Computer-aided detection scheme for identification of hypoattenuation of acute stroke in unenhanced CT

Abstract

Our purpose in this study was to develop a computer-aided detection scheme for identification of hypoattenuation of acute stroke on unenhanced CT images to select patients for thrombolysis of acute stroke. This method is based on a z-score mapping method. The algorithm of the developed method consisted of five main steps: anatomic standardization, calculation of the z-score with a normal reference database, extraction of candidate voxels for hypoattenuation, feature extraction, and classification. The territory of the middle cerebral artery was divided into ten specified regions, according to a visually quantitative CT scoring system, the Alberta Stroke Programme Early CT Score (ASPECTS) method. Each of the ASPECTS-defined regions was classified as hypoattenuation or normality by linear discriminate analysis. The method was applied to 26 patients who had hypoattenuation areas (<6h). The performance of this scheme for classification of hypoattenuation was evaluated using a leave-one-case-out method. As a result, an average sensitivity of 89.7% and an average specificity of 85.2% for automatically classifying hypoattenuation regions in the lentiform nucleus and the insular regions were obtained, and the average accuracy for the classification of hypoattenuation per patient was 84.6% (range 55.6-100%). The newly developed method has the potential accurately to identify hypoattenuation of acute stroke in the ASPECTS-defined regions.