A Degradation-Robust Deep Learning Framework for MRI Brain Tumor Diagnosis

Ricardo Bauchspiess, Mylène C. Q. Farias

Abstract: Brain Magnetic Resonance Imaging (MRI) is a reliable imaging technique for the diagnosis of brain tumors. However, the accuracy of tumor detection based on brain magnetic resonance images may be affected by the expertise and experience of healthcare professionals, particularly in the early stages of the disease. To address this, several deep learning systems have been proposed for the detection and classification of brain tumors. However, the presence of common MRI artifacts can reduce their precision and reliability. We propose a deep learning-based framework for brain tumor diagnosis that both improves the quality of input MRI scans and classifies the type of brain cancer. Specifically, a Uformer-based restoration model is used to mitigate artifacts before they are fed into the classification model. The framework was designed by creating a dataset of brain MRI scans with combinations of artifacts and designing a way to train and test deep learning models. Results show that although tumor classification is greatly affected by MRI artifacts, the framework is able to improve accuracy. Although this work targets brain tumors, the proposed framework can be applied to any diagnosis problem.