MULTI-TISSUE SEGMENTATION ALGORITHM FOR KIDNEY CT IMAGES BASED ON THE HYBRID METHOD OF DEFORMABLE LARGE-KERNEL ATTENTION AND WAVELET TRANSFORM
Zhixian Tang, Yuwei Yin, WENHAO GUAN, Zhezhifeng Wang, Jinyang Zhang, Xufeng Yao, LU REN, Huachun Weng · Journal of Mechanics in Medicine and Biology · 2025
Kidney diseases have emerged as a significant and growing threat to human health. CT imaging is widely used in the diagnosis of kidney diseases, and the accurate segmentation of kidney CT images is crucial for disease diagnosis and treatment. This study proposes a multi-tissue segmentation algorithm for kidney CT images based on the hybrid method of deformable large-kernel attention and wavelet transform. By improving certain convolutional modules in the YOLOv8 network and introducing a Haar-wavelet transform-based down-sampling method, our model achieves more effective multi-scale feature extraction and better adaptability to complex scenarios. Experiments on the KiPA22 dataset and in-house data show that the proposed model outperforms deep-learning-based segmentation models in terms of DSC coefficient, Hausdorff distance, and average volume difference. It achieves higher segmentation accuracy, more precise boundary prediction, and more accurate anatomical structure volume prediction, demonstrating strong robustness and precision in medical image segmentation tasks and providing reliable support for clinical applications.