Conference Proceedings

International E-Conference on Engineering, Technology and Management - ICETM 2020

TRIPLANAR-CNN FOR AUTOMATED GRADING OF GLIOMAS USING PREOPERATIVE MULTI-MODAL MR IMAGES

Author(s) : Abdela Ahmed Mossa, Ulus Cevik

Abstract

Glioma has been one of the most common life-threatening brain tumor diseases all over the world with different levels of aggressiveness: Low Grade Glioma (LGG) and High Grade Glioma (HGG), and consequently automated glioma grade prediction methods based on multi-modal MRI images are of great interest. However, the development of effective automated methods, and in particular convolutional neural networks (CNN) for fast and accurate medical image analysis has relied on the availability of large annotated training datasets. The purpose of this study was to develop a 2D CNN model, Triplanar-CNN, to a fully automated and accurate glioma grade prediction, using a small training dataset of less than 300 glioma patients who underwent pre-operative volumetric MRI exams, which included FLAIR, T1Gd, T1, and T2 modalities. Our approach operates on all of the MRI modalities and plane slices (axial, coronal, and sagittal) based on reconstructing the volumetric MRI as a set of 2D stacked slices in the sagittal, coronal and axial planes, and allows to leverage pre-trained CNN models for feature extraction, which is essential given the inadequate amount of labeled training dataset. The proposed Triplanar-CNN architecture consists of three sub-networks, each based on leveraging CNN model pre-trained on natural images, and separately applied to axial, coronal and sagittal view of a 3D MRI, respectively, followed by a common fusion-layer to integrate the extracted features by each sub-networks, which is an input to a fully connected layer used for prediction. On the BraTS 2017 dataset, the Triplanar-CNN were trained separately for each modality, and each corresponding model yields more than 0.9 AUC for classifying glioma into two groups. Moreover, averaging the probability of glioma grading by all four MRI modalities boosts the classification performance compared to either of the four separately, achieved a patient-level grading result of 95.8% and 0.985 in accuracy and AUC, respectively, outperforming state-of-the-art results. Five-fold-cross validation was used to evaluate the models. To sum up, we developed a fully automated CNN-based model that can be translated into a clinical tool for non-invasive diagnosis of glioma using pre-operative MRI scans in a rapid and accurate way, leading to better outcomes for patients

Conference Title : International E-Conference on Engineering, Technology and Management - ICETM 2020
Conference Date(s) : 31, May 2020
Place : Online (Via Video Conference)
No fo Author(s) : 2
DOI : 10.15224/978-1-63248-188-7-05
Page(s) : 21-27
Electronic ISBN : 978-1-63248-188-7
Views : 267   |   Download(s) : 186