Synopsis

Keywords: Machine Learning/Artificial Intelligence, Brain

Using an automated method of convolutional neural networks (CNNs), we aimed to predict the IDH mutation status of gliomas from conventional preoperative MRI in this study. We conclude that the Markov Random Field-U-Net network can accurately segment the tumor region. Using a modified 34-layer Resnet network, we were subsequently able to predict IDH mutation status effectively. Consequently, the model has the potential to be utilized more broadly as a practical tool with reproducibility, this model has the potential to be a practical tool for the non-invasive characterization of gliomas to help the individualized treatment planning.

Summary of Main Findings

An automated deep learning model was proposed that can automatically predict the IDH mutation status of gliomas using pre-operative conventional MR images.

Synopsis

Using an automated method of convolutional neural networks (CNNs), we aimed to predict the IDH mutation status of gliomas from conventional preoperative MRI in this study. We conclude that the Markov Random Field-U-Net network can accurately segment the tumor region. Using a modified 34-layer Resnet network, we were subsequently able to predict IDH mutation status effectively. Consequently, the model has the potential to be utilized more broadly as a practical tool with reproducibility, this model has the potential to be a practical tool for the non-invasive characterization of gliomas to help the individualized treatment planning.

Introduction

Glioma is the most common malignant primary brain tumor in adults. It is a highly heterogeneous disease with varied molecular subtypes that have resulted in diverse treatments and clinical prognoses ^1,2.It has been reported that IDH mutation status is a crucial factor in glioma tumor behavior. Specifically, the molecular profile and prognosis of low-grade gliomas with wild-type IDH were reported to be comparable to those of glioblastomas. But there is no reliable and non-invasive way to achieve it. Therefore, considerable attention has been dedicated to developing image-based diagnostic methods to determine IDH mutation status.CNNs are a superior technique for extracting high-dimensional numerical information from images by learning relevant features directly from image signal intensities.CNNs offer diagnostic value for glioma IDH status prediction. The purpose of the study was to predict the IDH mutation status of patients with gliomas (grades II-IV) from preoperative MR images using an automated approach that integrates the following: (i) a CNN for automated tumor segmentation and (ii) a CNN-based classifier for IDH status prediction(Figure 1).

Material and Methods

170 patients (105 males,50.6±3.9years，65 males,47.3±1.3years) were retrospectively included in our study. The inclusion criteria were as follows: (i) pathologically confirmed glioma, (ii) known IDH status, (iii) preoperative MRI inclusive and (iv) age ≥18 years. All patients underwent MR exams on a 3.0 T MR scanner (SIGNATM Architect; GE Healthcare, Milwaukee, WI, USA) with a 48-channel head coil. The scan protocol and detailed parameters were listed in Table 1. Our automatic process includes 2 models. Model 1 is a network that combines Markov Random Field (MRF) with U-Net, which was used to segment tumor into edema, hemorrhage and necrosis in tumor（Figure 2).The images were standardized and signal intensity normalized firstly. Then the images were randomly divided into training set, verification set and test set according to the ratio of 6:2:2. The training set images with sizes of 512×512×1 as network parameters of the model. Our CNN classifier for IDH mutation status prediction (model 2) is derived from the well-known 34-layer Resnet architecture (hereinafter referred as the G-Resnet34). Improvements made on this architecture were shown in Figure 1. The model images input comprised tumor masks of 512 × 512 size and axial CE-T1WI, T2WI, T2-FLAIR and T1WI images. The performance of Model 1 for tumor segmentation was measured using the dice similarity coefficient (Dice), PPV (positive predictive value), sensitivity. The diagnostic performance of model 2 was measured in terms of accuracy, area under the receiver operating characteristic curve (AUC) and F1. The 10-fold cross-validation was used to verify the stability of model 2.

Results

All parameters were calculated in PyCharm using Python. Model 1 (CNN for tumor segmentation) yielded dice coefficients of 0.976 averagely. The PPV and sensitivity were 0.947 and 96.8%.. Figure 3 shows the result of segmentation. Our model 2 which predicted the IDH status achieved accuracies of 98.2% (AUC=0.961), 94.5% (AUC=0.909), and 95.6% (AUC=0.915), with F1 of 96.4%, 92.9%, 93.4% in the training set, verification set, and test set, respectively (Figure 4). The average accuracy of 10-fold cross validation was 96.0%.

Discussion and Conclusion

Regardless of the histologic grade, gliomas with an IDH mutation station had a better prognosis and treatment response ^3,4. Our study demonstrated that the automatic approach based on conventional MRI images can accomplish the task from glioma tumor segmentation to the prediction of IDH mutation status. Time-consuming and subjective differences can be avoided because of automatic segmentation of tumor regions. Our model of segmentation derived from U-Nets, and was combined with MRF to extract the hierarchical features of the spatial location of the tumor region to improve the segmentation accuracy further. Besides, the gaussian filter on both sides of the network to refine the tumor area, so as to improve accuracy of the results. We build the prediction model based on Resnet34. Besides, we combined the first and the third layer, the second and the fourth layer of the network. So as to fuse the deep and shallow layers better, which allows obtain more features to improve the accuracy of the model. In addition, a large amount of data is required to deep learning model that ensure the stability and prevent over fitting. One of the major limitations in our study is the relatively small number of lesions. Hence future study with bigger cohort of subject is warranted. To conclude, we can predict the IDH mutation status of gliomas using a fully automatic CNNs based on conventional MR imaging.

Acknowledgements

No acknowledgement found.