MRI-based deep learning model for differentiation of hepatic hemangioma and hepatoblastoma in early infancy.
Journal:
European journal of pediatrics
PMID:
37462798
Abstract
UNLABELLED: Hepatic hemangioma (HH) and hepatoblastoma (HBL) are common pediatric liver tumors and present with similar clinical manifestations with limited distinguishing value of serum AFP in early infancy. An accurate differentiation diagnostic tool is warranted for optimizing treatments and improving prognosis. The present study aimed to develop an innovative and cost-effective diagnostic tool to differentiate HH and HBL in early infancy using advanced deep learning (DL) techniques. One hundred forty patients ≤4 months old diagnosed as HH or HBL with histological specimens were recruited from two institutions assigned into a training set with cross-validation and a testing set for external validation, respectively. Based on MRI images, imaging diagnoses were interpreted by two radiologists, and imaging-derived radiomic features were extracted by pretrained convolutional neural networks (CNNs)-Xception extractor via DL analysis. A nomogram model was constructed integrating predictive clinical variables, radiologist-based interpretation, and DL features, evaluated comprehensively on diagnostic and calibration accuracy. The DL-based model performed an area under the receiver operating characteristic curve (AUC) of 0.966 for the training cohort and 0.864 for the testing cohort. The radiologist-interpreted differentiation model showed an AUC of 0.837 in the testing cohort. The integrated nomogram model represented an increasing performance with an AUC of 0.887, accuracy of 78.57%, sensitivity of 76.19%, and specificity of 80.95% in the testing cohort.