Comparative deep learning CNN architectures for breast cancer detection from thermal imaging

Abstract

It has been observed that breast cancer is a severe disease among women globally. Mammography is the most effective screening method for detecting this severe illness. Over the last thirty years, mammography has been widely recognized as a preventive measure against breast cancer. In recent years, convolutional neural networks (CNN) and artificial intelligence (AI) have become more common in digital mammography for automated breast cancer detection. For classifying breast cancer, this study examines the five CNN models: LeNet-5, AlexNet, VGG-16, ResNet-50, and Inception-v3, using the database for mastology research with infrared images (DMR-IR) dataset's thermal image. These models were trained and validated using accuracy, recall, F1-score, specificity, and AUC as evaluation criteria after the dataset was preprocessed using normalization and data augmentation. Among the experimental models, Inception-v3 achieved 99.44% accuracy, outperforming other CNNs by 1–2%, while other models performed with accuracy levels above 97%. These results show the tremendous efficacy of CNN-based deep learning methods for breast thermogram analysis. The research points out thermography as a useful support for traditional imaging and InceptionV3 as a potential option for correctly detecting clinical breast cancer.