ScNet: a lightweight CNN with depthwise and SE modules for skin lesion classification

Accurate and early classification of skin lesions is important for improving patient management and clinical decision support. In this study, a lightweight deep learning model, referred to as ScNet, was developed using transfer learning for the classification of dermoscopic images. The architecture uses a MobileNetV3 backbone (with depthwise separable convolutions and Squeeze-and-Excitation modules) topped with an Adaptive Classification Module (ACM) designed to balance accuracy and computational efficiency. The HAM10000 dataset, containing 10,015 dermoscopic images across seven lesion categories, was used for training, validation, and testing under a lesion-wise split protocol. The proposed model achieved a test accuracy of 97.05%, with macro precision of 88.68%, macro recall of 91.41%, and macro F1-score of 89.91%. Weighted precision, recall, and F1-scores were 97.16%, 97.05%, and 97.09%, respectively. The results demonstrate that ScNet provides competitive performance while maintaining a compact architecture, making it a suitable candidate for deployment in environments with limited computational resources. While maintaining high overall accuracy and melanoma sensitivity, performance on underrepresented classes (e.g. DF) was lower; external validation on independent cohorts is planned.