Avaliação de arquiteturas de redes neurais convolucionais na identificação de tumores cerebrais por ressonância magnética

Abstract

Brain cancer is an aggressive type of neoplasm that requires early diagnosis for effective treatment. Magnetic resonance imaging (MRI) is a non-invasive diagnostic imaging technique widely used for the visualization of brain tumors due to its ability to provide detailed images of brain structures. However, the early and accurate identification of different tumor types remains a challenge due to their visual similarities and the reliance on invasive diagnostic methods. This study evaluates convolutional neural network architectures for classifying brain tumors in magnetic resonance images, aiming to improve diagnostic accuracy and efficiency. The DenseNet201, EfficientNetB0, and ResNet50 models were employed using Transfer Learning and fine-tuning techniques and applied to public datasets containing images distributed into four classes: meningioma, glioma, pituitary tumor, and no tumor. The images underwent a preprocessing stage that included resizing, grayscale conversion, and bilateral filtering. During the training stage, k-fold cross-validation was employed to assess the stability and generalization capability of the models. After training, the final evaluation of the proposed models was performed using a public dataset previously unseen by the models. Accuracy, precision, recall, F1-score, and AUC were used to quantify model performance. EfficientNetB0 achieved an accuracy of 75.63%, a precision of 86.52%, and an F1-score of 73.30%. ResNet50 achieved an accuracy of 75.13% and the highest AUC (92.86%), while DenseNet201 achieved an accuracy of 73.10% and an F1-score of 67.82%. The models exhibited sensitivity ranging from 69.04% to 73.46% and specificity above 91% across all architectures. The results demonstrate the capability of the models to classify brain tumors in magnetic resonance images, with EfficientNetB0 achieving the best overall performance.