Breast Thermal Images Classification into Three Classes Using Statistical Approach and CNN-XGBoost

Breast cancer remains a significant health concern. Many works explored computer-aided methods to detect breast cancer using thermal imaging but none using three classes. Our work developed an inference model to classify breast thermal images into three categories: normal, benign, and malignant. We implemented a statistical approach to separate abnormal images into benign and malignant classes. For the inference model, we combined Convolutional Neural Network architectures (VGG16 and ResNet50) with XGBoost, where CNN serves as a feature extractor and XGBoost as the classifier. Results show that the VGG16-XGBoost model achieved 87.44% training accuracy and 72.89% validation accuracy, while the ResNet50-XGBoost model reached 91.24% training accuracy and 73.17% validation accuracy. Confusion matrix analysis reveals the model's ability to classify images effectively despite some misclassifications between benign and malignant classes. More abnormal thermal images are needed to improve the inference model of the three classes and allow for a more accurate classification.