Computer-Aided Diagnosis of Hyperacute Stroke with Thrombolysis Decision Support Using a Contralateral Comparative Method of CT Image Analysis

Abstract

New and improved techniques have been continuously introduced into CT and MR imaging modalities for the diagnosis and therapy planning of acute stroke. Nevertheless, non-contrast CT (NCCT) is almost always used by every institution as the front line diagnostic imaging modality due to its high affordability and availability. Consequently, the potential reward of extracting as much clinical information as possible from NCCT images can be very great. Intravenous tissue plasminogen activator (tPA) has become the gold standard for treating acute ischemic stroke because it is the only acute stroke intervention approved by the FDA. ASPECTS scoring based on NCCT images has been shown to be a reliable scoring method that helps physicians to make sound decisions regarding tPA administration. In order to further reduce inter-observer variation, we have developed the first end-to-end automatic ASPECTS scoring system using a novel method of contralateral comparison. Due to the self-adaptive nature of the method, our system is robust and has good generalizability. ROC analysis based on evaluation of 103 subjects who presented to the stroke center of Chang Gung Memorial Hospital with symptoms of acute stroke has shown that our system’s dichromatic classification of patients into thrombolysis indicated or thrombolysis contraindicated groups has achieved a high accuracy rate with AUC equal to 90.2 %. The average processing time for a single case is 170 s. In conclusion, our system has the potential of enhancing quality of care and providing clinical support in the setting of a busy stroke or emergency center.

Appendix

Calculation of goodness-of-match metric between target image and reference template

The calculation of goodness of match between the target image and the reference template based on the four red boundaries in Fig. 9 can be expressed mathematically as follows. At each trial displacement, (Δx, Δy), the goodness of match for UR_reg_boundary, G_UR(Δx, Δy), is defined by Eq. (5):

$$ \begin{array}{c}\hfill {G}_{\mathrm{UR}}\left(\varDelta x,\varDelta y\right)={\displaystyle \sum_{\mathrm{pi}\kern0.5em \in \kern0.5em \mathrm{UR}\_\mathrm{reg}\_\mathrm{boundary}}{\left({B}_{\mathrm{pi}}\right)}_{\varDelta x,\varDelta y}}\hfill \\ {}\hfill +\left(\mathrm{More}\_\mathrm{than}\_10\%\_\mathrm{of}\_\mathrm{UR}\_\mathrm{reg}\_\mathrm{boundary}\_\mathrm{in}\_\mathrm{lateral}\_\mathrm{ventricles}\right)\times \left(-1,000\right)\hfill \end{array} $$

(5)

where B _pi = 1 if pixel(x _pi + Δx, y _pi + Δy) has at least one but not all of its eight neighbors overlapping with the lateral ventricles of the target image

B _pi = 0 otherwise

(More_than_10%_of_UR_reg_boundary_in_lateral_ventricles) = 1

if more than 10 % of UR_reg_boundary pixels fall within the lateral ventricles,

(More_than_10%_of_ UR_reg_boundary_in_lateral_ventricles) = 0 otherwise;

At each trial displacement, (Δx, Δy), a pixel, pi, may find itself in one of three possible locations with respect to the lateral ventricle in the right hemisphere: (1) the pixel is away from the lateral ventricle if none of the eight neighbors of (x _pi + Δx, y _pi + Δy) belong to the lateral ventricle, (2) the pixel is on the boundary of the lateral ventricle if at least one but not all of its eight neighbors of (x _pi + Δx, y _pi + Δy) belong to the lateral ventricle, and (3) the pixel is inside the lateral ventricle if all of its eight neighbors of (x _pi + Δx, y _pi + Δy) belong to the lateral ventricle. The first term counts the number of pixels of the UR_reg_boundary of the reference template that coincide with the right lateral ventricle boundary on the target image. Each such coincidence is given a point. On the other hand, the number of pixels of UR_reg_boundary of the reference template that fall within the lateral ventricle on the target image is also counted. A very large penalty of −1,000 points is imposed if the number of such pixels exceeds 10 % of UR_reg_boundary. The trial displacement, (Δx, Δy)_UR, that has the largest value of G_UR(Δx, Δy) is the best match.

Likewise, at each trial displacement, (Δx, Δy), the goodness of match for LR_reg_boundary, G_LR(Δx, Δy), is defined by Eq. (6) below:

$$ \begin{array}{c}\hfill {G}_{\mathrm{LR}}\left(\varDelta x,\varDelta y\right)={\displaystyle \sum_{\mathrm{pi}\kern0.5em \in \kern0.5em \mathrm{LR}\_\mathrm{reg}\_\mathrm{boundary}}{\left({B}_{\mathrm{pi}}\right)}_{\varDelta x,\varDelta y}}\hfill \\ {}\hfill +\left(\mathrm{More}\_\mathrm{than}\_10\%\_\mathrm{of}\_\mathrm{LR}\_\mathrm{reg}\_\mathrm{boundary}\_\mathrm{in}\_3\mathrm{V}\_\mathrm{or}\_\mathrm{QC}\right)\times \left(-1,000\right)\hfill \end{array} $$

(6)

where B _pi = 1 if pixel(x _pi + Δx, y _pi + Δy) has at least one but not all of its eight neighbors overlapping with the conglomerate region comprising the third ventricle and the quadrigeminal cistern

B _pi = 0 otherwise;

(More_than_10%_of_LR_reg_boundary_in_3V_or_QC) = 1 if more than 10 % of LR_reg_boundary pixels fall within the conglomerate region comprising the third ventricle and the quadrigeminal cistern,

(More_than_10%_of_upper_boundary_ in_3V_or_QC) = 0 otherwise;

Goodness-of-match metrics for the two boundaries on the left hemisphere, G _UL(Δx, Δy) and G _LL(Δx, Δy), can be calculated in a similar manner.