A Case for Automated Segmentation of MRI Data in Milder Neurodegenerative Diseases
Connor J Lewis1, Jean M Johnston1, Precilla D'Souza1
1Office of the Clinical Director and Medical Genetics Branch, National Human Genome Research Institute, 10 Center Drive, Bethesda MD USA.
Background:
Volumetric analysis and segmentation of magnetic resonance imaging (MRI) data is an important tool for evaluating neurological disease progression and neurodevelopment. Fully automated segmentation pipelines offer faster and more reproducible results. However, since these analysis pipelines were trained on or run based on atlases consisting of neurotypical controls, it is important to evaluate how accurate these methods are for neurodegenerative diseases. In this study, we compared 5 fully automated segmentation pipelines including FSL, Freesurfer, volBrain, SPM12, and SimNIBS with a manual segmentation process in GM1 gangliosidosis patients and neurotypical controls.
Methods:
We analyzed 45 MRI scans from 16 juvenile GM1 gangliosidosis patients, 11 MRI scans from 8 late-infantile GM1 gangliosidosis patients, and 19 MRI scans from 11 neurotypical controls. We compared results for 7 brain structures including volumes of the total brain, bilateral thalamus, ventricles, bilateral caudate nucleus, bilateral lentiform nucleus, corpus callosum, and cerebellum.
Results:
We found volBrain's vol2Brain pipeline to have the strongest correlations with the manual segmentation process for the whole brain, ventricles, and thalamus. We also found Freesurfer's recon-all pipeline to have the strongest correlations with the manual segmentation process for the caudate nucleus. For the cerebellum, we found a combination of volBrain's vol2Brain and SimNIBS' headreco to have the strongest correlations depending on the cohort. For the lentiform nucleus, we found a combination of recon-all and FSL's FIRST to give the strongest correlations depending on the cohort. Lastly, we found segmentation of the corpus callosum to be highly variable.
Conclusion:
Previous studies have considered automated segmentation techniques to be unreliable, particularly in neurodegenerative diseases. However, in our study we produced results comparable to those obtained with a manual segmentation process. While manual segmentation processes conducted by neuroradiologists remain the gold standard, we present evidence to the capabilities and advantages of using an automated process including the ability to segment white matter throughout the brain or analyze large datasets, which pose feasibility issues to fully manual processes. Future investigations should consider the use of artificial intelligence-based segmentation pipelines to determine their accuracy in GM1 gangliosidosis, lysosomal storage disorders, and other neurodegenerative diseases.
