NIMG-27. GLIOBLASTOMA TUMOR SEGMENTATION USING DEEP CONVOLUTIONAL NEURAL NETWORKS

Glioblastoma (GBM), the most malignant primary brain cancer, is mainly diagnosed initially based on tumor appearance on magnetic resonance (MR) images. Segmentation of GBM tumors is an important early step in image analysis to characterize tumor phenotypic features. However, manual segmentation performed by radiologists is a laborious and time-consuming process. Also, automated lesion segmentation is a challenging task due to the heterogeneous appearance of GBM. Traditional image segmentation methods may not consistently show robust performance. Recently, deep convolutional neural networks (CNNs) demonstrate great potential in image classification and recognition tasks. Here we applied a CNN architecture called U-Net that had recently been developed to segment neuronal structures in electron microscopic images and had shown good performance in segmenting lung lesions on CT images, which had not been applied to brain lesion segmentation on MR images. We trained the CNN framework to segment whole tumors on 2-dimentional MR contrast-enhanced T1 axial images. We acquired hand-drawn regions of interest (ROI) circumscribing whole GBM tumors for a total of 8280 images from 490 cases in two independent cohorts (from the Stanford University Medical Center and the Cancer Genome Atlas), with multiple tumor image slices from each case. We trained our CNN segmentation framework on 80% of the combined dataset, evaluated on 10%, and tested on the remaining 10%. After training for 20 epochs, our framework achieved a mean dice coefficient of 79% on the test set, which is comparable in performance to the published state-of-the-art brain tumor segmentation deep learning algorithm evaluated on another dataset. Further analysis needs to be performed to compare the two frameworks using the same dataset. This deep learning framework in our analysis may be extended to segmenting lower grade gliomas or other solid tumors visualized on medical images.