Selecting and assessing quantitative early ultrasound texture measures for their association with cerebral palsy.

Abstract

Cerebral palsy (CP) develops as a consequence of white matter damage (WMD) in approximately one out of every 10 very preterm infants. Ultrasound (US) is widely used to screen for a variety of brain injuries in this patient population, but early US often fails to detect WMD. We hypothesized that quantitative texture measures on US images obtained within one week of birth are associated with the subsequent development of CP. In this retrospective study, using images from a variety of US machines, we extracted unique texture measures by means of adaptive processing and high resolution feature enhancement. We did not standardize the images, but used patients as their own controls. We did not remove speckle, as it may contain information. To test our hypothesis, we used the "random forest" algorithm to create a model. The random forest classifier achieved a 72% match to the health outcome of the patients (CP versus no CP), whereas designating all patients as having CP would have resulted in 53% error. This suggests that quantitative early texture measures contain diagnostic information relevant to the development of CP.