Arthur Goldsmith (Columbia University)
Ciprian Crainiceanu (Johns Hopkins University)
Dzung Pham (Henry M. Jackson Foundation)
Elizabeth Sweeney (Johns Hopkins University)
Navid Shiee (Henry M. Jackson Foundation)
Russell Shinohara (Johns Hopkins University)
This invention is a novel statistical method for automatically detecting lesions in cross-sectional brain magnetic resonance imaging (MRI) studies. OASIS uses multimodal MRI from one image acquisition session and produces voxel-level probability maps of the brain that quantifies the likelihood that each voxel is part of a lesion. Binary lesion segmentations are created from these probability maps using a validated population-level threshold. In this application, fluid attenuated inversion recovery (FLAIR), proton density (PD), T2-weighted, and Tl-weighted volumes were used. The OASIS lesion segmentations are robust to changes in imaging centers and scanning parameters.
Traditionally trained observers do lesion identification and quantification manually on each MRI image. But this manual quantification is slow and prone to human error. There are many available methods for automated lesion segmentation. These methods are often tuned to one dataset and do not generalize to new datasets with different scanning parameters. The OASIS statistical method provides a fully automated, sensitive, and specific method for lesion segmentation that easily generalizes to new datasets.