Stroke is frequently associated with delayed, long-term cognitive impairment (CI) and dementia. Recent research has focused on identifying early predictive markers of CI occurrence. We carried out a texture analysis of magnetic resonance (MR) images to identify predictive markers of CI occurrence based on a combination of preclinical and clinical data. Seventy-two-hour post-stroke T1W MR images of 160 consecutive patients were examined, including 75 patients with confirmed CI at the 6-month post-stroke neuropsychological examination. Texture features were measured in the hippocampus and entorhinal cortex and compared between patients with CI and those without. A correlation study determined their association with MoCA and MMSE clinical scores. Significant features were then combined with the classical prognostic factors, age and gender, to build a machine learning algorithm as a predictive model for CI occurrence. A middle cerebral artery transient occlusion model was used. Texture features were compared in the hippocampus of sham and lesioned rats and were correlated with histologically assessed neural loss. In clinical studies, two texture features, kurtosis and inverse difference moment, differed significantly between patients with and without CI and were significantly correlated with MoCA and MMSE scores. The prediction model had an accuracy of 88 ± 3%. The preclinical model revealed a significant correlation between texture features and neural density in the hippocampus contralateral to the ischemic area. These preliminary results suggest that texture features of MR images are representative of neural alteration and could be a part of a screening strategy for the early prediction of post-stroke CI.
Keywords: Cognitive impairment; Neuron loss; Predictive features; Radiomics; Stroke; Texture analysis.