Outcome prediction is critical in stroke patient management. We propose a novel approach combining imaging with parameters (including history, hospitalization, demographics, clinical and outcome) for a population of patients in the Probabilistic Stroke Atlas (PSA) along with prediction engine. The PSA aggregates multiplicity of data for a population of stroke patients and presents them in image format. The PSA is composed from a series of three-dimensional (3D) image volumes including scans and parameters. A cohort of over 700 ischemic stroke generally treated patients with 176 parameters per patient, and CT scan performed at admission and on day 7 was acquired. Outcome measurements were assessed up to one year after stroke onset. Cases with old infarcts, infarcts in both hemispheres, and hemorrhagic transformations were rejected. This data was post-processed to build the PSA and then the PSA was used for prediction. The infarcts were delineated on CT scans and their 3D surface models constructed and normalized. The PSA was calculated from the normalized 3D infarct models as frequency of stroke occurrence. Similar maps were calculated for the following parameters: Age; Sex; Survival; NIH Stroke Scale (NIHSS); Barthel Index (BI) at 30, 90, 180, 360 days; modified Rankin Scale (mRS) at 7, 30, 90, 180, 360 days; White blood cell count; C-reative protein; Glucose at emergency department; History of hypertension; and History of diabetes. The PSA was used for prediction of mRS and BI for 50 stroke subjects. For a given case to be predicted, the infarct was delineated and analyzed by the PSA mapped on the scan. The predicted values of the parameters from the PSA were compared with the actual values of the parameters measured in up to 1-year neurological follow up. The accuracy was defined as 100*(1-(actual value-predicted value)/actual value)%. The mean prediction accuracy of mRS at (7, 30, 90, 180, 360) days is (89.7, 90.7, 92.1, 87.0, 83.3)% and that for BI at (30, 90, 180, 360) days is (90.0, 95.4, 94.4, 92.2)% respectively. This novel prediction method has high prediction rates. It can be applied to any other parameters. The PSA is dynamic and its power can increase with additional cases.
ISLES 2016 and 2017-Benchmarking Ischemic Stroke Lesion Outcome Prediction Based on Multispectral MRI
