Análise do desempenho de redes neurais profundas na classificação e segmentação de tumores cerebrais em imagens de ressonância magnética

Abstract

The development of malignant brain tumors characterizes brain cancer. According to the National Cancer Institute (INCA), approximately 11,000 new cases are registered in Brazil every year, with a mortality rate of approximately 84% of cases. In this context, early detection of brain tumors can considerably increase the survival rate of patients. Advances in Artificial Intelligence (AI) have improved the analysis of medical images. However, the classification and segmentation of brain tumors, using magnetic resonance imaging (MRI), are still challenging tasks due to several factors, such as changes in contrast, size, shape, position and variation of the tumor region, depending on the patient. Therefore, the main objective of this work is to develop a training processing flow and analyze the performance of deep neural network architectures in the classification of brain tumors in Meningioma, Glioma, Pituitary, as well as cases without tumor and the detection and segmentation of the tumor region. For this, three publicly available datasets from the literature of magnetic resonance images were used. In the classification stage, fifteen pre-trained deep learning models were trained. For validation, quantitative metrics widely applied in classification problems were assumed. Then, a thorough analysis was conducted using statistical tests to verify significant differences between the architectures. In addition, to segment the brain tumors, the UNet, UNet++, and FPN networks were applied in combination with the best-performing encoders. Therefore, the results indicated that EfficientNetB7 presented the best classification performance, with an accuracy of 97.68%, precision of 97.63%, recall of 97.69%, F1-Score of 97.64%, and specificity of 99.21%. In segmentation, the combination of EfficientNetB7 with FPN was the most effective, achieving 99.52% accuracy, 85.23% in F1-Score, 74.29% for the Intersection over Union metric and 4.56 in Hausdorff Distance. The study presented significant results compared to what is explored in the literature, evidencing the effectiveness in the detection and classification of brain tumors.