Evaluation of semiautomated quantification of cranial ultrasound images in newborns as a predictor of Neonatal Behavioral Assessment Scale

Diagnosis of white matter damage by neonatal cranial ultrasound (CrUS) is subject to inter-observer variability and has a low sensitivity to detect late abnormal neurodevelopment in life. In the last decades there have been a significant effort reporting that statistical features of ultrasound images carry important information associated with changes of tissue microstructure. In this work we explored the ability of a semi-automated image processing method to associate ultrasound texture patterns with Neonatal Behavioral Assessment Scale (NBAS) performance in premature neonates. A total of ninety infants born at a median gestational age of 29 weeks were included. The infants underwent one CrUS scan performed at the same day that NBAS test. In this work, we developed a feature selection algorithm to identify combination of features that correlated to NBAS clusters. Our algorithm was then able to predict individual underscored NBAS clusters with accuracy higher than 80% in a “blind” sample.