Preprocedural CT Radiomic Features of the Basal Ganglia Are Associated with Post-Thrombectomy Hyperdense Transformation: A LASSO-Based Logistic Regression Analysis.
Bayrak A, Kesimal U, et al. • Tomography (Ann Arbor, Mich.) • 2026
Radiomic analysis of visually unremarkable basal ganglia on preprocedural non-contrast CT identified an exploratory two-feature imaging signature associated with subsequent post-thrombectomy hyperdense transformation, with modest discrimination in temporal validation and evidence of model-development optimism, so the findings are hypothesis-generating.
Key Findings
Results
A LASSO logistic regression model retained two texture features from the ipsilateral basal ganglia as associated with post-thrombectomy hyperdense transformation (HT).
The retained features were Gray-Level Run Length Matrix Short Runs Emphasis and Gray-Level Size Zone Matrix Grey Level Variance.
Features were extracted with LIFEx from manually segmented ipsilateral basal ganglia on preprocedural non-contrast CT.
Candidate features underwent interobserver reproducibility assessment and correlation filtering before LASSO with 10-fold cross-validation.
Results
In the derivation cohort the two-feature model showed moderate discrimination with high specificity but low sensitivity.
Derivation cohort: 84 patients with ASPECTS ≥ 8, no visible basal ganglia abnormality, anterior circulation large-vessel occlusion, and complete reperfusion after thrombectomy (January 2022–July 2023).
AUC was 0.760 (95% CI, 0.652-0.855).
Sensitivity was 48.8%, specificity 95.3%, and accuracy 72.6%.
Results
When applied without modification to an independent temporal validation cohort, the model's performance was lower and the confidence interval was wide.
Validation cohort: 31 consecutive eligible patients treated between August and November 2023.
AUC was 0.677 (95% CI, 0.475-0.858).
Sensitivity was 33.3%, specificity 93.8%, and accuracy 64.5%.
Results
Nested cross-validation showed a lower mean AUC than the derivation estimate and unstable feature selection.
Nested cross-validation yielded a mean AUC of 0.673 ± 0.037.
Feature selection varied across resampling iterations.
The authors interpret this as evidence of model-development optimism and limited feature-selection stability.
Results
Learning-curve analysis showed an early plateau in cross-validated discrimination as training sample size increased.
There was little improvement in cross-validated discrimination as the training sample increased within the evaluated range.
This suggests that simply adding more samples within this range may not improve performance with this modeling approach.
Conclusions
The authors conclude the findings are exploratory and hypothesis-generating, requiring larger datasets, alternative modeling, and multicenter validation.
The study was retrospective and single-center.
The model retained only modest discriminatory ability in temporal validation.
Independent multicenter validation is called for.
What This Means
After clot-removal procedures (thrombectomy) for stroke, some patients develop bright areas on CT scans known as hyperdense transformation, which can reflect contrast leakage or bleeding in brain tissue. This study asked whether computer analysis of the texture of a brain region called the basal ganglia, on the CT taken before the procedure and looking normal to the eye, could be linked to later developing this change. It analyzed 84 patients to build a model and tested it on 31 later patients.
The model used just two texture measurements. It performed moderately in the original group (AUC 0.76) and somewhat worse in the later test group (AUC 0.68). It was good at correctly identifying patients who did not develop the change (specificity about 94-95%) but missed most who did (sensitivity 33-49%). Additional internal testing showed the results were likely somewhat overoptimistic and that the features picked by the method varied from run to run.
This research suggests that subtle CT texture patterns may carry information about later changes after thrombectomy, but the model is not accurate enough for clinical use. The authors describe it as hypothesis-generating. Larger studies across multiple centers would be needed to confirm whether such a signature is reliable.
Bayrak A, Kesimal U, Törenek &, Özdemir Y, Bilgin E, İmamoğlu &, et al.. (2026). Preprocedural CT Radiomic Features of the Basal Ganglia Are Associated with Post-Thrombectomy Hyperdense Transformation: A LASSO-Based Logistic Regression Analysis.. Tomography (Ann Arbor, Mich.). https://doi.org/10.3390/tomography12100149