Abstract
Fetal MRI provides superior tissue contrast and true 3D spatial information however there is only a limited number of number of MRI studies investigating cervix during pregnancy. Furthermore, there are no clearly formalised protocols or automated methods for MRI cervical measurements. This work introduces the first deep learning pipeline for automated multi-layer segmentation and biometry for 3D T2w images of the pregnant cervix. Evaluation on 20 datasets from 0.55T and 3T acquisitions showed good performance in comparison to manual measurements. This solution could potentially minimise the need for manual editing, significantly reduce analysis time and address inter- and intra-observer bias. Next, we used the pipeline to process 270 normal term cases from 16 to 40 weeks gestational age (GA) range. The inlet diameter and length showed the strongest correlation with GA which is in agreement with the gradual remodeling and softening of the cervix prior to birth. We also generated 3D population-averaged atlases of the cervix that are publicly available online.