Modelo semi-supervisionado aplicado à previsão da eficiência da quimioterapia neoadjuvante no tratamento de câncer de mama

Abstract

Breast cancer is the most common cancer kind in women in all the world and it is the second most common kind of cancer, behind only of lung cancer. Exist many factors that can increase or reduce the risk of cancer like healthily alimentation and body exercises. However prevention is the most effective weapon against breast cancer. It's early detection increases patient cure chances and reduces exposition to damages of the treatment. The breast cancer treatment has three stages, consisting in pre-operative chemotherapy, surgery and post-operative chemotherapy. As is known, chemotherapy imposes too much suffering to the patients, and, in the case of pre-operative chemotherapy, only 30% of them show some response to the treatment. It's desirable to have some way to preview the pre-operative chemotherapy response of the patients before they are subjected to it, avoiding unnecessary sufferings. Clinical examinations are not enough to take the decision of subject the patient to the treatment or not, but nowadays new technics allows access to patient genetic information that can be used to try solve this problem and avoid them to suffers unnecessarily. Euler Horta's work cite{Horta2008} separeted the most important genetic feature to the breast cancer pre-operative chemotherapy efficiency preview problem. Our work is a semi-supervised method applied to data set defined in cited work. Patient genetic feature was collected in the USA and France. Some of this data set will be considered as labeled patterns and other as unlabeled patterns to be used in the semi-supervised method presented at this work. The idea is train one MLP network with the labeled data set, and use geometric margin calculated value for both labeled and unlabeled data set together, to select one solution that well classifies both data set. This work proposal is evaluate margin geometrically to be the distribution feature to the semi-supervised method, however, the big difficult is on evaluate it to the unlabeled data set. Was necessary develop some way to identify the patterns limits between class to evaluate margin geometrically. Was used the FCM method and the feature space slicing proposed in this work. This method called extit{geometric margin based semi-supervised method (MSMG from portuguese método semi-supervisionado baseado na margem geométrica)}, had god results and inspires new approaches and tests.