A Deep Learning Based Framework for Skin Cancer Detection Using a Layered Software Architecture with Transfer Learning

Abstract

Skin cancer has become a serious public health issue worldwide. The number of cases is rising due to higher UV exposure and changing lifestyle patterns. Early detection is the best way to save lives. However, many areas still lack specialized doctors and equipment needed for a quick diagnosis. In this paper, we develop and test an automated skin cancer detection method based on standard smartphone photography. It provides a practical solution for resource-limited settings by removing the need for expensive clinical dermoscopy equipment, making diagnosis accessible in areas with limited infrastructure. The system was trained and tested on PAD-UFES-20, which contains 2,298 clinical smartphone images of skin lesions. The images were grouped into two categories: malignant and benign. Malignant cases involved Basal Cell Carcinoma, Squamous Cell Carcinoma, and Melanoma. Benign lesions consisted of Actinic Keratosis, Nevus, and Seborrheic Keratosis. Transfer learning was used to fine-tune two models: EfficientNetB3 and VGG16. Both were pre-trained on ImageNet. EfficientNetB3 achieved a test accuracy of 82.61%. It outperformed VGG16, which achieved an accuracy of 79.78%. The system follows a modular pipeline covering data ingestion, preprocessing, model training, and inference. Unlike many existing approaches that require high-end hardware, our approach provides high accuracy using standard technology. This design allows easy adaptation in clinical settings with limited infrastructure. The proposed four-layer modular pipeline embodies software engineering principles of separation of concerns and component independence, enabling maintainable and scalable deployment in resource-constrained environments.