Detecting Skin Disease Using A Hybrid Deep Learning Approach

Abstract

Diseases of the skin have taken on critical importance for global health, bringing into focus a necessity for accurate and quick diagnostics. This work demonstrates the development of both a Hybrid Convolutional Neural Network and Artificial Neural Network (Hybrid CNN-ANN) model for automated classification of skin diseases using dermatoscopic images. We trained our model on the HAM10000 database, which is a collection of 10,000 high-quality images representing various types of skin diseases (including melanoma, nephews and benign and malignant pinches). The approach includes the data acquisition, preprocessing, model architecture construction, and training, as well as evaluation. In addition, we conducted the preprocessing of image treatments that provided better feature extraction and learning efficiency for our model using traditional methods such as resizing, normalize and data augmentation. The proposed Hybrid CNN-ANN leverages convolutional layers for hierarchical spatial feature learning and a fully connected ANN model for fine-grained classification. Such combination can help the model effectively learn complex representations and subtle changes of skin lesion patterns. We trained and improved our models on Kaggle because it has great computing power and deep learning capabilities. The accuracy, precision, recall and F1-score were well studied for the Hybrid CNN-ANN model. The results of the experiments indicated that the model was robust and had a high accuracy in classification with an average recognition rate of 77.94%, which showed its potential in accurate detection for skin diseases. This work emphasizes the importance of integrating CNN and ANN architectures to improve feature representations learning, which yield better classification results. The results will support progress in the direction of AI-assisted dermatological diagnosis, and lay groundwork for broader applications to automated medical image analysis tasks. The objective of this work is to offer a scalable yet cost-effective approach of training intelligent diagnostic models for skin disease classification, leveraging publicly available datasets and cloud-based training.