Detection of Injury and Automated Triage of Preterm Neonatal MRI Using Patch-Based Gaussian Processes

Abstract

Automatic detection or highlighting of neonatal brain injury could be a valuable adjunct to radiological interpretation. Here we propose a normative modeling-based detection method for preterm neonatal neuroimaging using gaussian processes (GPs). These GPs incorporates local image intensity information from image patches and demographics such as age. Z-score images can then be created from the scaled difference between the model predictions and a neonate’s T1 and T2 weighted MRI. To test the use of these GP Z-scores as a form of automated triage, we trained a logistic regression classifier to separate normal and abnormal images. We used 133 preterm neonatal images with normal-reported MRI to train a GP model and optimized lesion detection performance on 36 preterm neonatal images with manually annotated lesion masks. The automated triage model was trained on 100 preterm neonates with normal reported MRI and 109 preterm neonates with MRI detectable lesions. It was tested on the same 36 manually annotated abnormal MRI preterm neonates and 33 normal-reported preterm neonates. Using a patch diameter of 7 voxels and integrating both T1w and T2w Z-score images provided our highest performing GP model for within image lesion detection, achieving an AUC of 0.961. By combining the output probabilities of a T1w and a T2w Z-score histogram classifiers allows for the correctly identification of 32/36 abnormal and 28/33 normal images. These results indicate patch-based normative model can accurately detect lesions in a highly interpretable fashion in preterm neonates with abnormal MRI. Using outputs from these predictions, the classifier is effective at separating abnormal and normal images.