State-of-The-Art Placenta Segmentation in Fetal MRI: A Bootstrap Self-Training Method for Sequence Transfer
2Sagol Brain Institute, Tel Aviv Sourasky Medical Center, Israel
3Sackler Faculty of Medicine and Sagol School of Neuroscience, Tel Aviv University, Israel
4Division of Pediatric Radiology, Tel Aviv Sourasky Medical Center, Israel
5University of Ottawa, Childrens Hopsital of Eastern Ontario, Canada
Background: Quantitative volumetric evaluation of the placenta in fetal MRI scans is an important component of the fetal health evaluation. However, manual segmentation of the placenta is a time-consuming task that requires expertise and suffers from high observer variability. Deep learning methods for automatic segmentation are effective but require manually annotated datasets for each scanning sequence.
Methods: We have developed a new method for bootstrapping automatic placenta segmentation by deep learning on different MRI sequences. The method consists of automatic placenta segmentation with two networks trained on labeled cases of one sequence followed by automatic adaptation using self-training of the same network to a new sequence with new unlabeled cases of this sequence. It uses a novel combined contour and soft Dice loss function for both the placenta ROI detection and segmentation networks. The method was evaluated on a dataset of retrospective MRI placenta scans of patients acquired with FIESTA (40 cases) and TRUFI sequences (15 cases) as part of routine fetal assessment from the Sourasky Medical Center with gestational ages of 28-39 weeks. Ground truth placenta annotations were obtained from scratch by manual annotation and validated by two expert radiologists. FIESTA and FIESTA+TRUFI networks were trained with the annotations.
Results: Our experimental studies for the FIESTA sequence yields a Dice score of 0.847 on 21 test cases with only 16 cases in the training set. Transfer to the TRUFI sequence yields a Dice score of 0.78 on 15 test cases, a significant improvement over the network results without transfer learning.
Conclusion: Our method achieves state-of-the art placenta segmentation results by sequence transfer bootstrapping. Its key advantage is that it streamlines the annotation for the new sequence since annotating data from initial network results requires less time than annotation from scratch ( 60 mins).