Xiaoxue Liu1, Jianrui Li1, Xiang Liao1, Zhongqiang Luo1, Qiang Xu1, Hao Pan2, Qing Zhou3, Yan Tao3, Feng Shi3, Guangming Lu1,4, Zhiqiang Zhang5,6. 1. Department of Diagnostic Radiology, Affiliated Jinling Hospital, Medical School of Nanjing University, 305#, Eastern Zhongshan Rd, Nanjing, 210002, China. 2. Department of Neurosurgery, Affiliated Jinling Hospital, Medical School of Nanjing University, Nanjing, 210002, China. 3. Department of Research and Development, Shanghai United Imaging Intelligence Co., Ltd, Shanghai, 200232, China. 4. State Key Laboratory of Analytical Chemistry for Life Science, Nanjing University, Nanjing, 210093, China. 5. Department of Diagnostic Radiology, Affiliated Jinling Hospital, Medical School of Nanjing University, 305#, Eastern Zhongshan Rd, Nanjing, 210002, China. zhangzq2001@126.com. 6. State Key Laboratory of Analytical Chemistry for Life Science, Nanjing University, Nanjing, 210093, China. zhangzq2001@126.com.
Abstract
PURPOSE: EphA2 is a key factor underlying invasive propensity of gliomas, and is associated with poor prognosis of tumors. We aimed to develop a radiomics-based imaging index for predicting EphA2 expression in diffuse gliomas, and further estimating its value for grading of tumors. METHODS: A total of 182 patients with diffuse gliomas were included. All subjects underwent pre-operative MRI and post-operative pathological diagnosis. EphA2 expression of tumors was scored on pathological sections with immunohistochemical staining using monoclonal EphA2 antibody. MRI radiomics features were extracted from three-dimensional contrast-enhanced T1-weighted imaging and diffusion kurtosis imaging. Predictive models were constructed using machine learning-based radiomics features selection and three classifiers for predicting EphA2 expression and tumor grade. Features of best EphA2 expression model were subsequently used to construct another model of tumor grading. For each model, 146 cases (80%) were randomly picked as training and the rest 36 (20%) were testing cohorts. EphA2 expression was further correlated to the radiomics features in both grade models using Spearman's correlation. RESULTS: Logistic regression model presented highest performance for predicting EphA2 expression (AUC: 0.836/0.724 in training/validation set). Tumor gradings model guided by features from EphA2 expression model demonstrated comparable performance (AUC: 0.930/0.983) to that constructed directly using imaging radiomics features (AUC: 0.960/0.977). Two radiomics features which included in both LR-grade models showed strong correlation (P < 0.05) with EphA2 expression. CONCLUSION: The expression of EphA2 in gliomas could be predicted by radiomics features extracted from diffusion kurtosis MRI, which could also be used to assist tumor grading.
PURPOSE: EphA2 is a key factor underlying invasive propensity of gliomas, and is associated with poor prognosis of tumors. We aimed to develop a radiomics-based imaging index for predicting EphA2 expression in diffuse gliomas, and further estimating its value for grading of tumors. METHODS: A total of 182 patients with diffuse gliomas were included. All subjects underwent pre-operative MRI and post-operative pathological diagnosis. EphA2 expression of tumors was scored on pathological sections with immunohistochemical staining using monoclonal EphA2 antibody. MRI radiomics features were extracted from three-dimensional contrast-enhanced T1-weighted imaging and diffusion kurtosis imaging. Predictive models were constructed using machine learning-based radiomics features selection and three classifiers for predicting EphA2 expression and tumor grade. Features of best EphA2 expression model were subsequently used to construct another model of tumor grading. For each model, 146 cases (80%) were randomly picked as training and the rest 36 (20%) were testing cohorts. EphA2 expression was further correlated to the radiomics features in both grade models using Spearman's correlation. RESULTS: Logistic regression model presented highest performance for predicting EphA2 expression (AUC: 0.836/0.724 in training/validation set). Tumor gradings model guided by features from EphA2 expression model demonstrated comparable performance (AUC: 0.930/0.983) to that constructed directly using imaging radiomics features (AUC: 0.960/0.977). Two radiomics features which included in both LR-grade models showed strong correlation (P < 0.05) with EphA2 expression. CONCLUSION: The expression of EphA2 in gliomas could be predicted by radiomics features extracted from diffusion kurtosis MRI, which could also be used to assist tumor grading.