Objective Dermoscopy is a key non-invasive diagnostic technique that significantly improves the early detection rate and diagnostic accuracy of skin tumors. However, its diagnostic efficacy heavily depends on the clinician's individual experience, and the interpretation process involves obvious subjective variability, which limits its widespread clinical application and standardized promotion. With the continuous advancement of artificial intelligence in the medical field, deep learning has become an important tool for medical image detection and classification, significantly promoting the development of computer-aided diagnosis. Mainstream convolutional neural networks are limited by their local receptive field design and lack the capability to extract global multi-scale features. Although combining CNNs with Transformers can compensate for this limitation, it introduces significantly higher model complexity. Therefore, achieving high-precision lesion segmentation while reducing model complexity has become a key challenge in current research. To address the above problem, this paper proposes a dual-encoder skin lesion segmentation method incorporating depthwise separable convolutions, aiming to achieve a better balance between segmentation accuracy and model efficiency.

Methods The proposed method consists of four main components. 1) A dual-branch encoder based on CNN and Transformer extracts local detailed representations and global contextual information. This design not only provides strong segmentation performance but also intentionally introduces high complexity to facilitate subsequent validation of lightweight design strategies. 2) A multi-scale spatial perception module is designed at the network bottleneck. It addresses the receptive field discrepancy between CNN and Transformer through multi-scale receptive fields while capturing spatial information along both horizontal and vertical directions. It effectively integrates local and global features and minimizes the parameter count at the stage with the highest channel dimensionality, serving as a bridge connecting the complex encoder and the lightweight decoder. 3) To tackle the structural complexity inherent in the dual-encoder architecture, a bottleneck decoder is constructed. This decoder reduces the number of channels before upsampling, significantly enhancing model efficiency, and constitutes the core of the lightweight design. 4) To mitigate the loss of fine-grained details in skin lesions, an adaptive feature fusion module is incorporated into the skip connections. This module enhances critical features in image segmentation through interaction between encoder and decoder features. It uses adaptive residual connections to align feature distributions between encoder and decoder and employs element-wise addition instead of concatenation for feature fusion. This reduces model parameters, suppresses noise, and acts as an effective supplement to lightweight operations, preventing performance degradation caused by lightweight design.

Results and Discussions The proposed DSC-Net was systematically compared with several state-of-the-art segmentation models on four public datasets: ISBI2016, ISBI2017, ISIC2018, and PH². Experimental results show that DSC-Net achieves the best performance across all comprehensive metrics. Accuracy reached 96.51%, 95.13%, 94.39%, and 96.41% on the four datasets, respectively. Dice similarity coefficients were 93.44%, 87.06%, 89.20%, and 96.12%, respectively. Kappa coefficients were 96.24%, 94.95%, 93.99%, and 95.42%, respectively. Visualization analysis further demonstrates that DSC-Net effectively alleviates common problems in skin lesion segmentation, including feature information loss, lesion region mis-segmentation, and foreign object interference, significantly improving segmentation completeness and robustness. On the ISBI2016, ISBI2017, and ISIC2018 datasets, DSC-Net was compared with other recent models under fair and objective data conditions, and the results show that DSC-Net maintains a clear performance advantage with good generalization capability. Ablation studies on the ISBI2016 and ISBI2017 datasets clearly demonstrate the individual contributions and synergistic effects of each module. The results show that the modules work together to effectively select and fuse features, reduce redundant information, and enhance key detail expression, achieving simultaneous improvement in segmentation accuracy and parameter efficiency, thereby validating the rationality and effectiveness of the lightweight design strategy. Compared with baseline models, DSC-Net achieves superior segmentation performance with lower computational overhead, demonstrating the practical value of the proposed lightweight architecture.

Conclusions To address the problems of feature information loss, regional segmentation deviation, and large variation in lesion scale in skin lesion image segmentation, this paper proposes a dual-encoder skin lesion segmentation method incorporating depthwise separable convolutions. The model fully exploits the inductive bias of CNNs in spatial modeling, complemented by the capability of Transformers in global relationship modeling. Through depthwise separable convolution for effective feature selection, a bottleneck decoder for redundant feature elimination, and adaptive feature fusion for preventing feature loss, the model achieves selective feature learning and utilization, ensuring high parameter efficiency. Experimental results on four datasets show that the segmentation performance of DSC-Net is superior to existing advanced methods. In visual comparisons, the segmentation results of DSC-Net are highly consistent with ground truth labels and demonstrate excellent robustness in challenging scenarios such as blurred boundaries and low contrast. The proposed method provides an effective solution for accurate skin lesion segmentation under limited computational resources. Future work will consider improving the encoder structure to further enhance the accuracy of skin lesion image segmentation.