Improving skin lesion classification using k-means, Transfer Learning and Mobilenet Architecture.

Abstract

The accurate and early classiication of skin lesions remains a critical concern in dermatological diagnostics due to the visual similarity between benign and malignant conditions. This study proposes an enhanced skin lesion classiication framework that combines K-Means image segmentation, transfer learning, and the lightweight MobileNet architecture. The HAM10000 dataset was employed, and preprocessing involved augmentation, normalization, and class balancing to improve model generalization. K-Means clustering was utilized to segment the lesion area, improving the model’s focus on the region of interest. Transfer learning was applied using a pre-trained MobileNet model, which was ine-tuned on the dataset to enable eicient and accurate classiication. The model was evaluated across multiple training epochs, with the best results achieved at epoch 50: 98.20% accuracy, 93.75% precision, 100% recall, and 97.10% F1-score, alongside an AUC of 98.00%. Comparative analysis demonstrates that the proposed model outperforms existing deep learning models in both diagnostic accuracy and computational eiciency. These results indicate that the integration of K-Means segmentation with MobileNet and transfer learning provides a robust solution for real-time and resource-eicient skin lesion classiication in clinical applications. Keywords Skin lesion classiication · MobileNet · K-Means segmentation · Transfer learning · HAM10000 · Deep learning