Abstract : Breast cancer continues to be among the major causes of cancer mortality among women worldwide, with more than 2 million cases and around 680,000 deaths recorded in 2020, as per the World Health Organization (WHO). Early diagnosis is very important, especially in low- and middle-income countries (LMICs) where diagnostic facilities are scarce. In this paper, a non-invasive breast cancer classification system based on thermal infrared imaging and deep learning is suggested. The DMR-IR dataset, which includes ground-truth thermal images, was utilized to train an ensemble model combining VGG16 and EfficientNet using transfer learning. Image enhancement methods and Grad-CAM visualizations were employed to enhance interpretability. The ensemble model had a classification accuracy of 99.8% and an AUC value of 1.00. These outcomes demonstrate high potential for precise early detection. The research shows how interpretable AI can aid radiologists during diagnosis with a decrease in reliance on intrusive procedures, particularly in healthcare settings with limited resources.