Portable, low-field-strength MRI systems have the potential to transform neuroimaging – provided that their low spatial resolution and low signal-to-noise (SNR) ratio can be overcome. Researchers at Harvard Medical School are harnessing artificial intelligence (AI) to achieve this goal. They have developed a machine learning super-resolution algorithm that generates synthetic images with high spatial resolution from lower resolution brain MRI scans.
The convolutional neural network (CNN) algorithm, known as LF-SynthSR, converts low-field-strength (0.064 T) T1- and T2-weighted brain MRI sequences into isotropic images with 1 mm spatial resolution and the appearance of a T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) acquisition. Describing their proof-of-concept study in Radiology, the researchers report that the synthetic images exhibited high correlation with images acquired by 1.5 T and 3.0 T MRI scanners.
Morphometry, the quantitative size and shape analysis of structures in an image, is central to many neuroimaging studies. Unfortunately, most MRI analysis tools are designed for near-isotropic, high-resolution acquisitions and typically require T1-weighted images such as MP-RAGE. Their performance often drops rapidly as voxel size and anisotropy increase. As the vast majority of existing clinical MRI scans are highly anisotropic, they cannot be reliably analysed with existing tools.
“Millions of low-resolution brain MR images are produced every year, but currently cannot be analysed with neuroimaging software,” explains principal investigator Juan Eugenio Iglesias. “The main goal of my current research is to develop algorithms that make low-resolution brain MR images look like the high-resolution MRI scans that we use in research. I am particularly interested in two applications: enabling automated 3D analysis of the clinical scans and use with portable, low-field MRI scanners.”
Training and testing
LF-SynthSR is built upon SynthSR, a method developed by the team to train a CNN to predict 1 mm-resolution MP-RAGE isotropic scans from routine clinical MR scans. Previous findings reported in NeuroImage showed that SynthSR-generated images could be reliably used for subcortical segmentation and volumetry, image registration and, if some quality requirements are met, even cortical thickness morphometry.
Both LF-SynthSR and SynthSR are trained on synthetic input images of highly varying appearance generated from 3D segmentations, and can thus be used to train CNNs for any combination of contrast, resolution and orientation.
Iglesias points out that neural networks perform best when data appear approximately constant, but every hospital uses scanners from different vendors that are configured differently, resulting in highly heterogeneous scans. “To tackle this problem, we are borrowing ideas from a field of machine learning called ‘domain randomization’, where you train neural networks with synthetic images that are simulated to constantly change appearance and resolution, in order to obtain trained networks that are agnostic to the appearance of the input images,” he explains.
To assess the performance of LF-SynthSR, the researchers correlated brain morphology measurements between synthetic MRIs and ground-truth high-field strength images. For training, they used a high-field-strength MRI dataset of 1-mm isotropic MP-RAGE scans from 20 subjects. They also used corresponding segmentations of 36 brain regions-of-interest (ROIs) and three extracerebral ROIs. The training set was also artificially augmented to better model pathologic tissue such as stroke or haemorrhage.
The test set comprised imaging data from 24 participants with neurological symptoms who had a low-field strength (0.064 T) scan in addition to a standard-of-care high-field strength (1.5–3 T) MRI. The algorithm successfully generated 1-mm isotropic synthetic MP-RAGE images from the low-field strength brain MRIs, with voxels more than 10 times smaller than in the original data. Automated segmentation of the synthetic images from a final sample of 11 participants yielded ROI volumes that were highly correlated with those derived from the high-field strength MR scans.
“LF-SynthSR may improve the image quality of low-field strength MRI scans to the point that they are usable not only by automated segmentation methods but potentially also with registration and classification algorithms,” the researchers write. “It could also be used to augment the detection of abnormal lesions.”
This ability to analyse low-resolution brain MRIs using automated morphometry would enable the study of rare diseases and populations that are under-represented in current neuroimaging research. In addition, improving the quality of images from portable MRI scanners would enhance their use in medically underserved areas, as well as in critical care, where moving patients to an MRI suite is often too risky.
Iglesias says that another challenge is the wide range of abnormalities found in clinical scans that need to be handled by the CNN. “Currently, SynthSR works well with healthy brains, cases with atrophy, and smaller abnormalities like small multiple sclerosis lesions or small strokes,” he tells Physics World. “We are currently working to improve the method so it can effectively deal with larger lesions, like larger strokes or tumours.”
Writing in an accompanying editorial in Radiology, Birgit Ertl-Wagner and Matthias Wagner from the Hospital for Sick Children in Toronto comment: “This exciting technical development study demonstrates the potential to go low on field strength and aim high for spatial and contrast resolution using artificial intelligence.”
The post AI creates high-resolution brain images from low-field strength MR scans appeared first on Physics World.