Enhancing Breast Cancer Detection from Thermographic Images: A Hybrid Approach Using Transfer Learning and Generative Adversarial Networks

Based on the most recent statistics, breast carcinoma stands as the most widespread cancer globally, claiming the lives of nearly 900,000 individuals annually. Detecting this disease at an early stage and providing an accurate diagnosis can enhance the likelihood of favorable outcomes, subsequently reducing the mortality rate. Early diagnosis indeed plays a role in hindering its proliferation and safeguards individuals from falling victim prematurely. When researchers delve into the differentiation of benign and malignant tumors, and when they seek to draw conclusions about early and advanced stages of breast cancer, they confront a myriad of challenges. The limited availability of training data samples has been recognized as a fundamental challenge in this field. Due to the novelty of this research area, large datasets for breast cancer detection using thermography are not yet available, which necessitates the use of supplementary methods to compensate for the data scarcity. In this paper, to enhance the model and address the data scarcity challenge in breast cancer detection from thermographic images, we employed a combination of transfer learning and the application of Generative Adversarial Network (GAN) neural networks on the well-known DMR-IR dataset. The performance of neural networks, with and without this combination, was tested and compared. The results demonstrate that the utilization of the ResNet-152 transfer learning model on this dataset achieved an Accuracy of 83% in breast cancer detection. Furthermore, after applying the GAN neural network, the Accuracy of this same scenario increased to 90%.