A comparative study for glioma classification using deep convolutional neural networks

Abstract

Gliomas are a type of central nervous system (CNS) tumor that accounts for the most of malignant brain tumors. The World Health Organization (WHO) divides gliomas into four grades based on the degree of malignancy. Gliomas of grades I-II are considered low-grade gliomas (LGGs), whereas gliomas of grades III-IV are termed high-grade gliomas (HGGs). Accurate classification of HGGs and LGGs prior to malignant transformation plays a crucial role in treatment planning. Magnetic resonance imaging (MRI) is the cornerstone for glioma diagnosis. However, examination of MRI data is a time-consuming process and error prone due to human intervention. In this study we introduced a custom convolutional neural network (CNN) based deep learning model trained from scratch and compared the performance with pretrained AlexNet, GoogLeNet and SqueezeNet through transfer learning for an effective glioma grade prediction. We trained and tested the models based on pathology-proven 104 clinical cases with glioma (50 LGGs, 54 HGGs). A combination of data augmentation techniques was used to expand the training data. Five-fold cross-validation was applied to evaluate the performance of each model. We compared the models in terms of averaged values of sensitivity, specificity, F1 score, accuracy, and area under the receiver operating characteristic curve (AUC). According to the experimental results, our custom-design deep CNN model achieved comparable or even better performance than the pretrained models. Sensitivity, specificity, F1 score, accuracy and AUC values of the custom model were 0.980, 0.963, 0.970, 0.971 and 0.989, respectively. GoogLeNet showed better performance than AlexNet and SqueezeNet in terms of accuracy and AUC with a sensitivity, specificity, F1 score, accuracy, and AUC values of 0.980, 0.889, 0.933, 0.933, and 0.987, respectively. AlexNet yielded a sensitivity, specificity, F1 score, accuracy, and AUC values of 0.940, 0.907, 0.922, 0.923 and 0.970, respectively. As for SqueezeNet, the sensitivity, specificity, F1 score, accuracy, and AUC values were 0.920, 0.870, 0.893, 0.894, and 0.975, respectively. The results have shown the effectiveness and robustness of the proposed custom model in classifying gliomas into LGG and HGG. The findings suggest that the deep CNNs and transfer learning approaches can be very useful to solve classification problems in the medical domain.

Keywords

• glioma
• clinical scans
• retrospective study
• magnetic resonance imaging
• classification
• deep convolutional neural networks
• transfer learning