Differentiation of Treatment-Related Effects from Glioma Recurrence Using Machine Learning Classifiers Based Upon Pre-and Post-Contrast T1WI and T2 FLAIR Subtraction Features: A Two-Center Study

Abstract

Xin-Yi Gao,1– 3,* Yi-Da Wang,4,* Shi-Man Wu,5,* Wen-Ting Rui,5 De-Ning Ma,1 Yi Duan,4 An-Ni Zhang,1– 3 Zhen-Wei Yao,5 Guang Yang,4 Yan-Ping Yu1– 3 1Institute of Cancer and Basic Medicine (ICBM), Chinese Academy of Sciences, Hangzhou, People’s Republic of China; 2Department of Radiology, Cancer Hospital of the University of Chinese Academy of Sciences, Hangzhou, People’s Republic of China; 3Department of Radiology, Zhejiang Cancer Hospital, Hangzhou, People’s Republic of China; 4Department of Physics, Shanghai Key Laboratory of Magnetic Resonance, Shanghai, People’s Republic of China; 5Department of Radiology, Huashan Hospital, Fudan University, Shanghai, People’s Republic of China*These authors contributed equally to this workCorrespondence: Yan-Ping YuDepartment of Radiology, Cancer Hospital of the University of Chinese Academy of Sciences, East Banshan Road, 1#, Hangzhou City, Zhejiang Province, People’s Republic of ChinaTel/Fax [phone]Email [email protected] YangDepartment of Physics, Shanghai Key Laboratory of Magnetic Resonance, North Zhongshan Road, 3663#, Shanghai, People’s Republic of ChinaTel/Fax [phone]Email [email protected]: We propose three support vector machine (SVM) classifiers, using pre-and post-contrast T2 fluid-attenuated inversion recovery (FLAIR) subtraction and/or pre-and post-contrast T1WI subtraction, to differentiate treatment-related effects (TRE) from glioma recurrence.Materials and Methods: Fifty-six postoperative high-grade glioma patients with suspicious progression after radiotherapy and chemotherapy from two centers were studied. Pre-and post-contrast T1WI and T2 FLAIR were collected. Each pre-contrast image was voxel-wise subtracted from the co-registered post-contrast image. Dataset was randomly split into training, and testing on a 7:3 ratio, accordingly subjected to a five fold cross validation. Best feature subsets were selected by Pearson correlation coefficient and recursive feature elimination, whereupon a radiomics classifier was built with SVM. The discriminating performance was assessed with the area under receiver-operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV).Results: In all, 186 features were extracted on each subtraction map. Top nine T1WI subtraction features, top thirteen T2 FLAIR subtraction features and top thirteen combination features were selected to build optimal SVM classifiers accordingly. The accuracies/AUCs/sensitivity/specificity/PPV/NPV of SVM based on sole T1WI subtraction were 80.00%/80.00% (CI: 0.5370– 1.0000)/100%/70.00%/62.50%/100%. Those results of SVM based on sole T2 FLAIR subtraction were 86.67%/84.00% (CI: 0.5962– 1.0000)/100%/80%/71.43%/100%. Those results of SVM based on both T1WI subtraction and T2 FLAIR subtraction were 93.33%/94.00% (CI: 0.7778– 1.0000)/100%/90%/83.33%/100%, respectively.Conclusion: Pre- and post-contrast T2 FLAIR subtraction provided added value for diagnosis between recurrence and TRE. SVM based on a combination of T1WI and T2 FLAIR subtraction maps was superior to the sole use of T1WI or T2 FLAIR for differentiating TRE from recurrence. The SVM classifier based on combination of pre-and post-contrast subtraction T2 FLAIR and T1WI imaging allowed for the accurate differential diagnosis of TRE from recurrence, which is of paramount importance for treatment management of postoperative glioma patients after radiation therapy.Keywords: glioma recurrence, treatment-related effects, T2 FLAIR enhancement, image subtraction, support vector machines

Keywords

• glioma recurrence
• treatment related effects
• t2 flair enhancement
• image subtraction
• support vector machines