Nevin Manimala Statistics

Texture Features of Magnetic Resonance Images Predict Poststroke Cognitive Impairment: Validation in a Multicenter Study

Stroke. 2022 Jul 13:101161STROKEAHA122039732. doi: 10.1161/STROKEAHA.122.039732. Online ahead of print.


BACKGROUND: Imaging features derived from T1-weighted (T1w) images texture analysis were shown to be potential markers of poststroke cognitive impairment, with better sensitivity than atrophy measurement. However, in magnetic resonance images, the signal distribution is subject to variations and can limit transferability of the method between centers. This study examined the reliability of texture features against imaging settings using data from different centers.

METHODS: Data were collected from 327 patients within the Stroke and Cognition Consortium from centers in France, Germany, Australia, and the United Kingdom. T1w images were preprocessed to normalize the signal intensities and then texture features, including first- and second-order statistics, were measured in the hippocampus and the entorhinal cortex. Differences between the data led to the use of 2 methods of analysis. First, a machine learning modeling, using random forest, was used to build a poststroke cognitive impairment prediction model using one dataset and this was validated on another dataset as external unseen data. Second, the predictive ability of the texture features was examined in the 2 remaining datasets by ANCOVA with false discovery rate correction for multiple comparisons.

RESULTS: The prediction model had a mean accuracy of 90% for individual classification of patients in the learning base while for the validation base it was ≈ 77%. ANCOVA showed significant differences, in all datasets, for the kurtosis and inverse difference moment texture features when measured in patients with cognitive impairment and those without.

CONCLUSIONS: These results suggest that texture features obtained from routine clinical MR images are robust early predictors of poststroke cognitive impairment and can be combined with other demographic and clinical predictors to build an accurate prediction model.

PMID:35862196 | DOI:10.1161/STROKEAHA.122.039732

By Nevin Manimala

Portfolio Website for Nevin Manimala