Ensemble Deep Learning for Automated Skin Lesion Classification: A GAN-Augmented Multi-Model Approach with Clinical Deployment

Abstract

Introduction: This paper presents an end-to-end automated system for skin lesion classification using ensemble deep learning. Our pipeline begins with dermatoscopic image preprocessing (hair removal via inpainting) and lesion segmentation (K-means clustering), followed by Generative Adversarial Network (GAN) based data augmentation to address class imbalance.

Objectives: Four CNN architectures (MobileNet, ResNet50, Xception, NASNet) process images in parallel, with predictions combined through weighted ensemble learning. The system achieves robust classification of lesions as benign, malignant, or other subtypes, deployed via a flask web interface for clinical use.

Methods: Experimental results on the ISIC dataset demonstrate superior accuracy compared to single-model approaches, with particular improvements in melanoma detection. This work bridges the gap between computer vision research and clinical dermatology applications.

Results: Our experiments on the ISIC 2020 dataset demonstrate state-of-the-art performance, achieving 97.5% accuracy and 98.3% sensitivity for malignant cases, outperforming existing single-model approaches. The system’s modular design ensures scalability, allowing integration with teledermatology platforms and mobile health applications.

Conclusions: The skin lesion plays an important role in human health, skin related issues causes the most common and concerning medical conditions. Our ensemble approach outperformed individual models, achieving superior accuracy, precision, recall, and F1-scores. On the original ISIC dataset, the ensemble model attained an accuracy of 89.5%, while the balanced dataset further boosted performance to 97.5%. These results highlight the effectiveness of ensemble learning in combining the strengths of multiple architectures, producing a more reliable and precise classification system. Additionally, the model’s ability to mitigate class imbalance through oversampling demonstrates its robustness, making it a promising tool for clinical applications in early skin cancer detection and diagnosis.