### 2332

Radiomic analysis to determine glioma’s IDH1 gene status based on multi-MR sequences
Yi Liu1,2, Tong Han2, Xiaoling Yan3, Xuebin Zhang3, and Zhengting Cai4

1graduate school, Tianjin Medical University, Tianjin, China, 2Department of Radiology, Tianjin Huanhu Hospital, Tianjin, China, 3Department of Pathology, Tianjin Huanhu Hospital, Tianjin, China, 4Innovation Department, Huiying Medical Technology Co., Ltd, Beijing, China

### Synopsis

The purpose of this retrospective study was to demonstrate the feasibility of radiomic methods to determine glioma’s IDH1 gene status based on MR imaging. We used a training set (99 patients)with a test set (29 patients), and extracted 1029 radiomic features from each sequence of T2WI, ADC, FLAIR, T1WI-CE and the combined, then reduced by Least absolute shrinkage and selection operator. Five logistic regression classifiers were built based on training set, evaluated using test set and compared by DeLong test. The results indicated the radiomics of combined four sequences had the best performance in distinguishing IDH1 gene status.

### Introduction

IDH1 gene as a molecular marker plays an important role in determining the molecular subtype of glioma, individualizing treatment and estimating prognosis. The current determination of molecular typing (including IDH1 status) mainly depends on pathology. Radiomics transformed MR images into minable and quantitative data, bringing new hope for non-invasive molecular typing. Previous radiomics studies used single MR sequence images with one feature to evaluate glioma classification or molecular subtype[1-3], so the consequences were limited and not good enough. In this study, the use of multi-MR sequences (T2WI, ADC, FLAIR and T1WI-CE) attempted to improve the accuracy of glioma’s IDH1 gene status diagnosis.

### Materials and Methods

Our study retrospectively included 128 patients (male/female: 76/52, average age: 51.28±13.24, IDH1 wild/mutant: 76/52) who were pathologically confirmed as primary glioma in TianJin HuanHu hospital between Jul 2016 and Nov 2017. The MR images of 99 patients (male/female: 62/38, average age: 52.27±13.30, wild/mutant: 62/38) were collected from the Skyra 3T MR scanner (Siemens Healthcare, Erlangen, Germany) with a 20-channel head coil assigned as the training set, while the remaining 29 patients (male/female: 14/14, average age: 47.82±12.14, wild /mutant: 14/14) were from Trio MR scanner (Siemens Healthcare, Erlangen, Germany) with a 20-channel head coil as the test set. The same parameters were set for the same sequence of the two machines. The slice thickness of all axial MR sequences was 5mm, the slice interval was 1.5mm and the field of vision (FOV) was 240×240mm.

To obtain volume of interest (VOI) for further analysis, all the data were uploaded to the Radcloud platform (Huiying Medical Technology Co.Ltd, Beijing, http://radcloud.cn/). The VOI of glioma was the tumor core region manually defined by two experienced radiologists with overall consideration of image features of four sequences, then manually delineated by the junior one in one sequence(eg. T1WI-CE) and mapped into the other three sequences (T2WI, ADC and FLAIR). Figure 1 showed an example of the manual segmentation.

For each sequence, 1029 radiomic features were extracted from VOI, which were divided into four groups: (I) tumor image intensity, (II) shape and size features, (III) texture features (IV) wavelet feature. Next, the features were standardized to avoid the variety of ranges adding excessive weights of different features. In order to reduce the redundancy of features and improve the robustness of results, least absolute shrinkage and selection operator (LASSO) method reduced the count of features into proper number.

Five logistic regression classifiers were constructed based on the training set using the remaining radiomic features of T2WI, ADC, FLAIR, T1WI-CE and the combined, respectively. The diagnostic performance of aforementioned classifiers was evaluated by the test set utilizing the quantitative indices of receiver operating curves (ROC), the area under ROC curves (AUC), accuracy, sensitivity and specificity. The confidence interval of AUC was computed by exact binomial method[4]，and the comparison of different ROCs was conducted by DeLong test[5].

### Result

There were no significant differences in age (p = 0.235, independent-samples t test) or gender (p = 0.253, chi-square test) between the training set and test set. After feature selection by LASSO method, the numbers of remaining features of T2WI, ADC, FLAIR, T1WI-CE and the combined were 24, 23, 31, 24 and 42, respectively (Figure 2A). The diagnostic performances of the five classifiers were shown in Figure 2B and Table 1. As we can see, the combination of four sequences (ALL) had the best prediction performance with the highest AUC (0.962, 95% CI: 0.817-0.999), accuracy (0.897), specificity (0.929). Of single sequence, the T1WI-CE showed better performance with AUC (0.962, 95% CI: 0.817-0.999), accuracy (0.897) while the T2WI showed relatively lower AUC (0.814, 95% CI: 0.627-0.933), accuracy (0.793). According to DeLong test, there were no significant differences among the ROCs of five classifiers (p = 0.063-0.840), except the one between T2WI and ALL (p = 0.031).

### Discussion and Conclusion

In this study we evaluated the IDH1 gene status of glioma using multi-MR sequences based on radiomics, and illuminated that the diagnostic performance of combined four sequences were improved with highest AUC (0.962) compared to each single sequence (AUC = 0.814-0.938). And our study provided further evidence for the association between MR image features and glioma IDH1 gene status.

In summary, we had established classifiers with the good ability to predict IDH1 genotyping preoperatively. These classifiers were expected to assist clinicians to make better clinical diagnosis and treatment strategies, and indicated that radiomics enabled to accelerate the development of personalized medicine.

### Acknowledgements

No acknowledgement found.

### References

[1]Brynolfsson P, Nilsson D, Henriksson R, et al. ADC texture—An imaging biomarker for high-grade glioma?[J]. Medical Physics, 2014, 41(10):101903.

[2] Yu J, Shi Z, Lian Y, et al. Noninvasive IDH1 mutation estimation based on a quantitative radiomics approach for grade II glioma[J]. European Radiology, 2016:1-14.

[3]Ryu Y J, Choi S H, Park S J, et al. Glioma: application of whole-tumor texture analysis of diffusion-weighted imaging for the evaluation of tumor heterogeneity[J]. Plos One, 2014, 9(9):e108335.

[4] Sean Wallis. Binomial Confidence Intervals and Contingency Tests: Mathematical Fundamentals and the Evaluation of Alternative Methods[J]. Journal of Quantitative Linguistics, 2013, 20(3):178-208.Biometrics 1 (September) (1988) 837–845

[5] E.R. DeLong, D.M. DeLong, D.L. Clarke-Pearson, Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach, Biometrics 1 (September) (1988) 837–845.

### Figures

Figure1. An example of the manual segmentation in glioma tumor region

Figure2. A. The histogram of the numbers of remaining features in each sequence after feature selection. B. The ROC curves of five classifiers based on test set.

Note: ALL, All sequences, the combination of four sequences; TC, T1WI-CE; T2, T2WI.

Table 1. The diagnostic performances of five classifiers based on test set

Note: AUC, area under the curve; CI, confidence interval.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
2332