Fetal Brain Segmentation Using Deep Learning

Deep learning is no longer groundbreaking in the field of medical imaging. As outstanding developers have publicly released their source codes in public, it has been widely exploited not only for segmentation of tumor and brain injuries but also for clinical diagnostic services assisted by artificial intelligence.

Pediatric radiologists have also kept their eyes on the applicability of deep learning technologies. In in utero magnetic resonance imaging (MRI) of pregnant women, brain segmentation is an important step to accurately assess the volumetric growth of the fetal brain, but technically challenging due to fetal and maternal motion and other physiological artifacts. Excellent scientists in Harvard Medical School began to demonstrate the feasibility of automatic segmentation of fetal brain using convolutional neural network (CNN) which is one of major deep learning technologies for image analyses [1].

Since last year, I have initiated a new project of building the automatized pipeline of segmenting fetal brain and placenta in T2-weighted MR images (which are used to visualize the anatomical structure of organs). I took the lead to organize a deep learning team, and have studied the state-of-the-arts theories and applications of deep learning. As the first outcome of this wonderful teamwork, we could first report our preliminary progress on fetal brain segmentation using CNN in October 2018 at the Workshop on Machine Learning hosted by the International Society for Magnetic Resonance in Medicine (ISMRM) [2].

We used an open source deep learning package called DeepMedic, developed by the biomedical Image Analysis Group in Imperial College London, in which neural networks are composed of multiple convolutional pathways of different image scales. We trained the neural network using the ground truth data of manual segmentation we've already created. Although we used a small number of training samples, the performance of predicting the brain region from test samples was quite acceptable compared to the ground truth.

This study demonstrates the applicability of deep learning for fully automatic segmentation of the fetal brain in anatomical MR images. This work offers an important technical advance for developing future non-invasive, and early MRI biomarkers for abnormal in vivo fetal brain growth disturbances in high-risk pregnancies. It is just the first step to move forward to our dream. We dream to build a fully automatized pipeline of MRI data processing to quickly provide clinical information of imaging data analysis to clinicians and patients.

References

1. Mohseni Salehi SS, Erdogmus D, Gholipour A. Auto-Context Convolutional Neural Network (Auto-Net) for Brain Extraction in Magnetic Resonance Imaging. IEEE Trans. Med. Imaging 2017;36:2319–2330.
2. Wonsang You, Kushal Kapse, Yao Wu, Dhineshvikram Krishnamurthy, Catherine Limperopoulos, "Automatic segmentation of the fetal brain using multi-scale 3D convolutional neural network: a pilot study," ISMRM Workshop on Machine Learning, Part II, Washington, DC, October 2018.