SKINC-NET: an efficient Lightweight Deep Learning Model for Multiclass skin lesion classification in dermoscopic images

Diagnosing skin cancer through visual image examinations is time-consuming and error-prone, as the similar appearance and variations within each type of skin cancer challenge the efficiency and accuracy. Computer-aided techniques, notably convolutional neural networks (CNNs), demand substantial computational resources and exhibit a high level of time complexity, rendering them impractical for real-world applications necessitating rapid and precise skin cancer detection. To address this issue, this research introduces SKINC-NET, a streamlined deep learning (DL) model specifically designed to achieve precise diagnoses of various skin cancer types while minimizing computational resource requirements. SKINC-NET can classify seven types of skin lesions using dermoscopic images, designed with dense layers, batch normalization (BN), and LeakyReLU activation functions. Numerous experiments were conducted on the HAM10000 dataset, using a data augmentation approach to tackle the class imbalance problem. The study compared the SKINC-NET model’s performance with four transfer learning models, namely ResNet50, VGG16, MobileNetV2, and EfficientNetB0. Several ablation studies were conducted to identify the optimal hyperparameters that lead to precise skin cancer diagnosis. The SKINC-NET model demonstrates its superiority by achieving the highest overall classification accuracy of 98.54% with minimal computation. The experimental findings affirm that the SKINC-NET model showcases remarkable robustness and presents several significant benefits, including lower computational complexity, a smaller number of trainable parameters, and superior performance compared to existing state-of-the-art models. SKINC-NET can assist dermatologists in making appropriate decisions regarding the early detection of skin cancer.