<p>Convolutional neural networks (CNNs) have limited capability in capturing long-range dependencies due to their localized receptive fields, while pure Transformer architectures often struggle to preserve fine-grained spatial details essential for precise boundary delineation in medical image segmentation. To address these complementary limitations, we propose CS-Net, a novel hybrid segmentation network that synergistically integrates modified ConvNeXt blocks with Swin Transformer stages within a U-Net architecture. Specifically, the modified ConvNeXt blocks employ depthwise separable convolutions with a 7 × 7 kernel and batch normalization to efficiently extract local features with expanded receptive fields, while the Swin Transformer stages capture global context through shifted window-based self-attention mechanisms. To enhance feature representation, we incorporate a Coordinate Attention Module (CAM) that encodes spatial positional information into channel attention, enabling precise localization of lesion regions. Furthermore, we propose an Encoder–Decoder Feature Fusion (EDFF) module that adaptively recalibrates channel responses using efficient channel attention, facilitating selective fusion of low-level encoder features with high-level decoder semantics. Extensive experiments on five public datasets demonstrate that CS-Net consistently outperforms state-of-the-art methods, achieving Dice coefficients of 83.02%, 75.94%, 86.89%, 87.57%, and 96.95% on BUSI, UDIAT, Kvasir, ISIC18, and LiTS2017 datasets. These results validate the effectiveness of combining local convolutional feature extraction with global Transformer-based context modeling for medical image segmentation. The model's computational demands necessitate GPU acceleration and parallel processing for real-time clinical deployment, demonstrating the critical role of high-performance computing in advanced medical image analysis.</p>