By Emily Leclerc | Waisman Science Writer
Movement is the arch nemesis of MRI (magnetic resonance imaging) scans. This type of imaging has become an important tool for studying the brain and disorders related to the brain. Motion during a scan creates blurry images and greatly impacts the information that can be gleaned. Individuals must remain as still as possible while getting a scan to ensure the best results. This can be particularly challenging when scanning infants and young children where movement is almost inevitable. To help account for motion-created artifacts during an MRI scan, Doug Dean III, PhD, assistant professor of pediatrics and medical physics and Waisman investigator, and Jayse Merle Weaver, a graduate student in Dean’s lab, created a neural network, a type of machine learning, to accurately identify and remove motion-corrupted images from infant diffusion MRI data sets. This tool will help improve the quality of data that can be gathered from diffusion MRI scans and subsequently improve research.
The recently released paper showcasing the neural network, ‘Automated motion artifact detection in early pediatric diffusion MRI using a convolutional neural network’, was published in the journal Imaging Neuroscience.
In general, movement is detrimental to medical imaging. It affects image quality as well as the data pulled from those images. With diffusion MRI, which is a specific type of MRI scanning that is sensitive to the motion of water in different tissues, motion can create issues with image quality and affect the information that can be measured from these data. Diffusion MRI involves taking a series of repeated images over a span of time and when processing the images, they have to be aligned in order to gather information from them. “We have techniques to correct for this but they don’t always work super well,” Weaver says. “Especially in the case of infants and young pediatric subjects, the movement tends to be a lot greater so these automated correction techniques don’t really work at all.”
Up to this point to counteract this, an important step in processing MRI scans and data is to manually examine the quality of the images. This involves people going through every image and removing the ones with artifacts or those corrupted by movement. That is currently the gold standard for cleaning data sets. According to Weaver, removal of the corrupted data and images is a crucial step because it prevents results from being skewed due to bad data. But human-based quality control comes with its own issues.
“We have a team of folks that go through our data and identify which images are bad. But imagine you are scrolling through hundreds of images, that’s going to take time,” Dean says. “Time is a big factor. Also, if you have multiple people doing this, they may have different criteria for what they believe is good and what is bad. So, the process is somewhat subjective and based on the reviewer.” People are also going to make mistakes that an artificial intelligence tool would not.
Now, the new neural network program developed by Dean and Weaver overcomes the shortcomings of manual quality control. The neural network can complete the entire process in a matter of seconds and does so based on objective criteria. AI is less prone to making mistakes as it does not suffer from reader fatigue the way a person would be. Dean and Weaver trained the network by feeding it data sets that their team had gone through and tagged which images were good and which were bad. This allowed the network to learn what constituted a good image and a bad one. That way they could give the network a baseline to work from. Once the network was fully trained, they found that it was processing the data sets with accuracies in the upper 90%.
“We got above 95% accuracy on both of the data sets we trained the network on, which is pretty good,” Weaver says. In fact, the neural network was able to catch mistakes that the people had made in tagging the data. “I was looking at the results and was like why are there so many false positives, where the network would tag an image with motion when there wasn’t any? So, I went into the data and looked and for most of them there actually was motion. Even though it was trained on these manual labeling mistakes, it was able to see past them and realize that the images were actually motion corrupted.”
Dean and his lab focus on studying the microstructure of the brain as it develops from infancy into childhood and how changes in the structure are related to behaviors and conditions. Accurate data sets are crucial to his overall work as well as other who work with imaging. Having more accurate quality control measures helps to create better data sets. Better and more accurate data sets then allow for more accurate results from research studies. This would then help researchers improve their understanding of intellectual and developmental disabilities (IDDs) as they continue to study the brain. Then hopefully, this would lead to improved quality of life for individuals with IDDs.
“If you don’t do anything about motion artifacts, the errors in our metrics can increase. And when you are doing comparison studies, say comparing children with IDDs versus children without IDDs, if we observe statistical results that may or may not show significant differences, you have to ask yourself if the findings are real or if the findings may be influenced by artifacts that didn’t get corrected,” Weaver says. That can be even more of a problem with infants or individuals with IDDs who can be less likely to stay still in the scanner.
While the neural network performed well it is still rather limited in its scope as it was specifically trained using data from infant scans. “We should begin thinking about training it on larger data sets to help improve it and more generalize it. Expanding the range of the model, so that it covers a broad scope of data, will help make it more useful for a wide spread of studies here at the Waisman Center and elsewhere,” Dean says. “Then we can really start to integrate the network into our existing frameworks for processing data.”
Dean and Weaver have already started working on using the neural network in their data processing pipelines. Diffusion MRI requires many steps to process its data and implementing the network into their pipeline would help create a more automated process that is quicker and more accurate, Weaver says. From there, the hope is then to start using it in larger studies such as the Healthy Brain and Child Development Study (HBCD), which is a large longitudinal multisite study looking at infant brain development that Dean and his lab are a part of.
“We want to have the most reliable measures and estimates of brain markers. Motion artifacts influence those measures and the data we get from MRI scans,” Dean says. “So, we are trying to do the best we can at removing those artifacts and getting the best quality data possible that will hopefully be informative for a wide range of studies, including studies of intellectual and developmental disabilities.”
|Your support makes a difference. Donate now to advance knowledge about human development, developmental disabilities, and neurodegenerative diseases through research, services, training, and community outreach.