Skin Cancer Detection Using Data Fusion and Deep Transfer Learning Techniques

Abstract

Skin cancer remains a major healthcare challenge. Early detection is important because it increases the chances of successful treatment. Deep learning has become an effective approach for automated skin lesion classification. However, many existing studies rely on a single dataset, which limits the diversity of training images and may reduce model generalization. This study proposes a transfer learning framework for binary skin cancer classification using dataset-level fusion. Two public datasets, PAD-UFES-20 and ISIC~2016, are combined to create a fused dataset that contains both clinical smartphone images and dermoscopic images. The images are resized to 224 × 224 pixels, converted to RGB format, and divided into training, validation, and testing sets using stratified splitting. Data augmentation and class weighting are applied to improve model learning. Two pretrained models, ResNet50 and DenseNet121, are fine-tuned using a two-phase training strategy. Model performance is evaluated using accuracy, sensitivity, specificity, positive and negative predictive values, AUC-ROC, and confusion matrices. Experimental results show that ResNet50 achieved an accuracy of 82.19%, while DenseNet121 achieved the best accuracy of 82.98%. The findings suggest that dataset-level fusion improves training data diversity and provides reliable performance for binary skin cancer classification. The proposed framework is organized into a Training Layer and a Testing Layer, which separate model development from inference. This design provides a flexible framework for future medical image classification applications.