Classification of Isocitrate Dehydrogenase (IDH) Mutation Status in Gliomas Using Transfer Learning

Abstract

Isocitrate Dehydrogenase (IDH) mutation is a significant genetic alteration that is found in brain tumors. Its diagnosis is vital for the prognosis of low-grade Glioma and secondary grade Glioma patients. Physicians used invasive methods to diagnose the Gliomas, which was an unsafe method but now advanced magnetic resonance imaging techniques are being used for tumor visualization and treatments. Classical machine learning and deep learning methods have been used by some studies for the problem of IDH mutation status detection using magnetic resonance images. Recent studies have used concatenation of deep and handcrafted features to achieve superior performance. This study has used the concatenations of the deep features extracted through pre-trained convolution neural networks (CNNs) for the task of IDH mutation status detection using magnetic resonance images. To select the pre-trained CNNs, five top accuracies on the ImageNet dataset were considered. Magnetic resonance images were acquired from The Cancer Genome Atlas Glioblastoma Multiforme and The Cancer Genome Atlas Low-Grade Glioma. All experiments (performed using features extracted from each CNN and their concatenation) were compared with each other and state-of-the-art. The proposed technique achieved 99% accuracy while being efficient in terms of data and computational resources.