Diagnosis and Classification of Skin Cancer Types Based on Improved EfficientNet Network

Skin cancer has emerged as a major global health challenge, where timely and precise identification plays a vital role in optimizing clinical outcomes and mortality reduction. Traditional diagnostic approaches, which rely heavily on clinical experience, often suffer from limited accuracy and efficiency due to variations in observer assessment and the complexity of skin lesions. EfficientNet, currently recognized as the most advanced convolutional neural network architecture, demonstrates promising capabilities in the automated classification of dermatological malignancies. Nevertheless, practical implementation encounters several challenges, particularly due to heterogeneity in visual characteristics and confounding variables from imaging conditions. In response, this paper proposes an enhanced version of the EfficientNet model by incorporating Dynamic Convolution (Do-Conv) for better feature extraction and Coordinate Attention (CA) to improve spatial feature perception. These enhancements allow the model to concentrate on critical regional information. Extensive experiments conducted on the ISIC 2019 and HAM10000 datasets demonstrate that the combination of Do-Conv and CA achieves outstanding classification accuracies and exceeds the performance of mainstream models like MobileNet and ResNet in comparative tests. The present research provides a robust framework for the automated detection of skin cancer while simultaneously contributing to the advancement of deep learning methodologies in medical image analysis.