Deep Learning Radiology for Fetal Development Assessment in MRI

Human fetal development is a complex process where the fetus changes in size, structure, and maturation occur in a unique spatio-temporal pattern. The early detection of developmental disorders relies on the evaluation of ultrasound (US) images and on a few manual measurements performed on them. Magnetic Resonance Imaging (MRI) is increasingly used to obtain additional and more accurate biomarkers when US results are inconclusive. However, manual acquisition of MRI-based biomarkers is time consuming, requires expertise and is prone to annotator variability.

In this talk, we present a novel annotation methodology and deep learning methods for the automatic computation of standard and novel biometric measures in fetal MRI. Specifically, we describe: 1) a bootstrapping approach for optimizing the manual annotation and correction of fetal structures and for training deep learning models with very few annotated datasets; 2) pipelines for computing standard linear and volumetric biometric measures on MRI scans; 3) new ocular and fetal fat MRI-based biometric markers and methods to compute them. We demonstrate the clinical potential of our methods on fetal whole-body weight estimation and placenta and fetal brain structures segmentation.