Identification of mammary lesions in thermographic images: feature selection study using genetic algorithms and particle swarm optimization

The incidence of breast cancer increases every year. Early detection of the disease is critical since the sooner the disease is discovered the better are the treatments and the chances of cure. Mammography is now the gold standard for the diagnosis of breast cancer, but this screening tool has some limitations. Infrared thermography is being studied as a complementary tool due to its benefits. The combination of specialized professionals with methods of digital image analysis in breast thermography can contribute to improve diagnosis performance. From this, several research groups have been proposing methods to treat these data. Feature selection plays a fundamental role in this process, since it may optimize the machine learning process. In this study, we propose a feature selection approach using genetic algorithms (GA) and particle swarm optimization (PSO) in thermographic images with breast lesions. The main goal of this approach is to optimize the identification and classification of breast lesions. We used several classifiers to assess the performance of the subsets with selected features. Support vector machines were more effective in these experiments. It was possible to reduce from 169 features with accuracy of 91.12% to 57 features with accuracy of 87.08% using the genetic algorithm. We also found a subset of 60 features with an accuracy of 86.16% using PSO. The features selection step could optimize classification time and computational cost. By varying input parameters, we could achieve a significant reduction in the amount of features. Features reduction could be done without losing much in accuracy, when compared to the whole set of features.