Authors: SONAL GORE, JAYANT JAGTAP
Abstract: As per 2016 updates by World Health Organization (WHO) on cancer disease, gliomas are categorized and further treated based on genomic mutations. The imaging modalities support a complimentary but immediate noninvasive diagnosis of cancer based on genetic mutations. Our aim is to train a deep convolutional neural network for isocitrate dehydrogenase (IDH) genotyping of glioma by auto-extracting the most discriminative features from magnetic resonance imaging (MRI) volumes. MR imaging data of total 217 patients were obtained from The Cancer Imaging Archives (TCIA) of high and low-grade gliomas. A 3-pathway convolutional neural network was trained for IDH classification. The multipath neural network, consisting of one shallow and two deep neural network paths, is used to auto-extract the significant imaging features for successful IDH discrimination into IDH mutant and wild type. An accuracy of 93.67% and cross-entropy loss of 0.052 is achieved for IDH classification. The results of 3-pathway convolutional neural network (CNN) are better than the results achieved from individual paths of 3-pathway model. The results have demonstrated the multipath convolutional neural networks as state-of-the-art method with simple design to predict IDH genotypes in glioma with auto-extraction of radiogenomic features.
Keywords: Glioma, magnetic resonance imaging, radiogenomic analysis, multipath convolutional neural network, isocitrate dehydrogenase
Full Text: PDF