Characterization and Classification of Brain Tissue and Stroke Lesions in Non-Contrast Computed Tomography Images of Stroke Patients Using Statistical Texture Descriptors and Artificial Neural Network

Abstract

Aim: To characterize and classify stroke lesions and normal brain tissue in computed tomography (CT) images using statistical texture descriptors.

Patients and methods: Two experienced radiologists blinded to each other inspected CT images of 164 stroke patients to identify and categorize stroke lesions into ischaemic and haemorrhagic subtypes. Four regions of interest (ROIs) in each CT slice that demonstrated the lesion; two each representing the lesion and normal tissue were selected. Statistical texture descriptors namely, co-occurrence matrix, run-length matrix, absolute gradient and histogram were calculated for them.  Raw data analysis was performed to identify the parameters that best discriminate between normal brain tissue and stroke lesions. Artificial neural network (ANN) was used to classify the ROIs into normal tissue, ischaemic and haemorrhagic lesions using the radiologists’ identification and categorization as the gold standard, and further analyzed using the receiver operating characteristic curve.

Results: Three parameters in each texture class discriminated between normal tissue, ischaemic and haemorrhagic stroke lesions. The discriminating co-occurrence matrix parameters were sum average parameters namely S1-1 SumAverg, S1-0 SumAverg and S0-1 SumAverg.  For the run-length matrix, short run emphasis in horizontal, 135⁰ and 45⁰ directions were the discriminating features. The discriminating absolute gradient parameters were gradient non-zeros, gradient variance and gradient mean. For the histogram class, the mean, 90^th and 99^th percentiles 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 co-occurrence matrix. With run-length matrix the sensitivity was 0.544, specificity 0.607, FPR 0.393, and FNR 0.456 while with absolute gradient the sensitivity was 0.546, specificity 0.586, FPR 0.414, FNR 0.454. With histogram, the sensitivity was 0.947, specificity 0.962, FPR 0.038, and FNR 0.053.

Conclusion: The histogram texture features showed the highest sensitivity and specificity in the classification of brain tissue and stroke lesions using the artificial neural network.