The Effect of Deep Learning Segmentation on Placental T2* Estimation

The placenta plays a crucial role in ensuring adequate oxygen delivery to the fetus. Despite being one of the most important organs in pregnancy, the placenta remains poorly understood. Research into placental developmental structure and function has accelerated in recent years. The use of a preclinical pregnant nonhuman primate model allows researchers to quantify blood and oxygen transfer through the placenta and potentially improve our understanding of it. Furthermore, deploying Magnetic Resonance Imaging (MRI) to better understand the explicit function and mechanism of the placental vascular supply has become commonplace. Among these MRI techniques, T2, T2* and diffusivity are the most commonly deployed for accessing the pseudo-quantitative physical properties of the placenta. To enable better analysis of the placenta from MR images, and reduce the labour-intensive workload of manually delineating regions of interest, here we deployed a robust deep learning model trained on T2* MR images acquired from pregnant rhesus macaques, with the ground truth placenta segmentation drawn by our experts. After obtaining the trained model, we experimented with three segmentation fusion strategies aimed at merging the estimated segmentations from multi-echo times into one final outcome. For each fusion method, we conducted experiments to evaluate the effect of the automated segmentation on T2* estimation, including varying the echo time and the number of imaging slices. The best Dice score (averaging fusion method over 6 segmentations from different TE in one scan) reached 0.869. This work highlights the robustness of our trained segmentation model on nonhuman primate datasets and its potential for application in future preclinical studies. With appropriate fine-tuning, it could also be adapted for analysing human placental datasets.