Texture Characterization of Stroke Lesions in Non-Contrast Computed Tomography Images of Nigerian Patients

Abstract

The aim of this study was to characterize stroke lesions and normal brain tissue in computed tomography (CT) images of Nigerian patients using statistical texture descriptors, and to identify the class of texture descriptor that is most suitable for computer-aided diagnosis of stroke. Non-contrast CT images of 164 stroke patients were obtained in contiguous slices from the base of the skull to the vertex from two private radiodiagnostic centres. Initially, two experienced radiologists blinded to each other, visually inspected the images to identify and categorize the lesions into ischaemic and haemorrhagic subtypes. Four regions of interest (ROIs) were selected on each CT image that demonstrated the lesion; two each represented the lesion and normal tissue respectively. Statistical texture descriptors of co-occurrence matrix, run-length matrix, absolute gradient and histogram, representing spatial distribution of grey levels in the images were calculated. Raw data analysis was carried out to identify the best parameters that discriminated between normal brain tissue and stroke lesions. Three parameters in each texture class discriminated between normal brain tissue, ischaemic and haemorrhagic stroke lesions. Artificial neural network (ANN) and k-nearest neighbour (k-NN) algorithms were used to classify the ROIs into normal tissue, ischaemic and haemorrhagic lesions using the radiologists’ identification and categorization as the gold standard. The classification of ROIs was compared with the radiologists’ categorization of lesions and normal tissues, and further analyzed using the receiver operating characteristic curve to establish the sensitivity and specificity of ANN and k-NN in identifying stroke lesions. The discriminating co-occurrence matrix parameters were sum average parameters namely S1-1 SumAverg with feature value of -3.54 to 4.35, S1-0 SumAverg -4.19 to 4.39 and S0-1 SumAverg -3.87 to 4.30. For the run-length matrix, short run emphasis in the horizontal, 1350 and 450 directions with feature values of -9.08 to 2.27, -9.61 to 2.13 and -9.13 to 2.16 were the discriminating features. The discriminating absolute gradient-derived parameters were gradient non-zeros with feature value of -14.33 to 0.83, gradient variance -2.71 to 4.00 and gradient mean -3.96 to 2.58. For the histogram class, the mean with feature value of -1.77 to 2.59, 90 percentile -1.83 to 2.19 and 99 percentile -1.99 to 1.91 were the discriminating parameters. The ANN achieved a sensitivity of 0.637, specificity 0.753, false positive rate (FPR) 0.247, and false negative rate (FNR) 0.363 with the co-occurrence matrix. With the run-length matrix it achieved a sensitivity of 0.544, specificity 0.607, FPR 0.393, and FNR 0.456 while with the absolute gradient it achieved a sensitivity of 0.546, specificity 0.586, FPR 0.414, FNR 0.454. With the histogram it achieved a sensitivity of 0.947, specificity 0.962, FPR 0.038, and FNR 0.053. The k-NN achieved a sensitivity of 0.644, specificity 0.759, false FPR 0.241, and FNR 0.356 with the co-occurrence matrix. With the run-length matrix it achieved a sensitivity of 0.481, specificity 0.676, FPR 0.324, and FNR 0.519 while with the absolute gradient it achieved a sensitivity of 0.445, specificity 0.651, FPR 0.349, and FNR 0.555. With the histogram it achieved a sensitivity of 0.929, specificity 0.955, FPR 0.045, and FNR 0.071. The histogram-based classification was significantly better than other statistical texture descriptors using the ANN and k-NN (p < 0.05). The histogram class of texture features also showed the highest sensitivity and specificity in classification of brain tissue and therefore is adjudged most suitable for computer-aided diagnosis of stroke. The results suggest that histogram-derived features can be used in computer-aided diagnosis of stroke on non contrast brain CT and can improve diagnosis.