Semi-automated process for fetal functional brain fMRI isolation, motion correction and censoring
Poster No:
1954
Submission Type:
Abstract Submission
Authors:
Amyn Majbri1, Lanxin Ji2, Ellyn Kennelly3, Cassandra Hendrix4, Mark Duffy2, Iris Menu2, Tanya Bhatia5, Moriah Thomason6
Institutions:
1New York University Grossman School of Medicine, New York, NY, 2NYU Langone Health, New York, NY, 3Wayne State University, Detroit, MI, 4NYU Langone, New York, NY, 5New York University Medical Center, New York, NY, 6NYU Langone Medical Center, New York, NY
First Author:
Co-Author(s):
Introduction:
Fetal brain functional MRI is a non-invasive technique that measures correlations of BOLD signals between different brain regions in the developing fetus, providing a unique window into the organization and maturation of the fetal brain networks4.Although well developed in the adult brain, image processing in fetal brains poses more challenges than adult brains due to fetal motion, artifacts, and reduced signal-to-noise. Our research group previously published a manual processing pipeline for fetal brain fMRI, but it is labor intensive and can discards large quantities of fMRI data2This study aims to address these challenges by adapting a novel semi-automated process for fetal functional brain masking, motion correction and censoring.
Methods:
Data from 49 fetuses were acquired using a Siemens Verio 70-cm open-bore 3T MR system with a 550 g abdominal 4-Channel Siemens Flex Coil. The data had the following scanning parameters: TR/TE = 2000/18 ms; resolution = 3.4 × 3.4 × 4 mm3), flip angle = 80º.
A single mask was manually drawn onto a reference frame within a section of low motion for each acquired run. For every volume in the timeseries, a convolutional neural network-trained model automatically segmented the brain from the maternal compartment, generating a rough 4D mask for the entire timeseries3. The brain was then extracted using this rough mask for motion estimation using FSL mcflirt1, where transformation matrices for mapping each volume to the reference frame were generated. The rough mask is necessary because motion estimation requires a clear background without maternal tissue. We then applied the inverted transformation matrices to the manually drawn mask to generate a more closely defined 4D mask. The raw data was masked again using the improved 4D manual-based mask. We repeated the steps above one time resulting in more refined transformation matrices and a smartly-generated semi-automated mask for the 4D timeseries. Once the final 4D mask is obtained, we apply it to the raw fMRI data for a precise brain extraction and conduct the motion correction on the masked data. See a flowchart in Figure 1.
A three-step approach to censoring was adopted. First, frames that not aligned with the reference were identified via Sørensen–Dice coefficient (DC) calculation. Five volumes around the reference volume were averaged and binarized to produce an ideal image-shape mask. The DC between each volume across the whole time series and this mask was calculated. Volumes with a DC below 0.9 were censored. Next, any volume above two standard deviations above the mean framewise displacement (1.5) across all inputted runs was censored. Finally, any volumes with DVARS values greater than two standard deviations above the mean (132) across all inputted runs was censored. The same dataset was pre-processed in parallel with our previously published pipeline for comparison. Retained volumes were calculated for both pipelines and compared.
A single mask was manually drawn onto a reference frame within a section of low motion for each acquired run. For every volume in the timeseries, a convolutional neural network-trained model automatically segmented the brain from the maternal compartment, generating a rough 4D mask for the entire timeseries3. The brain was then extracted using this rough mask for motion estimation using FSL mcflirt1, where transformation matrices for mapping each volume to the reference frame were generated. The rough mask is necessary because motion estimation requires a clear background without maternal tissue. We then applied the inverted transformation matrices to the manually drawn mask to generate a more closely defined 4D mask. The raw data was masked again using the improved 4D manual-based mask. We repeated the steps above one time resulting in more refined transformation matrices and a smartly-generated semi-automated mask for the 4D timeseries. Once the final 4D mask is obtained, we apply it to the raw fMRI data for a precise brain extraction and conduct the motion correction on the masked data. See a flowchart in Figure 1.
A three-step approach to censoring was adopted. First, frames that not aligned with the reference were identified via Sørensen–Dice coefficient (DC) calculation. Five volumes around the reference volume were averaged and binarized to produce an ideal image-shape mask. The DC between each volume across the whole time series and this mask was calculated. Volumes with a DC below 0.9 were censored. Next, any volume above two standard deviations above the mean framewise displacement (1.5) across all inputted runs was censored. Finally, any volumes with DVARS values greater than two standard deviations above the mean (132) across all inputted runs was censored. The same dataset was pre-processed in parallel with our previously published pipeline for comparison. Retained volumes were calculated for both pipelines and compared.
·Figure 1: Proposed semi-automated process for fetal functional brain fMRI isolation, motion correction and censoring flowchart.
Results:
For every run in the dataset, the new approach resulted in an increased number of retrained volumes, while maintaining a low FD. The median number of volumes retained using the legacy method was 157.0 volumes; IQR = 67.0, whereas the novel method proposed here yields a higher median of 264 volumes; IQR = 119.0 volumes (Figure 2). The disparity between the two methods is notable (p < 0.01), revealing the new approach increases the median volumes retained, highlighting the effectiveness of the new method in achieving a more consistent and controlled retention of volumes compared to the older approach.
·Figure 2: Median number of volumes retrained using novel and legacy processing approaches. The Wilcoxon signed-rank test was employed to assess the significance of differences between the median volum
Conclusions:
This semi-automated pipeline for fetal fMRI processing improves on previous methods by retaining more volumes. This pre-processing pipeline is also more efficient because it requires a single manual mask per run as opposed to the previous method, which required multiple manually drawn masks per run.
Modeling and Analysis Methods:
Methods Development 2
Motion Correction and Preprocessing 1
Neuroinformatics and Data Sharing:
Workflows
Novel Imaging Acquisition Methods:
BOLD fMRI
Keywords:
Computational Neuroscience
Computing
Data analysis
FUNCTIONAL MRI
MRI
Workflows
1|2Indicates the priority used for review
Provide references using author date format
Ji (2022), ‘Empirical evaluation of human fetal fMRI preprocessing steps’, Network Neuroscience, vol. 6, no. 3, pp. 702-721.
Rutherford (2022). Automated Brain Masking of Fetal Functional MRI with Open Data. Neuroinformatics. vol. 20 no.1, pp.173-185.
Thomason (2021), Interactive relations between maternal prenatal stress, fetal brain connectivity, and gestational age at delivery. Neuropsychopharmacol. vol. 46, pp. 1839–1847.