EIR – sistema inteligente para detecção e classificação de calcificações em mamografias

EIR Intelligent System for Detection and Classification of Calcifications in Mammograms Author: João Paulo Virǵılio Marinho Martins Supervisor: Prof(a). Dra. Lourdes Mattos Brasil Co-supervisor: Dra. Janice Magalhães Lamas Post-Graduation Program in Biomedical Engineering – Qualify of Master Degree BRASÍLIA/DF 2016. The breast cancer it is the cancer that kills more women in Brazil. The early diagnosis of breast cancer is still the best weapon against this disease and the main technique used for the identification of a possible carcinoma is the screening mammogram. The screening mammogram is used for asymptomatic women, the goal of this exam is to find breast cancer before their growth and spread in the breast. This is due to the fact that most tumors are detected, mammographically by the presence of calcification. Nevertheless a correct diagnosis of a possible calcification is very complex because of the high levels of noise and high density of breast tissues in the mammography, which may cause decision errors by the expert who analyzes the image. To tackle this problem and enable more efficient and accurate diagnosis, Computer Aided Diagnosis Systems (CAD) have been increasingly studied. The main purpose of this work is to develop a CAD system, named EIR, for location, feature extraction and classification of regions of interest in mammograms of asymptomatic women subjected to screening mammogram. The system was trained with 37 mammograms, classified in category 4 of Breast Imaging-Reporting and Data System (BIRADS), from these, 85 regions of interest were extracted by a radiologist. To ensure a better training for EIR, the dataset were artificially augmented through affine transformations, generating a total of 255 images for training. The EIR training was evaluated through various metrics, such as accuracy, precision, recall and F1 Score. At the ending of the training phase, EIR achieved a rate of 99% of accuracy, 99% of precision, a sensitivity of 99% and a F1 Score of 0.99 with the test set. These results suggest a potential relevance to assist, as a second opinion or support, the expert in the diagnose of breast cancer that is manifested by the presence of calcifications in mammograms. In the next stages of the research it is intended increase the quality of the used models, develop a user-friendly graphical interface for the use of experts and the parallelization of the code for a faster training and prediction is intended. Key-words: Machine Learning, Feature Extraction, Feature Classification, CAD System.