Classifying breast lesions in Brazilian thermographic images using convolutional neural networks

Breast cancer is the leading cause of death from malignant tumors in women worldwide. Early diagnosis is essential for the treatment and cure of patients. Breast anomalies, such as cysts, cancers and benign tumors, show an increase in blood supply in their region, causing temperature variations in the area, which can be detected through thermographic images. Thermography has shown to be a promising tool in the detection of breast cancer as it is low cost, harmless to the patient and it can be performed in younger people, whose breast tissue is denser, making the diagnosis more difficult through mammography, which is currently the gold standard for detecting this disease. The aim of this work is to develop a computer vision technique based on a convolutional neural network in order to detect breast cancer using thermographic images. Thus, a single dataset with thermographic data obtained from 97 patients was used with two different class assignments. First, the dataset was separated into three classes: benign, malignant and cyst, resulting in a global error rate of 7.5% and a sensitivity of 98.46%. Afterward, a binary classification was performed in order to label the images into cancer and non-cancer, obtaining a 21.94% global error rate and 81.66% sensitivity. The method proposed in this work had the best performance in both cases when compared with the results obtained by existing algorithms in the literature.