Deep Classification of Glioma Grade using 3D Wavelet Features

Abstract

Accurate grading of a brain tumor helps the physician to choose the most appropriate dosage of chemotherapy and radiation therapy to treat the brain tumor, which increases the prognosis of the disease. The purpose of this study is to classify a brain tumor into low-grade glioma and high-grade glioma automatically before brain surgery. Normally, a biopsy of a brain tumor is done to predict the grade of glioma. However, a biopsy of the brain may eventually result in adverse morbidity or mortality. Therefore, to diagnose tumor grade before surgery, Magnetic Resonance Imaging is used. Fully automated prediction of glioma grading based on MRI thereby reduces the adverse effects caused by biopsy. Since the grade of a tumor is determined conventionally based on the histological appearance of the tissues, wavelet-based radiomic features and unfiltered image features from multimodal brain 3D MR images are extracted. To enhance the classification accuracy, the best subset of these extracted features are selected using the sequential forward feature selection technique. In the forward feature selection method, different induction algorithms are experimented with to determine the best feature subset. These selected features are then trained and tested using a deep neural network classifier with the Adam optimization algorithm. Performance of the DNN classifier model is evaluated using F1-score, precision, recall and accuracy. Deep neural network classifier model using 3D wavelet features selected with random forest-based forward feature selection outperforms the existing state-of-the-art machine learning classifier models.