171 - Automatic segmentation of umbilical cord to aid fetal MRI safety assessment

Publication Number: 3165.171

Adi Titelman Ashkenazy, Boston Children's Hospital, Watertown, MA, United States; Neel Dey, Massachusetts General Hospital, Cambridge, MA, United States; Esra Abaci Turk, Boston Children's Hospital, Harvard Medical School, Boston, MA, United States; P. Ellen grant, Boston Children's Hospital, Boston, MA, United States; Polina Golland, Massachusetts Institute of Technology, Cambridge, MA, United States; Filiz Yetisir, Harvard Medical School, Boston, MA, United States

Background: Radiofrequency (RF) energy deposition during Magnetic Resonance Imaging (MRI) can cause tissue heating and is regulated for safety. Fetal MRI has a history of safe use, but new technologies that can improve image quality can alter RF energy deposition patterns and thus require safety evaluation. While whole-body heating can be limited by controlling RF power, estimating local tissue heating is challenging and relies on electromagnetic simulations using anatomical body models. As fetal motion can considerably affect tissue heating estimations, body models incorporating fetal motion are needed. This requires segmentation of uterine contents for 100s of fetal positions that sample fetal motion, which can be acquired for example by 4D spatiotemporal echo-planar imaging (EPI) of the uterine volume. However, previous work merges the segmentations for fetal body and umbilical cord (UC) into a single label due to the extreme difficulty in disambiguating these structures in EPI, thereby preventing accurate fetal tissue heating estimations.

Objective: Automatic segmentation of umbilical cord to help generation of detailed anatomical pregnant body models and to enable accurate estimations of fetal tissue heating during MRI.

Design/Methods: This work builds on a previously developed 5-label segmentation network which segments the placenta, amniotic fluid, fetal brain, fetal eyes, and fetal body combined with UC. This network was trained on 176 annotated EPI uterine volumes from 99 subjects. We manually annotated the UC in 36 volumes from 36 subjects (3-10 h per volume) to create new training data for UC segmentation. As 36 volumes are too few for performant training, we finetuned the previously developed 5-label segmentation network. We tested UC segmentation performance (using the Dice score) on four new EPI volumes that were not trained on.

Results: Fig.1 shows an example case from the training data before and after the UC is manually delineated from the fetal body label. In the 4 test cases, our model achieved Dice scores of 0.96, 0.83, 0.91 and 0.72 (Fig.2).

Conclusion(s): We developed a segmentation model to automatically segment the umbilical cord to help the generation of detailed anatomical pregnant body models used for fetal MRI safety assessment. In a preliminary test dataset of 4 cases, our model achieved acceptable Dice scores. Our dataset was limited due to the extremely labor-intensive task of manual labeling of the UC. Future work involves expanding our dataset with more manual segmentations of the umbilical cord in a variety of pregnant subjects.

Figure 1
Whole uterus segmentation as they appear without and with the UC (manually segmented) in an example case from the training data of the UC segmentation model, in 2D (first row) and 3D (second row). Magenta: umbilical cord, cyan: fetal body, dark blue: fetal brain, yellow: fetal eyes, green: amniotic fluid, red: placenta.

Figure 2
Automatic (top row) and manual (bottom row) segmentations of the umbilical cord in 3D for the 4 test cases. Gestational ages are noted at the top and Dice scores are noted at the bottom for each test case. GA: gestational age, w: weeks, d: days.