Quinn Thomson MS, Presenter: Nothing to Disclose

Jayesh Ashok Modi MD, Abstract Co-Author: Nothing to Disclose

Shelagh Coutts MD, Abstract Co-Author: Nothing to Disclose

Andrew Demchuk MD, Abstract Co-Author: Nothing to Disclose

Michael D. Hill MD, Abstract Co-Author: Nothing to Disclose

Mayank Goyal MD, FRCPC, Abstract Co-Author: Shareholder, Calgary Scientific, Inc Research Grant, Bayer AG Consultant, Covidien AG Shareholder, NONo Inc

Steven K Boyd PhD, Abstract Co-Author: Nothing to Disclose

Ross Mitchell PhD, Abstract Co-Author: Co-founder, Calgary Scientific, Inc Intellectual property, Calgary Scientific, Inc Shareholder, Calgary Scientific, Inc Research, Calgary Scientific, Inc

Brain imaging must be assessed prior to treatment of acute stroke. Accurate assessment requires extensive training. Even so, there remains variability between and within assessors. Our objective is to develop a computerized tool to improve the accuracy and reliability of ischemic brain region labeling in acute, non-contrast CT.

Our method compares contralateral sides of the brain to identify ischemic tissue. We compute three different textures (corresponding to image features used for differential diagnosis) in matched spherical regions mirrored across the brain mid-plane. Potential ischemia identified in overlapping spheres results in a voxel being identified multiple times. A threshold is applied to produce a binary map for each texture characteristic. Numerical optimization is used to tune the algorithm to a truth dataset that we constructed. Five expert readers outlined ischemic brain regions in acute CT scans from 20 patients. The process was repeated 3 times, for a total of 300 measurements. The true ischemic region was defined as the majority classification for each voxel. Ischemic areas found by our algorithm visually matched with the regions identified by expert readers. Initial parameter tuning resulted in a Dice coefficient of 0.495 (98.5% accuracy, 53.1% sensitivity, and 99.1% specificity) using only one texture characteristic. The regions identified by the experts resulted in a Dice coefficient of 0.641 (99.4% accuracy, 75.2% sensitivity, 99.6% specificity).  

No gold standard exists for acute ischemic stroke. Consequently, we used 300 measurements to estimate truth. The expert’s Dice coefficient of 0.641 suggests there is considerable variability between experts. A Dice coefficient of 0.495 for our algorithm is promising. Future work will focus on using additional texture information to increase the algorithm sensitivity and Dice coefficient.

Our texture analysis method is able to localize ischemic brain tissue in CT scans of acute ischemic stroke patients. Early results are encouraging. However, it does not yet perform as well as stroke experts.

Thomson, Q, Modi, J, Coutts, S, Demchuk, A, Hill, M, Goyal, M, Boyd, S, Mitchell, R, Acute Ischemic Stroke Texture Analysis: Finding Ischemic Brain Tissue in Early CT Images.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12025540.html