CDMT-Unet: a dual-frequency cross-fusion model for skin cancer segmentation

In this paper, we propose a novel cross-fusion model, CDMT-Unet, which leverages the MobileNet and Transformer architectures for precise segmentation and diagnosis of skin cancer images. This model integrates the strengths of the Haar wavelet transform, capable of capturing global texture and edge information through multi-scale analysis, with an attention mechanism that focuses on more pertinent information. By combining these methodologies, the model effectively utilizes both high-frequency detail and low-frequency structural information to comprehensively characterize the skin lesion area. Building upon this foundation, the CDMT-Unet model adopts a hybrid architecture encompassing CNN, Transformer, and Unet components. The CNN module is employed to extract local spatial features, while the Transformer module captures global contextual information via a self-attention mechanism. Furthermore, the Unet structure facilitates the fusion and reconstruction of multi-level features through skip connections. Utilizing a dual-frequency cross-fusion strategy, the model adeptly integrates features extracted by the discrete Walsh transform and Haar wavelet transform, thereby enhancing its ability to segment complex skin lesions. Experimental results demonstrate that CDMT-Unet achieves remarkable performance across two public skin cancer datasets (ISIC2017 and ISIC2018). Specifically, on ISIC2017, it achieves mIoU of 80.51%, DSC of 87.88%, ACC of 95.68%, Spe of 96.98%, and Sen of 92.13%; on ISIC2018, it reaches mIoU of 80.31%, DSC of 87.60%, ACC of 95.37%, Spe of 97.93%, and Sen of 89.64%. These results highlight the model’s excellence in accurate segmentation of lesion edges and preservation of small-scale features. This model offers an efficient and reliable solution for the automated diagnosis of skin cancer, demonstrating significant potential for a wide range of clinical applications. Our source code is publicly available at: https://github.com/zzhzd/CDMT-Unet .