Abstract
This model, which was based on CT radiomics features, could help to predict hemorrhagic transformation (HT) in the setting of acute ischemic stroke (AIS) for any infarct size and provide guiding suggestions for clinical treatment and prognosis evaluation.
Introduction
In the management of acute ischemic stroke (AIS), rapid intervention via intravenous thrombolysis (IVT) or mechanical thrombectomy (MT) is critical for restoring blood flow. However, these life-saving treatments carry a significant risk of Hemorrhagic Transformation (HT)—bleeding into the infarcted brain tissue. Predicting which patients are at high risk for HT remains a clinical challenge. While clinical scales like the NIHSS and visual CT inspections provide some guidance, they often fail to detect the subtle, microscopic changes in the blood-brain barrier (BBB) that precede a hemorrhage.
This study, published in Frontiers in Neuroscience, investigates the utility of radiomics—the high-throughput extraction of quantitative features from medical images—to identify "invisible" markers of tissue vulnerability on standard non-contrast CT (NCCT) scans.
Methodology
The researchers developed a retrospective radiomics pipeline using data from 118 AIS patients across two medical centers. The workflow consisted of four primary stages:
- Segmentation: Radiologists manually outlined infarcted areas on initial NCCT scans using 3D-Slicer.
- Feature Extraction: 851 features (shape, intensity, and texture) were extracted using the Pyradiomics plugin.
- Feature Selection: A multi-step process was used to ensure robustness and prevent overfitting. This included Interclass Correlation Coefficient (ICC) filtering (>0.75), univariate analysis to identify statistically significant differences between HT and non-HT groups, and Least Absolute Shrinkage and Selection Operator (LASSO) regression to isolate the most impactful predictors.
- Model Construction: A "Rad-score" was derived from the final five features using logistic regression.
Results and Key Insights
The final predictive model demonstrated strong performance, achieving an Area Under the Curve (AUC) of 0.845 in the training cohort and 0.750 in the validation cohort.
The Significance of Wavelet Features
Notably, all five features selected for the final Rad-score were wavelet-based. Wavelet transforms decompose images into various scales and frequencies, suggesting that the critical markers for HT are not found in simple brightness levels but in complex, multi-scale textures. These textures likely reflect the heterogeneity of damaged brain tissue and the physical disruption of the blood-brain barrier.
Radiomics vs. Clinical Predictors
While clinical factors such as "massive cerebral stroke" and high NIHSS scores were associated with HT, the Rad-score emerged as the only independent predictor in multivariate analysis. This indicates that radiomic data captures unique information about the state of the BBB that clinical examinations cannot.
Model Robustness
The model maintained an accuracy of 71.4% to 78.3% regardless of infarct size or treatment method (IVT vs. MT). Furthermore, for patients scanned within the critical 4.5-hour "golden window," the model remained effective with an AUC of 0.754.
Discussion and Limitations
While the results are promising, several steps are required before clinical implementation:
- Sample Size: The cohort of 118 patients is relatively small; larger, multi-ethnic studies are necessary to validate these findings.
- Automation: Manual segmentation is time-consuming. Future work must focus on automated segmentation to make the tool viable for emergency settings.
- Imaging Constraints: CT resolution limits can affect the segmentation of small or poorly defined infarcts.
- Study Design: As a retrospective study, prospective trials are needed to test the model's real-time efficacy in a clinical environment.
Conclusion
The radiomics-based Rad-score offers a potent decision-support tool for emergency stroke interventions. By identifying subtle markers of tissue vulnerability on standard CT scans, this model could help clinicians better navigate the risks of hemorrhagic transformation and optimize treatment strategies for AIS patients.
Summary updated