Table_1_Machine Learning Assisted MRI Characterization for Diagnosis of Neonatal Acute Bilirubin Encephalopathy.DOCX (274.21 kB)

Table_1_Machine Learning Assisted MRI Characterization for Diagnosis of Neonatal Acute Bilirubin Encephalopathy.DOCX

Download (274.21 kB)
posted on 01.10.2019 by Zhou Liu, Bing Ji, Yuzhong Zhang, Ge Cui, Lijian Liu, Shuai Man, Ling Ding, Xiaofeng Yang, Hui Mao, Liya Wang

Background: The use of magnetic resonance imaging (MRI) in diagnosis of neonatal acute bilirubin encephalopathy (ABE) in newborns has been limited by its difficulty in differentiating confounding image contrast changes associated with normal myelination. This study aims to demonstrate the feasibility of building a machine learning prediction model based on radiomics features derived from MRI to better characterize and distinguish ABE from normal myelination.

Methods: In this retrospective study, we included 32 neonates with clinically confirmed ABE and 29 age-matched controls with normal myelination. Radiomics features were extracted from the manually segmented region of interest (ROI) on T1-weighted spin echo images, followed by the feature selection using two-sample independent t-test, least absolute shrinkage and selection operator (Lasso) regression, and Pearson's correlation matrix. Additional feature quantifying the relative mean intensity of ROI was defined and calculated. A prediction model based on the selected features was built to classify ABE and normal myelination using multiple machine learning classifiers and a leave-one-out cross-validation scheme. Receiver operating characteristics (ROC) analysis was used to evaluate the prediction performance with the area under the curve (AUC) and feature importance ranked based on the Fisher score.

Results: Among 1319 radiomics features, one radiologist-defined intensity-based feature and 12 texture features were selected as the most discriminative features. Based on these features, decision trees had the best classification performance with the largest AUC of 0.946, followed by support vector machine (SVM), tree-bagger, logistic regression, Naïve Bayes, discriminant analysis, and k-nearest neighborhood (KNN), which have an AUC of 0.931, 0.925, 0.905, 0.891, 0.883, and 0.817, respectively. The relative mean intensity outperformed other 12 texture features in differentiating ABE from controls.

Conclusions: The results from this study demonstrated a new strategy of characterizing ABE-induced intensity and morphological changes in MRI, which are difficult to be recognized, interpreted, or quantified by the routine experience and visual-based reading strategy. With more quantitative and objective measurements, the reported machine learning assisted radiomics features-based approach can improve the diagnosis and support clinical decision-making.