VITILIGO DETECTION USING MACHINE LEARNING ALGORITHMS

Abstract

Vitiligo is a chronic skin condition involving the progressive depigmentation of the skin because of melanocyte destruction or malfunction. Prompt and proper diagnosis of vitiligo is critical, since early treatment can considerably enhance outcome and improve the quality of life for those suffering from the disease. Manual diagnosis, however, tends to be qualitative and time-consuming, especially in low-resource clinical settings. In this thesis, a diagnostic framework based on deep learning is presented for automatically diagnosing vitiligo from skin images. The research starts with the evaluation of five leading convolutional neural networks (CNNs), namely VGG16, ResNet50, InceptionV3, EfficientNet, and DenseNet121, on a publicly distributed vitiligo dataset retrieved from Kaggle. The initial phase involved training and fine-tuning each CNN model to identify the top performers based on metrics such as accuracy, precision, recall, and F1-score. Among the evaluated models, VGG16, ResNet50, and DenseNet121emerged as the most effective, and were selected for further ensemble modeling. To enhance predictive reliability, three ensemble strategies were employed: bagging using Random Forest, boosting using XGBoost, and stacking with a logistic regression meta-learner. Beyond traditional ensemble methods, a Multilayer Perceptron (MLP)-based architecture was developed that fused deep features extracted from the three CNNs and learned complex inter-feature representations. Experimental evaluations demonstrated that the proposed MLP-based model significantly outperformed all other approaches, achieving a classification accuracy of 99.22%, along with 99% precision, 99% recall, and 99% F1-score. Traditional ensembles such as Random Forest also performed well (98.43% accuracy), but were slightly less effective in terms of overall balance across evaluation metrics. These results confirm that feature-level fusion combined with neural modeling can yield superior classification outcomes in medical image analysis. This research not only demonstrates the viability of deep ensemble learning for vitiligo detection but also sets the foundation for developing intelligent dermatological screening tools. The proposed framework is scalable and may be extended to support multi-class classification and other skin conditions in future clinical decision support systems.