Introduction: Accurate segmentation of lesions is crucial for medical image analysis and diagnosis. Traditional U-Net methods, limited by their receptive field, struggle to capture global contextual semantic information. Although transformers can model contextual dependencies, their high computational complexity restricts their practical application. Method: To address these limitations, we developed X-Net, a dual-stream hybrid network integrating U-Net’s local representation with ConvNeXt’s global context, featuring a Cross Feature Fusion module for dynamic feature refinement. Results: Extensive experiments on the DRIVE dataset demonstrated that X-Net achieved a Dice score of 99.3%, while reducing parameter count by 98.99% and FLOPs by 75.55% compared to TransU-Net. Discussion: These results indicate that X-Net effectively balances segmentation accuracy with computational efficiency, offering significant advantages over existing hybrid models. However, evaluation was limited to retinal vessel segmentation; broader validation across diverse lesion types remains necessary. X-Net demonstrates excellent generalization ability and computational efficiency, representing a reliable solution for clinical medical image segmentation applications.
Wang et al. (Tue,) studied this question.