Enhancing Skin Cancer Classification with Mixup Data Augmentation and Efficientnet

Skin lesion classification and segmentation are two crucial tasks in dermatological diagnosis, here automated approaches can significantly aid in early detection and improve treatment planning. The proposed work presents a comprehensive framework that integrates K-means clustering for segmentation, Mixup augmentation for data enhancement, and the EfficientNet B7 model for classification. Initially, K-means clustering is applied as a pre-processing step to accurately segment the lesion regions from the background, ensuring that the model focuses on processing the most relevant and informative features. This segmentation enhances the model’s ability to differentiate between subtle lesion boundaries and surrounding skin textures. To address the common issue of class imbalance and to improve the overall robustness of the classification model, Mixup augmentation is employed. This technique generates synthetic samples by linearly interpolating between pairs of images and their corresponding labels, effectively enriching the training dataset and promoting better generalization. For the classification task, EfficientNet B7 is utilized due to its superior feature extraction capabilities, optimized scalability, and excellent performance across various image recognition challenges. The entire pipeline was evaluated on a dataset comprising 10,015 dermatoscopic images covering seven distinct categories of skin lesions. The proposed method achieved outstanding performance, demonstrating a precision rate of 95.3% and maintaining a low loss of 0.2 during evaluation. Compared to traditional machine learning and earlier deep learning approaches, the proposed framework showed significant improvements, particularly in handling complex patterns and imbalanced datasets, making it a promising solution for real-world clinical deployment in dermatology.