Automatic fetal brain extraction from 2D in utero fetal MRI slices using deep neural network
Publication in refereed journal


Times Cited
Altmetrics Information
.

Other information
AbstractBackground
In utero fetal MRI has been developing in common medical prenatal practice for nearly two decades. But the applications and research on fetal MRI still lag behind due to the lack of specialized image processing and analysis tools. Brain extraction, as an initial preprocessing step for many brain MRI-based processing methods, is an important basis for accurate fetal MRI analysis. However, it is very challenging to automatically extract fetal brains from fetal MRI due to the large variation in fetal brains across different gestational weeks and complex maternal tissues surrounding the fetal brains.

Method
We proposed a novel two-step framework using the deep learning method for solving the challenging problem of automatic fetal brain extraction in 2D in utero fetal MRI slices. The proposed framework consisted of two fully convolutional network (FCN) models, i.e., a shallow FCN and an extra deep multi-scale FCN (M-FCN). The first shallow FCN rapidly located the fetal brain and extracted the region of interest (ROI) containing the brain. Then, within the brain ROI, the M-FCN further refined the segmentation and produced the final brain mask by leveraging the multi-scale information and residual learning blocks. Dilated convolutional layers were employed in both FCNs to control the size of feature maps and increase the field of view.

Result
Eighty-eight 2D fetal MRIs were collected for experiments. We compared our method with the state-of-the-art methods on extracting fetal brains. It has been evaluated that our proposed framework outperformed the other methods in both fetal brain localization and segmentation tasks. With the proposed method, we located the fetal brain with an accuracy of 100%. The brain segmentation performance was measured based on the overlap between the automatic segmentations and the manual segmentations. Our proposed method achieved an average of 0.958 Dice score, 0.950 sensitivity rate, and 0.968 precision on the testing dataset, and it took an average of 6 s to process one fetal MRI stack on a workstation with TITAN X GPU and i7-6700 CPU.

Conclusion
In this paper, we proposed an effective and efficient deep learning framework for automatic fetal brain extraction from fetal MRI. It has been validated with solid experiments that the proposed method can be used as a practical and useful tool in clinical practice and neuroscience research.
Acceptance Date02/02/2020
All Author(s) ListJinpeng Li, Yishan Luo, Lin Shi, Xin Zhang, Ming Li, Bing Zhang, Defeng Wang
Journal nameNeurocomputing
Year2020
Month2
Volume Number378
PublisherElsevier
Pages335 - 349
ISSN0925-2312
eISSN1872-8286
LanguagesEnglish-United States

Last updated on 2020-26-10 at 00:11