Deep Regression Model for Total Brain Volume Estimation in Preterm Neonates

Abstract

This study investigates the use of Deep Learning (DL) models for estimating total brain volume (TBV) in preterm infants utilizing 3D ultrasound (US) imaging data. A custom model with 3D convolution kernels and residual connections is developed, trained on 370 3D raster images from a cohort of 51 infants of postnatal age below 98 days (Mean age: 23.31 days, Standard Deviation: 21.03 days). The impact of incorporating postnatal age, either as an input parameter or a predicted variable, is briefly analyzed. Model uncertainty representations based on dropout are explored, leading to the development of a novel method centered on input transformations. The practical application of this research culminates in the creation of a graphical user interface that can load 3D US images, estimate TBV within seconds, and autonomously reject images deviating significantly from the model’s training distribution. This work contributes to the understanding of DL-based TBV estimation in preterm infants, highlighting the potential benefits of incorporating uncertainty measures and providing a user-friendly tool for medical practitioners.