Convolutional Neural Networks for breast cancer detection using regions of interest from infrared images

Abstract

Breast cancer is the most common cancer among women in Iraq and worldwide. The incidence in young women has been increasing over the years, and the gold standard test for diagnosis, mammography, is contraindicated for people under 40. Thermography appears in this scenario as a promising technique for early detection and a higher survival rate in this group of women. The analysis of thermographic images by Convolutional Neural Networks has good results in increasing the reliability and sensitivity of diagnoses. This work uses the Densenet201 and Resnet50 networks based on 72 images of different patients, 38 of whom are sick and 38 healthy. These images underwent pre-processing before being analyzed, and in one of the pre-processing steps, there was a manual clipping of only the region of interest of the breasts, intending to evaluate whether detection is superior to the image whole. The best average accuracy rate was obtained with the Densenet201 network, learning rate of 0.001 and 30 epochs, which reached 89%. Regarding the F1-score, the network with the best performance was Resnet50, with a learning rate of 0.0001 and 30 epochs, which reached 76%.