Lesion segmentation poses significant difficulties, primarily attributed to the highly irregular morphology and distinct scale variations of target regions. For this purpose, we introduce DConvUNeXt, a new convolution network combining deformable convolution v4 (DCNv4) with an efficient attention mechanism. Our architecture combines CNNs' strength in capturing fine-grained local details with the adaptability of Transformers. The former ensures stable performance under limited training data, while the latter enables data-driven adjustment of receptive fields, helping to address CNNs’ typical shortcomings in capturing global context and deformable objects. Experimental evaluations reveal that DConvUNeXt consistently achieves superior performance over existing methods in lesion segmentation tasks.