Detection of Breast Cancer in Thermal Images Using Convolutional Neural Network

Resumo

Worldwide, breast cancer is the most commonly diagnosed cancer in women. Early detection of this
type of cancer can help women to have a more appropriate treatment and, consequently, reduce the mortality rate.
Today, there are several techniques and algorithms for the diagnosis of breast cancer, but techniques that provide
greater agility in diagnosis and precision in results are still widely studied. Thermography is a recent technique to
record the image of the breast, measuring the temperature based on infrared radiation, and has been an attraction
for research. In this context, a Convolutional Neural Network (CNN) was set up to process a data set of thermal
breast image inputs in order to classify healthy and breast cancer patients with good accuracy. The data were
obtained from the Mastology Database (DMR), with a total of 5,602 images divided into 80% for training and
20% for validation. The CNN architectures used in the experiments were Xception, ResNet101, ResNet101V2,
MobileNet, MobileNetV2, DenseNet201 and InceptionV3. Each architecture had the same configurations, with a
learning rate of 0.001, using SGD and a maximum of 20 times. The results found showed that the best proposed
CNN architecture was Xception with an accuracy of 95.89%, while InceptionV3 obtained an accuracy of 94.73%
and DenseNet201 had an accuracy of 93.22%.