Transfer Learning for Breast Cancer Classification in Terahertz and Infrared Imaging

Proactive treatment of cancer, characterized by early detection and intervention is one of the main focus of the next-generation healthcare systems for predictive, timely detection and seamless, customized care. Recent breakthroughs in new molecular imaging tools that are noninvasive and nonionizing coupled with artificial intelligence-enabled healthcare systems through data analytic techniques enable the realization of these systems. Breast cancer is the most common and prevalent cancer in women. The development of new imaging modalities based on low photon radiation, particularly the Terahertz and Infrared radiation-based imaging, shows potential for early and more accurate cancer detection with an improved treatment exploitation. Due to the unavailability of sufficient training datasets, the research on the potential application of deep learning techniques for the analysis of image datasets in this domain has not yet been fully explored. Nevertheless, the convolutional neural networks (CNNs) are capable of complex visual patterns recognition but their performance depends on the availability of large datasets, however, the transfer learning (TL) based on pre-trained CNN models allow improved accuracy performance over the smaller datasets. In this paper, we have explored the deep learning frameworks based on TL through fine-tuning of the pre-trained CNN architectures: ALEXnet, VGG16, ResNet18, ResNet50, GoogleNet, VGG19, MobileNetv2 and Densenet201 on a thermal breast cancer image dataset. The classification performances of the models are compared with respect to the evaluation metrics used. VGG16, Googlenet, Resnet50 and VGG19 outperformed the rest of the models with an accuracy of 100 %.