Radiomic Strategy For Preoperative Glioma Grading Based On Multi-modal Magnetic Resonance Images

Glioma is the most common type of primary intracranial malignant tumor.The increasingly morbidity and mortality of glioma have threatened human health seriously.The accurate differentiation between low-grade gliomas?LGGs;grade II?and high-grade gliomas?HGGs;grades III,IV?is critical,since the therapeutic strategy and the prognosis could differ substantially depending on the grade.Although some previous research indicate that magnetic resonance images?MRI?should be helpful for glioma grading,due to its conventional anatomical imaging and functional imaging,which may reflect the characteristic of gliomas multi-dimensional,but most of these studies were single modality or univariate analysis researches,and theirs results were controversial.Recently,the term radiomics,which aims to extract abundant quantitative features from medical images and analyze their predictive values for diagnostic decision making,has gradually drawn great attention,and maybe useful to reflect the structural,cell density and angiogenesis characteristics of gliomas,and further improve the accuracy of glioma grading.Purpose: This study aimed to discuss the valuable of Radiomic Strategy combined with Multi-modal MRI for the clinical practice of glioma grading,and furthermore,verified its advantage compared with routine MR diagnostic process.Materials and Methods: In this study,153 patients with postoperatively pathologically confirmed glioma were enrolled according to the inclusion and exclusion criteria.Conventional MR imaging,multi-b diffusion weighted imaging?DWI?and three-dimensional arterial spin labeling?3D-ASL?sequences were performed for each patient.Then,three structure sequences and seven functional parameter maps were obtained as ten single modalities and were classified into three combined modalities: structure combined modalities: T1-Weighted Imaging?T1WI?,T2-Weighted Imaging?T2WI?,Contrast Enhanced T1-Weighted Imaging?T1-CE?;diffusion combined modalities: Apparent Diffusion Coefficient?ADC?,Slow Apparent Diffusion Coefficient(ADC_slow),Alpha,Distributed Diffusion Coefficient?DDC?;perfusion combined modalities: Fast Apparent Diffusion Coefficient?ADCfast?,Perfusion Fraction?Fp?,Cerebral Blood Flow?CBF?.Volume of interest?VOI?was delineated manually on TI-CE images and was copied to the rest nine intensity or parameter maps which had been registered to T1-CE images.3D gray-level co-occurrence and curvature co-occurrence matrix?GLCM and GLGCM?textural features extracted from the VOIs were used to describe the intra-glioma heterogeneity in this study.Meanwhile,the histogram-based features and mean intensity value of VOIs were extracted to compare with previous studies.Then a support vector machine?SVM?based sample augmentation,feature selection and classification strategies were proposed to firstly obtain an optimal feature set and then verify its capacity of grading tasks?LGG vs.HGG and WHO III vs.WHO IV?,comparing with histogram features often used in previous studies.The Receiver Operating Characteristic?ROC?curve analysis was implemented to calculate the Area Under the Curve?AUC?and corresponding 95% confidence interval of grading results.Results: 153 patients with 42 LGG and 111 HGG?33 WHO III and 78 WHO IV?were eligible for this perspective study.Texture feature,histogram feature and mean value were extracted from the VOIs of ten modalities,and constituted the texture matrix with dimension as 420×153,90×153 and 10×153,respectively.To evaluate the performance of texture features in these glioma grading tasks and compare with other two feature groups,the classification models were established using the SVM classifiers trained by these feature groups.After the resampling of synthetic minority oversampling technique?SMOTE?and SVM-based feature selection strategy?SVM-RFE?,the optimal feature subsets for the two tasks were determined?the Top-30 feature set for LGG vs.HGG task,and the Top-28 feature set for Grade III vs.IV task?.SVM-based classification strategy was established.The classification performances of LGG vs.HGG task using texture feature group?accuracy: 96.8%,AUC: 0.987?on overall modalities were more outstanding than using histogram feature group?91.4%,0.984?and mean value group?89.6%,0.961?.Texture feature group also seemed more effective than other types of features on combined modality comparison,the accuracy and AUC of texture feature group on structure,diffusion and perfusion modalities were:?92.8%,0.975?,?93.2%,0.965?and?87.8%,0.931?,respectively.The accuracy and AUC of histogram group were:?91.9%,0.974?,?90.5%,0.971?and?86.9%,0.911?,and for mean value group were:?82.4%,0.889?,?86.0%,0.944?and?79.3%,0.840?,respectively.Moreover,no matter what feature groups were used,the performance of structure and diffusion modality was better than perfusion modality.Texture feature group was still better than other feature groups on single modality comparison.When texture features were applied,the grading performance on T1 CE was the best among the modalities,the accuracy and AUC were 98.1% and 0.992.The descending order of texture classification performance using most single-modalities was: texture,histogram,mean value,but there was an inverse tendency on Alpha map.The best classification performances of III vs.IV task using texture feature group?accuracy: 98.1%,AUC: 0.992?.According to the evaluation of which texture features and corresponding modality or parametric maps contribute most to the two grading tasks,T1-CE and ADC_slow were dominant with the highest weight.In this study,we found that the performance of LGG vs.HGG task using oversampling method?SMOTE?was better than that using undersampling method?T-Link?,and the performance of LGG vs.HGG task after feature selection and parameter optimization was greatly out-performed that before SVM-RFE feature selection method.Along with the inclusion of each ranked feature,the overall variation tendencies of classification accuracy and AUC for both two tasks were soared firstly and declined slowly,which indicated that the precondition of optimal feature subset were adequate number of features and feature selection method.Conclusion: This study presented a radiomic strategy for noninvasively discriminating glioma grades.It effectively incorporated 3D texture features from multi-parametric MRI,and SVM-based classification strategy.Results suggested that texture features were more potential than histogram feature or mean value from VOI in the glioma grading tasks.The proposed radiomic strategy could facilitate the clinical decision making in patients with different grades of glioma.