AUTOMATED SKIN DISEASE CLASSIFICATION USING HYBRID DEEP LEARNING AND ENSEMBLE LEARNING TECHNIQUES

Abstract

Skin diseases impact millions of people globally, from prevalent issues such as eczema and acne to life-threatening conditions such as melanoma and psoriasis. Due to the similarity in visual features of these diseases, extensive variation in skin types and dependence on subjective clinical observations, diagnosing these diseases timely and accurately is a challenging task. Conventional diagnostic techniques, such as manual palpation and biopsy, are invasive, time consuming, and subjective. To overcome the drawbacks, in this paper we present a hybrid architecture that combines deep learning-based feature extraction and ensemble machine learning classifiers for automated classification of skin diseases. We use a state-of-the-art convolutional neural network, i.e., EfficientNet to obtain hierarchical representations of breast cancer images which incorporate local and global patterns of lesions on skin. These deep features then are served as inputs to the ensemble classifiers including XGBoost, LightGBM based on multiple decision trees for better classification accuracy and avoiding over-fitting. The proposed system is tested on a publicly available dermatology dataset consisting multiple disease classes, resulting in the overall accuracy of 92.4%, which is higher compared to the conventional CNN, SVM and Random Forest techniques. Moreover, these results show high precision, recall, and F1-score in all classes (results not shown), suggesting that the proposed framework can perform well to discriminate visually similar diseases. The results show that integrating deep feature extraction and ensemble learning provides a robust, non-invasive, scalable solution for automated dermatological diagnosis and has the potential to be integrated into clinical decision support systems.