Comparative analysis of autoencoder architectures for breast cancer detection using dynamic infrared thermography

Breast cancer is the most diagnosed cancer among women worldwide. Early detection substantially improves treatment outcomes, especially when lesions are small and localized. Although conventional imaging modalities such as mammography, CT, MRI, and ultrasonography play a vital role in diagnosis, they often entail radiation exposure, high cost, and the use of contrast agents. These drawbacks have motivated increasing interest in non-invasive and cost-effective alternatives such as Infrared Thermal Imaging (ITI), which captures surface temperature variations that may indicate malignancy. This study proposes a novel ITI-based diagnostic framework integrating deep learning-driven feature extraction with conventional machine learning classifiers. Three autoencoder architectures—Vanilla Autoencoder (VanAE), Convolutional Autoencoder (CAE), and Variational Autoencoder (VAE)—were utilized to extract discriminative latent features from dynamic breast thermograms. The extracted features were subsequently classified using Support Vector Machine (SVM) and Random Forest (RF) algorithms. Experimental evaluation on a balanced DMR-IR dynamic dataset comprising 3,600 thermograms demonstrated that the CAE-SVM combination achieved the highest performance, reaching 92.28% accuracy, 89.11% sensitivity, 95.94% specificity, and a 92.26% F1-score. In addition to its superior classification performance, the CAE model exhibited the shortest training time, underscoring its potential for practical clinical implementation. Overall, the findings confirm the effectiveness of autoencoder-based architectures in learning meaningful representations directly from raw thermograms without relying on handcrafted or pre-trained features.