Supporting the Diagnosis of Brain Cancer in Post-treatment Patients Using Ensemble Learning and Transfer Learning Techniques

Brain cancer is a serious condition characterized by the uncontrolled growth of abnormal cells, where delay in diagnostic testing represents a significant challenge in global health. Deep Learning has advanced in biomedical applications, with Convolutional Neural Networks (CNN) showing promise for identifying and classifying brain tumors. Some recent studies propose based on Deep Learning approach to predict brain images and obtain preliminary diagnoses, developing Artificial Intelligence (AI) models to detect brain tumors using CNN and distinguish between different types of tumors from MRI scans. The goals of this work include developing methods for brain cancer diagnosis using Ensemble and Transfer Learning techniques, as well as validating these methods using Machine Learning quality measures. In addition, postreatment of gliomas is being carried out through segmentation, evaluating its effectiveness. For the training of the method, only the last folder of each patient has been taken to evaluate the effectiveness of the segmentation. The proposed methodology combines the study of images, Machine Learning and Deep Learning to predict and detect brain cancer. Normalization techniques, Data augmentation, and U-Net V2-FT models are used for accurate segmentation. Metrics such as precision, recall, F1-score and Dice coefficient are evaluated to validate the effectiveness of the model. The BraTS 2024 dataset, with 1350 brain scan samples, is used to train and evaluate the model. The results show an effective tool for the diagnosis of brain cancer, improving performance and reducing the time and resources required for diagnosis. For future work, the implementation of additional postreatment techniques and evaluation in other types of brain tumors is suggested.