Identificação de anomalias nas mamas através da extração de características de termogramas

Abstract

For some years, Thermography has been used with more frequency in several medical areas, like auxiliary tool by monitoring and detecting breast cancer, however, it still faces several limitations. This work aims to classify patients as healthy or sick people through breast thermography in 18 patients. The images were acquired by VISUAL-LAB database, which uses the images of patients from University Hospital Antonio Pedro located at Federal Fluminense University (UFF). An infrared camera, FLIR SC-620 model, with resolution of 640X480, was used for capturing the image of the patients. Every thermal image has a correspondent temperature matrix that associates every image pixel in one temperature. The Regions of interests (ROIs) are regions of an image which demonstrate significant characteristics for an analysis. In Thermography, ROIs corresponds to the left and right patients’ breasts. The segmentation of the Region of Interest (ROIs) was made through the manual use of GIMP. Considering that the body is divided symmetrically by sagittal plane, it is considered that left and right breast temperature distribution is symmetric too. In this context, analysis of breast symmetry, characteristics were divided in three huge groups. The groups used were: characteristics based in statistics, fractal geometric and histogram. The methods implementation to extraction of characteristics were made through IDE Microsoft Visual C++ and OpenCV library, by using SVM as classifier, which is a linear classifier based in statistical theories. The cross validation was used to evaluate the model generalization capacity. The data were analyzed through metrics of evaluation like Sensibility, Specificity, Youden index, Accuracy and Area under the curve ROC. The finals results achieved 77% of accuracy, 69% of sensibility, 100% of specificity, 0,6923 of Youden indexes and 0,694 of area under the curve ROC.