Leveraging Transfer Learning and Fine-Tuning for Improved Skin Cancer Detection in Dermatoscopic Images

Abstract

 This study presents a deep learning-based approach to improve skin cancer detection using dermoscopic images. The proposed system employs the ResNet50 architecture with transfer learning to enhance the classification accuracy of skin lesions. Training and evaluation were conducted on the HAM10000 dataset, with specialized techniques applied to mitigate class imbalance and improve model generalization. To ensure interpretability, the study integrates SHAP (Shapley Additive explanation), which identifies key image regions influencing classification decisions. The model leverages texture and morphological features extracted by the convolutional neural network (CNN) to distinguish between benign and malignant lesions effectively. A comprehensive evaluation framework was developed, incorporating random sampling and visualization methods to rigorously assess performance. The results demonstrate consistent accuracy across diverse lesion types, highlighting the model’s clinical applicability