Radiomics for predicting hematoma expansion in patients with hypertensive intraparenchymal hematomas. Issue 115 (June 2019)

• Record Type: Journal Article
• Authors: Ma, Chao; Zhang, Yupeng; Niyazi, Tuerdialimu; Wei, Jian; Guocai, Guo; Liu, Jianan; Liang, Shikai; Liang, Fei; Yan, Peng; Wang, Kun; Jiang, Chuhan
• Is Part Of: European journal of radiology. Issue 115(2019)
• Page Start: 10
• Page End: 15
• Publication Date: 2019-06
• Subjects: NECT non-contrast enhanced CT -- LASSO least absolute shrinkage and selection operator -- HE hematoma expansion -- NHE no hematoma expansion -- CTA CT angiography -- ICH intracerebral hemorrhage -- AUC area under the curve -- ICC intraclass correlation coefficient -- ROC receiver operating characteristics curve -- LoG laplacian of gaussian -- PPV positive predictive value -- NPV negative predictive value -- ROI region of interest -- GLCM gray-level co-occurrence matrix -- GLRLM gray-level run length matrix; Radiomics -- Computed tomography -- Hematoma expansion -- Hypertension -- Machine learning; Medical radiology -- Periodicals; Radiology -- Periodicals; Radiologie médicale -- Périodiques; Medical radiology; Periodicals; 616.075705
• DOI: 10.1016/j.ejrad.2019.04.001
• Languages: English
• ISSNs: 0720-048X

Abstract

Abstract: Purpose: To explore the feasibility of predicting hematoma expansion at acute phase via a radiomics approach. Methods: 254 cases with hypertensive intraparenchymal hematomas were retrospectively reviewed. Baseline non-contrast enhanced CT scan (NECT) were obtained on admission and compared to follow up CT to confirm the occurrence of hematoma expansion. Cases were split into training dataset with 149 cases and a test dataset with 105 cases. Radiomics features were extracted and informative features were selected by least absolute shrinkage and selection operator (LASSO) with 3-fold-cross validation. A radiomics score was then constructed with the selected features to discriminate enlarged hematomas from those that remained stable. Discriminative performance of the score was evaluated on the training and test dataset with area under the curve (AUC) and confusion matrix related metrics. Results: A total of 576 radiomics features were extracted from 6 feature groups on NECT, of which 484 were stable. 5 features were selected by LASSO and based on which a radiomics score were constructed. The radiomics score achieved high discrimination ability between hematoma expansion and no-expansion with AUC of 0.892 (95% CI: 0.824–0.959) and accuracy of 0.852 in the training dataset. In the test dataset, predicting sensitivity, specificity, PPV, NPV and accuracy were 0.808, 0.835, 0.618, 0.930 and 0.820, respectively. Conclusions: Radiomics features were effective in the prediction of hematoma expansion for patients with hypertensive intraparenchymal hematomas. Our radiomics score may provide a fast and quantitative risk assessment for these patients.