introduction and purpose

Hypoxic ischemic encephalopathy is caused by a prolonged hypoxic event or diffuse brain blood flow reduction in the perinatal period with high disability and mortality. Structural MR imaging is standard tool to detect the location, distribution and severity^[1]. Because of complicated patterns, different lesions of the injury and inconsistent MR imaging scan efficacy on different scanners, high interobserver variability influenced the diagnosis and therapeutic schedule seriously. We aimed to develop a deep learning network for hypoxic ischemic encephalopathy diagnosis by using structural MR imaging. Objective deep learning for high consistency in diagnosis contributes to precise clinical decision and risk stratification^[2].

Materials and methods

Patients: With the approval of the institutional review board, the study cohort comprised 645 term neonates (median gestational age 38.55+2.59weeks) including 303 patients of hypoxic ischemic encephalopathy and 342 healthy controls. Fulfillment of the criteria for the diagnosis of hypoxic ischemic encephalopathy served as ground truth.
Imaging: on several MRI scanners (Achieva 1.5T, Philips, Amsterdam, Netherlands; Discovery 750, GE Health Care, Waukesha, Wisconsin, USA; United Imaging 780, Shanghai, China;)
Data processing: An end-to-end 3D ResNet34 deep learning architecture served as the diagnosis model, respectively using axial T1-weighted images and T2-weighted images as input to detect the presence of hypoxic ischemic encephalopathy. All images were preprocessed with an MRI automatic processing pipeline. The preprocessing steps included bias correction, intensity normalization, and anisotropic diffusion filtering. First, 3D ResNet34 networks with u-shape framework was trained for brain tumor segmentation in BraTS2017 using T1 and T2 modal, respectively. And the pre-trained 3D ResNet34 networks was fine-tuned in the train set with smaller initial learning rate. Finally, the two trained 3D ResNet34 networks were formulated ensemble model and evaluated in test set.

Results

In the test set, the accuracy in combination of T1WI and T2WI achieved 81.4%, superior to single T1WI or T2WI (79.8%, 76.7%, respectively). Besides, the model reached an AUC of 0.89 for the combination of T1WI and T2WI, higher than single T1WI and T2WI (0.86, p<0.05). The sensitivity and specificity for single T1WI and T2WI was 82.0%, 77.9% and 82.0%, 72.1%, respectively, which was improved in combination of T1WI and T2WI (85.2%, 77.9%)

Discussion and conclusion

The deep learning network achieved high performance in HIE diagnosis for multimodal structural MRI (T1WI+T2WI), which outperformed T1WI or T2WI alone. These findings suggested multimodal MRI provide more features than single modal for diagnosis in neonate HIE, which was extracted by the network effectively. As for single MRI modal, T1WI showed higher diagnosis accuracy than T2WI in the deep learning network. That indicated that the network was more sensitive to hyperintense lesion such as basal ganglia injury in T1WI. The network could detect severe neonate HIE with characteristic appearance such as basal ganglia injury and periventricular leukomalacia, contributing to screening severe HIE for early medical intervention. Early diagnosis using objective methods is quite important for parental decision-making and neonatal management. To sum up, this study validates deep learning based radiomics for diagnosis in neonate HIE, suggesting deep learning based radiomics feature markers can be used to improve interobserver variability and diagnosis accuracy.

Acknowledgements

Funding: This project was supported by National Natural Science Foundation of China (No.81730049）