Deep Learning Automated Algorithms Accurately Segment Stroke Lesions
- Researchers previously developed an automated algorithm, consisting of an ensemble of three-dimensional convolutional neural networks (CNNs), for segmenting ischemic stroke lesions on magnetic resonance imaging (MRI) scans
- The multicenter data were obtained as part of routine clinical practice involving multiple MRI field strengths, sequences, vendors and acquisition protocols
- In this study, an ensemble algorithm trained with MRI data from a single center performed comparably well to an ensemble algorithm trained with multicenter data
- Even better results were obtained when using an ensemble algorithm that combined CNNs trained on diverse data sets
- Fully automated segmentation of multicenter MRI data appears to be feasible and accurate, and it should facilitate big data approaches to phenotyping ischemic stroke
The big data approach to studying genetic pathways underlying ischemic stroke requires pooling information on very large, heterogenous patient populations. Clinical brain scans offer invaluable information, but large variation of their quality poses the greatest challenge in their analysis. As part of these efforts, researchers need an automated way to segment stroke lesions on magnetic resonance imaging (MRI) scans obtained in multiple centers, across multiple machine vendors and acquisition protocols.
Subscribe to the latest updates from Neuroscience Advances in Motion
To this end, Natalia S. Rost, MD, MPH, of the J. Philip Kistler Stroke Research Center at Massachusetts General Hospital, launched the MRI-GENetics Interface Exploration (MRI-GENIE) study, an NIH-funded collaboration between 12 international academic stroke centers.
Earlier this year, Ona Wu, PhD, director of the Clinical Computational Neuroimaging Group at the Athinoula A Martinos Center at Mass General, and colleagues published a report in the American Journal of Neuroradiology about an ensemble of three-dimensional convolutional neural networks (CNN). Neural networks are sets of algorithms designed to recognize patterns; CNNs are used primarily to analyze visual images.
In that study, the researchers demonstrated that these deep learning algorithms, when trained on multiple parametric MRI maps, more accurately segmented ischemic lesions than a CNN that was trained on solo MRI parametric maps. However, the study's results were based on training data from only one center that used machines from one MRI vendor.
In a subsequent paper published in Stroke, Dr. Wu, Dr. Rost and colleagues build on those earlier observations, reporting that an ensemble of CNNs trained with data from a single-center can perform as well as an ensemble trained with multicenter data.
The researchers trained ensembles of six CNNs for each of two cohorts:
- 267 patients with acute ischemic stroke who presented to a single-center between 1996 and 2012 and had MRI done on a 1.5 T scanner from a single vendor
- 267 patients from MRI-GENIE (8 centers), a large data repository run by the Stroke Genetics Network of 12 centers. These patients had MRI done on 1, 1.5 and 3 T scanners from six vendors
The team also evaluated a mixed ensemble of CNNs that included three CNNs from each cohort
The researchers compared the performance of the single-center, MRI-GENIE (12 centers) and mixed data sets of CNNs on an independent data cohort from MRI-GENIE whose data was not used in training either ensemble with regard to:
- Dice score (a measure of overlap between automated and manually segmented lesions): The mixed ensemble was significantly superior to the single-center and MRI-GENIE ensembles. The single-center and MRI-GENIE ensembles were comparable to each other
- Precision (positive predictive value): The mixed ensemble was significantly superior to the single-center ensemble although not the MRI-GENIE ensemble. Again, the single-center and MRI-GENIE ensembles were comparable
- Sensitivity: There were no significant differences between the ensembles
Performance of the Mixed Ensemble
In further analyses, the researchers evaluated only the mixed ensemble because of its excellent results. For a separate 383-patient subset of MRI-GENIE, the mixed ensemble of CNNs outperformed all individual CNNs with regard to Dice score and precision, although not sensitivity. The median automated lesion volume correlated significantly with lesion outlines drawn by human readers of the MRI scans.
The mixed ensemble was then applied to the entire 2,270 MRI datasets in the MRI-GENIE repository. The median automated lesion volume was 3.7 cm3 and correlated with stroke severity.
Support for the Big Data Approach
This study demonstrates the feasibility of automated lesion segmentation, using large, multicenter stroke data repositories, to facilitate the genetic discovery of acute clinical MRI phenotypes. It lays the foundation for future investigations associated with large ischemic lesions on MRI scans.
Automated tools powered by artificial intelligence will also allow high-throughput studies of how imaging phenotypes are associated with genetics, stroke severity and long-term functional outcomes in large multicenter datasets.
view original journal article Subscription may be required
Learn more about the Department of Neurology
Learn more about the Athinoula A. Martinos Center for Biomedical Imaging