Quantification of Fetal Volume in Magnetic Resonance Images Using Deep Learning

Abstract: Background: Fetal Magnetic Resonance Imaging is an imaging technique that visualizes the fetus during pregnancy. It is used for assessing and diagnosing disease, for monitoring high risk pregnancies, and for conducting research. By segmenting three-dimensional fetal magnetic resonance (MR) images, volume estimation of intrauterine structures is possible. The quantification of fetal body volume and placental volume is of both clinical and scientific relevance. If performed manually, the segmentation is a time-consuming procedure. Purpose: The aim of this project was to create a fully automatic algorithm that segments and quantifies the volumes of the fetus and placenta in MR images. Method: Deep learning, or more specifically a two-dimensional U-Net, was used for solving the task. The data used for training, validation, and testing of the models were manually delineated volumetric MR images of 25 fetuses, which were considered as ground truth. In all images the fetus, placenta, umbilical cord and amniotic fluid were delineated. In addition to segmenting the fetus and placenta, the model was also trained to segment the umbilical cord and amniotic fluid. Result: For the automatic fetal body volume the median absolute volume difference was 2.3 % (std 4.9 percentage points) for five cases that were used for testing, when compared to the ground truth. For the placental volume, the median absolute volume difference was instead 23.9 % (std 34.4 percentage points). Conclusion: In conclusion, the results suggest that a fully automatic method based on the U-Net can be used for quantification of fetal body volume. However, the quantification of placental volume was not satisfactory.