Deep learning for magnetic resonance images of gliomas

Abstract

Gliomas are tumors that arise in the glial cells of the brain and spine. Gliomas are one of the most common brain cancers, and comprise 80% of all malignant brain tumours. Gliomas are classified by cell type, grade, and location. Prognosis of patients presenting with high grade gliomas remains poor. The gold standard for grading of gliomas remains histopathology, but a working radiological diagnosis can be established from a magnetic resonance imaging (MRI) scan. MRI is typically used throughout the patient pathway because routine structural imaging provides detailed anatomical and pathological information. In addition, advanced techniques can provide additional physiological detail. Traditionally, MRIs were read exclusively by radiologists, but improvements in machine learning has sparked considerable interest in its application to enhanced and automated diagnostic tools. Machine learning approaches are also of interest in monitoring the progression of low grade gliomas, and in monitoring of patients undergoing treatment. Convolutional neural networks (CNNs), especially when trained using transfer learning, have been shown to grade gliomas with up to 94% accuracy. Given an MR image of a brain tumour, we wish to manually segment the various tissues to aid diagnoisis and other assessments. This manual process is difficult and laborious; hence there is demand for automatic image segmentation of brain tumors. Public datasets and the BRATS benchmark have enabled clearer and more objective comparison of segmentation techniques. State-of-the art automated segmentation of gliomas is currently represented by deep learning methods. Deep learning is also now on the rise for prediction of molecular biomarkers. Novel approaches to explainable artificial intelligence are now required to aid the extraction of novel useful features from machine learning approaches. Finally, CNNs have been developed to make predictions of patient survival times. There are many exciting new directions for this field, from novel CNN architectures, to the integration of information from advanced MRI and complementary imaging modalities and spectroscopic techniques. In time, this may lead to clinically acceptable automation of a variety of radiological determinations, with positive consequences for patient outcomes. In this chapter, we will give an overview for engineers and computer scientists of the deep learning applications used in glioma detection, characterization/grading and overall survival prognosis of the patients. We will highlight the limitations and challenges of deep learning techniques as well as the potential future of these methods in prognosis, clinical diagnostics, and decision making.