[Synergistic learnable frequency and multi-scale spatial network for lightweight skin cancer classification]

Early screening of skin cancer is crucial to the survival rate of patients. Although deep learning has made significant progress in dermoscopic image analysis, the blurred edge of the lesion, the vulnerability to noise interference, and the limited computing resources at the time of model deployment are still the main bottlenecks. To this end, this paper proposes a frequency-space collaborative enhancement network (FSC-Net) based on lightweight classification. Aiming at the problem of blurred lesion edge and noise interference, the network first constructs a learning frequency enhancement module. Through the dynamic selective enhancement of frequency domain features, the lesion edge is finely characterized while suppressing high-frequency artifacts. Secondly, aiming at the scale heterogeneity of lesion morphology, this paper proposes a multi-scale aggregation module, which uses multi-branch pooling to reduce the loss of deep semantic features in the lightweight network. Finally, in order to solve the problem of difficult localization of complex lesion areas, this paper introduces a directional spatial calibration mechanism, which realizes accurate localization of lesion features through orthogonal decoupling coding and asymmetry factors. The experimental results on the 2019 international skin image collaboration challenge (ISIC2019) and the human against machine with 10000 training images (HAM10000) dataset show that FSC-Net achieves 93.41% eight-classification accuracy and 95.84% seven-classification accuracy with a lower number of parameters. Compared with the existing advanced models, the proposed method achieves a better balance between computational overhead and diagnostic performance, and provides a robust and efficient solution for auxiliary diagnosis in resource-constrained environments.