Thermal signatures in breast cancer: Deciphering latent biomarkers through deep learning and explainable AI

Breast cancer, characterized by its aggressive pro-gression and high mortality rates, continues to be among the most common types of cancer. While early detection significantly improves survival outcomes, the current standard diagnostic tool, mammography, presents limitations including high costs and radiation exposure. Thermography emerges as a promising alternative, offering a non-invasive and cost-effective screening approach. This study aims to develop machine learning models based on convolutional neural networks to analyze multiple thermal breast views from the Visual DMR dataset for cancer detection. Our research methodology leverages advanced image analysis techniques to enhance diagnostic accuracy. Our findings show that our approach achieves strong model performance, highlighting the effectiveness of image-based diagnostics in achieving reliable result. Through extensive experimentation and careful model selection, an architecture was selected that achieved strong diagnostic performance. Our final finetuned VGG16 model demonstrated a training accuracy of 95.86% and Area Under the Receiver Operating Characteristic Curve (AUROC) of 99.63 with a test accuracy of 92.7%. Key metrics included an F1 score of 91.73%, precision of 91.04% and a sensitivity of 92.42%. These results underscore the model's robustness and reliability, making it a promising tool for accurate case identification. By leveraging Explainable AI (SHAP), model interpretability was further en-hanced, offering clear insights into its decision-making process. Given that thermography remains underutilized in breast cancer diagnostics, with limited public datasets available, this research contributes to the growing body of evidence supporting its potential as a viable screening method and aims to stimulate further investigation in this promising field.