RLK-Net: An Efficient Residual Large Kernel Convolution with Channel-Wise Adaptive Feature Fusion for Medical Image Segmentation

While U-Net and its variants perform well in medical image segmentation, their encoders have limitations that hinder the capture of global context. Moreover, traditional skip connections do not fully utilize essential features. To address these issues, we propose a novel segmentation architecture called RLK-Net, which integrates Residual Large Kernel Convolution (RLK) block and Channel-wise Adaptive Feature Fusion (CAFF) module. RLK-Net effectively enhances feature extraction and context capture through RLK block, while the CAFF module adaptively adjusts channel importance, optimizing feature fusion and the performance of skip connections. Compared to Transformer-based models, RLK-Net achieves high segmentation performance while significantly reducing computational resource requirements. Experimental evaluations on three ultrasound medical image datasets and one dermoscopy dataset demonstrate that RLK-Net outperforms several leading medical image segmentation models.