DLKUNet: A Lightweight and Efficient Network With Depthwise Large Kernel for Medical Image Segmentation

Accurate multi-organ segmentation is crucial in computer-aided diagnosis, surgical navigation, and radiotherapy. Deep learning-based methods for automated multi-organ segmentation have made significant progress recently. However, these improvements often increase model complexity, leading to higher computational costs. To address this problem, we propose a lightweight and efficient network with depthwise large kernel, called DLKUNet. Firstly, we utilize a hierarchical architecture with large kernel convolution to effectively capture multi-scale features. Secondly, we constructed three segmentation models with different layers to meet different speed and accuracy requirements. Additionally, we employ a novel training strategy that works seamlessly with this module to enhance performance. Finally, we conducted extensive experiments on the multi-organ abdominal segmentation (Synapse) and the Automated Cardiac Diagnosis Challenge (ACDC) dataset. DLKUNet-L significantly improves the 95% Hausdorff Distance to 13.89 mm with 65% parameters of Swin-Unet on the Synapse. Furthermore, DLKUNet-S and DLKUNet-M use only 4.5% and 16.52% parameters of Swin-Unet, achieving Dice Similarity Coefficient 91.71% and 91.74% on the ACDC. These results underscore the proposed model's superior performance in terms of accuracy, efficiency, and practical applicability.