Abstract
BACKGROUND AND PURPOSE: Accurate determination of glioma grade leads to improved treatment planning. The criterion standard for glioma grading is invasive tissue sampling. Recently, radiomic features have shown excellent potential in glioma-grade prediction. These features may not fully exploit the underlying information in MR images. The objective of this study was to investigate the performance of features learned by a convolutional neural network compared with standard radiomic features for grade prediction. MATERIALS AND METHODS: A total of 237 patients with gliomas were included in this study. All images were resampled, registered, skull-stripped, and segmented to extract the tumors. The learned features from the trained convolutional neural network were used for grade prediction. The performance of the proposed method was compared with standard machine learning approaches, support vector machine, random forests, and gradient boosting trained with radiomic features. RESULTS: The experimental results demonstrate that using learned features extracted from the convolutional neural network achieves an average accuracy of 87%, outperforming the methods considering radiomic features alone. The top-performing machine learning model is gradient boosting with an average accuracy of 64%. Thus, there is a 23% improvement in accuracy, and it is an efficient technique for grade prediction. CONCLUSIONS: Convolutional neural networks are able to learn discriminating features automatically, and these features provide added value for grading gliomas. The proposed framework may provide substantial improvement in glioma-grade prediction; however, further validation is needed.