and interpretability. Additionally, the proposed Dual-Branch Convolutional Boundary Enhancement Module (DBE) focuses on pixel-level boundary refinement, improving edge detail preservation in shallow skip connections. Meanwhile, the Skip Connection Channel Spatial Attention Module (SCCSA) mechanism is applied in the deeper skip connections to strengthen cross-dimensional interactions between channels and spatial features, mitigating the loss of spatial information due to downsampling. Extensive experiments across multiple medical imaging datasets demonstrate that PDS-UKAN consistently achieves superior performance compared to state-of-the-art (SOTA) methods.