Meeting Banner
Abstract #1020

Validation of Automatic Segmentation Algorithms of DWI in Acute Stroke Patients in Independent Data

Steven Mocking1, 2, Raquel Bezerra1, Elissa McIntosh1, Izzudin Diwan1, Priya Garg1, William Taylor Kimberly3, Ethem Murat Arseva1, Hakan Ay1, Aneesh B. Singhal3, William A. Copen4, Pamela Schaefer4, Ona Wu1

1Athinoula A. Martinos Center for Biomedical Imaging, Dept. of Radiology, MGH, Charlestown, MA, United States; 2Image Sciences Institute, University Medical Center Utrecht, Utrecht, Netherlands; 3Dept. of Neurology, MGH, Boston, MA; 4Dept. of Radiology, MGH, Boston, MA

We validated five algorithms for segmenting DWI lesions in acute ischemic stroke on an independent dataset and investigated the proportion of acutely misclassified tissue with follow-up manual outlines. Performance of the algorithms on the validation dataset was comparable to results on the training dataset, with the naive Bayes approach providing best sensitivity and Dice similarity coefficient. In many cases, several voxels classified as lesion or normal by the automatic algorithms but not by the manual outliner were found respectively to be infarcted or normal on follow-up imaging. Consequently, apparent misclassification may partially be a result of reader variation.

Abstract PDF Poster PDF

Keywords

accurate acute affected agrees analyzed apparent appeared application approaches arterial assigning automated automatic automatically available background ball better biomedical blinded bottom choice class classification classifier clinical coefficient comparing computed computes considered correction cover dataset days defined delineated delineations dept determined deviation diagnosis dice diffusion disagreement early either ensures estimation evaluated experimental false feature features five flair fractions generated geometric greater green hart independent inferior institute interquartile intra intravenous investigate known labels lagged lesion lesions likelihood made manual manually marker media median medical mere mocking multiplied naive nearest negatives neighbor neither neurology none occurred odds operationally outline outlined outlines outlining output overall patients performance performances perfusion poor positives post preparation previously probability prognosis radiology rapid reader regard registered reported reproducibility respectively retrospectively sample segmentation sensitivity sided significantly similarity slice specificities specificity stroke substantial suggesting table terms theory therapy thresholding thresholds tissue together trace trained training true validation variance visually whereas whether yellow