Enhancing Breast Cancer Detection through Thermal Imaging and Customized 2D CNN Classifiers

Abstract

Breast cancer is one of the most prevalent and life-threatening forms of cancer due to its aggressive nature and high mortality rates. Early detection significantly improves a patient's chances of survival. Currently, mammography is the preferred diagnostic method, but it has drawbacks such as radiation exposure and high costs. In response to these challenges, thermography has become a less invasive and cost-effective alternative, gaining popularity. We aim to develop a cutting-edge model for breast cancer detection based on thermal imaging. The initial phase involves creating a customized machine-learning (ML) model built on convolutional neural networks (CNN). Subsequently, this model undergoes training using a diverse dataset of thermal images depicting breast abnormalities, enabling it to identify breast cancer effectively. This innovative approach promises to revolutionize breast cancer diagnosis and offers a safer and more accessible alternative to traditional methods. In our recent study, we leveraged thermal image processing techniques to forecast breast cancer precisely based on its external manifestations, particularly in cases where multiple factors are interconnected. This research employed various image classification methods to categorize breast cancer effectively. Our comprehensive approach encompassed segmentation, texture-based feature extraction from thermal images, and subsequent image classification, leading to the successful detection of malignant images. Our study harnessed the power of machine learning to create a tailored classifier, merging key components from GoogleNet, including the utilization of 2D CNNs and activation functions, with the ResNet architecture. This hybrid approach incorporated batch normalization layers following each convolutional layer and employed max-pooling to enhance classification accuracy. Next, we used a sample dataset of carefully selected images from DMR-IR to train our proposed model. The outcomes of this training demonstrated significant improvement over existing methods, with our suggested 2D CNN classifiers achieving an impressive classification rate of 95%, surpassing both the SVM and current CNN models, which achieved rates of 91% and 71%, respectively.