Deep Learning Framework for Optimizing Early Detection of Measles Using Transfer Learning

Abstract

Measles is a highly infectious viral disease that can have serious health consequences. Accurate and early diagnosis is crucial. This study aims to enhance automated classification and early detection of this disease. To address the class imbalance, we augmented the dataset of normal images. Spatial features were extracted using convolutional neural networks, and traditional classifiers, including support vector machine, Random Forest, logistic regression, and k-nearest neighbors were applied to these features. Initial classification accuracy based solely on spatial features was as follows: Random Forest 63%, SVM 63%, KNN 60%, and Logistic Regression 63%. Through 10-fold cross-validation, mean accuracies were recorded as 65% for RF, 62% for SVM, 60% for KNN, and 61% for LR. Despite these initial results, the implementation of transfer learning led to significant improvements. By extracting probabilistic features from spatial features using RF and KNN models and concatenating these derived features, classification accuracy was substantially enhanced. The improved model achieved 99% accuracy for RF, SVM, and LR, with KNN reaching 98%. Cross-validation confirmed the robustness of the models, with a mean accuracy of approximately 98% and minimal standard deviations of 0.01. The findings demonstrate that combining transfer learning with traditional classifiers improves the efficiency and accuracy of lesion images. This approach shows significant potential for clinical applications.