A two-stage deep learning system for skin lesion classification

Skin cancer remains one of the most prevalent oncological diseases worldwide, and its early detection is crucial for improving treatment outcomes and reducing mortality rates. In this regard, the application of artificial intelligence techniques for the automated analysis of dermatoscopic images has gained significant attention. This study presents a two-stage deep learning framework for the classification of skin lesions, integrating a preliminary skin region detection stage with a subsequent diagnostic classification stage. In the first stage, a MobileNetV3-Small model is employed to identify the presence of skin regions within input images, enabling the exclusion of irrelevant background areas and thereby reducing the likelihood of erroneous predictions. In the second stage, an EfficientNet-B4 convolutional neural network is utilized to classify lesions into seven diagnostic categories, including melanoma, basal cell carcinoma, dermatofibroma, and other clinically relevant classes. To enhance model performance and generalization capability, a comprehensive data processing pipeline is implemented, incorporating image preprocessing, data augmentation techniques, and class balancing strategies. These approaches mitigate the effects of class imbalance and improve the robustness of the model when applied to real-world data. Furthermore, model interpretability is addressed through the use of the Grad-CAM method, which enables visualization of the most influential regions contributing to the model’s predictions, thereby increasing the transparency and reliability of the system for medical applications. The proposed approach is evaluated on the publicly available HAM10000 dermatoscopic image dataset, comprising 10,000 images across seven diagnostic categories. Experimental results demonstrate that the EfficientNet-B4 model achieves high classification performance, with accuracy reaching up to 94%, outperforming several alternative architectures. The inclusion of the initial skin detection stage further enhances system robustness by reducing the impact of noise and irrelevant image content. In conclusion, the developed two-stage deep learning system demonstrates strong potential as a decision-support tool for preliminary skin lesion screening, facilitating early detection of oncological conditions and contributing to improved diagnostic workflows. The findings of this study highlight the effectiveness and highlight the effectiveness and potential of deep learning methods deep learning methods in biomedical engineering and medical imaging applications