Contrast-Free ADC Intensity Features for Early Neurological Deterioration in Lacunar Infarction: An Exploratory Comparison with Perfusion Imaging and Deep Learning.

Authors

Abstract

Early neurological deterioration (END) affects approximately one in five patients with acute lacunar infarction and is associated with poor functional outcome. Hypoperfusion on contrast-enhanced perfusion MRI has been linked to END, but perfusion imaging requires intravenous contrast and is not part of the routine acute workup for minor stroke. We explored whether intensity features of non-contrast structural MRI carry information on END, using the visually classified pattern from a contrast-enhanced perfusion study in the same patients as a reference. We analyzed a prospectively collected single-center cohort of patients with acute lacunar infarction who underwent MRI within 24 h of symptom onset, including diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) mapping, fluid-attenuated inversion recovery (FLAIR), and dynamic susceptibility contrast perfusion imaging. END was defined as an increase of ≥ 2 points on the National Institutes of Health Stroke Scale within 72 h. One hundred patients with complete imaging and outcome data were included, of whom 20 (20.0%) developed END. Candidate predictors comprised intensity features of the infarct core and an 8-mm perilesional rim normalized to whole-brain intensity, 201 DWI texture features, and 22 clinical and laboratory variables. Models were evaluated by 10-times-repeated 5-fold stratified cross-validation with the decision threshold fixed at 0.5; feature selection and standardization were performed within training folds only. The visually classified perfusion pattern served as a contrast-based clinical benchmark. ADC intensity features discriminated END with an area under the receiver operating characteristic curve (AUC) of 0.699 (approximate patient-level 95% CI 0.56-0.84) and an area under the precision-recall curve (AUPRC) of 0.473 against a base rate of 0.200; at the fixed threshold, sensitivity was 52.0% and positive predictive value 27.9%. The perfusion pattern alone yielded a sensitivity of 70.0%, a specificity of 60.0% and a negative predictive value of 88.9% (balanced accuracy 0.650). The difference between the two predictors was small relative to the sampling uncertainty at 20 events, so neither superiority nor equivalence could be established. The strongest individual feature was the 75th percentile of normalized ADC within the infarct core (univariate AUC 0.702), with higher values in patients who deteriorated (median 0.615 versus 0.555, P = 0.005). Adding the perfusion pattern to the ADC features did not improve discrimination (AUC 0.665, 95% CI 0.645-0.684). The signal was confined to the ADC map: models built on DWI or FLAIR intensity alone performed below chance (AUC 0.463 and 0.453). More complex models did not reach the ADC model at this sample size: 201 DWI texture features reached an AUC of 0.566, clinical and laboratory variables at best 0.625, and an end-to-end 2.5D ResNet-18 0.404 with random initialization, 0.482 with ImageNet pretraining and 0.611 for a small network sized for the data. In this exploratory single-center study, ADC intensity features obtained without contrast were associated with END in acute lacunar infarction, with discrimination similar in magnitude to the contrast-enhanced perfusion benchmark; the cohort was too small to establish equivalence, and threshold-level performance is insufficient to guide monitoring. Texture analysis and deep learning could not be adequately trained with 20 events, so their relative merit cannot be judged. ADC appears to be the informative non-contrast sequence; the findings are hypothesis-generating and require validation in larger, multicenter cohorts.

Tags

Topics