RESUMO
Early detection and diagnosis of brain tumors are crucial to taking adequate preventive measures, as with most cancers. On the other hand, artificial intelligence (AI) has grown exponentially, even in such complex environments as medicine. Here it's proposed a framework to explore state-of-the-art deep learning architectures for brain tumor classification and detection. An own development called Cross-Transformer is also included, which consists of three scalar products that combine self-care model keys, queries, and values. Initially, we focused on the classification of three types of tumors: glioma, meningioma, and pituitary. With the Figshare brain tumor dataset was trained the InceptionResNetV2, InceptionV3, DenseNet121, Xception, ResNet50V2, VGG19, and EfficientNetB7 networks. Over 97 % of classifications were accurate in this experiment, which provided a network's performance overview. Subsequently, we focused on tumor detection using the Brain MRI Images for Brain Tumor Detection and The Cancer Genome Atlas Low-Grade Glioma database. The development encompasses learning transfer, data augmentation, as well as image acquisition sequences; T1-weighted images (T1WI), T1-weighted post-gadolinium (T1-Gd), and Fluid-Attenuated Inversion Recovery (FLAIR). Based on the results, using learning transfer and data augmentation increased accuracy by up to 6 %, with a p-value below the significance level of 0.05. As well, the FLAIR sequence was the most efficient for detection. As an alternative, our proposed model proved to be the most effective in terms of training time, using approximately half the time of the second fastest network.
RESUMO
In the last decade, neurosciences have had an increasing interest in resting state functional magnetic resonance imaging (rs-fMRI) as a result of its advantages, such as high spatial resolution, compared to other brain exploration techniques. To improve the technique, the elimination of artifacts through Independent Components Analysis (ICA) has been proposed, as this can separate neural signal and noise, opening possibilities for automatic classification. The main classification techniques have focused on processes based on typical machine learning. However, there are currently more robust approaches such as convolutional neural networks, which can deal with complex problems directly from the data without feature selection and even with data that does not have a simple interpretation, being limited by the amount of data necessary for training and its high computational cost. This research focused on studying four methods of volume reduction mitigating the computational cost for the training of 3 models based on convolutional neural networks. One of the reduction techniques is a novel approach that we call Reduction by Consecutive Binary Patterns (RCBP), which was shown to preserve the spatial features of the independent components. In addition, the RCBP showed networks in components associated with neuronal activity more clearly. The networks achieved accuracy above 98 % in classification, and one network was even found to be over 99 % accurate, outperforming most machine learning-based classification algorithms.