Abstract
Neuroimaging measurements of brain structural integrity are vital surrogates for brain health, yet precise assessments often require advanced, time-consuming image acquisitions. This study evaluates a radiomic approach to assess brain structural integrity by predicting white matter hyperintensity (WMH) burden from conventional T2-FLAIR images and uncovering associations between these features and clinical phenotypes. Analyzing a multi-site cohort of 4,163 acute ischemic stroke (AIS) patients, we found that radiomic features extracted from normal-appearing brain tissue were highly predictive of WMH burden ($R^2 = 0.855 \pm 0.011$). Furthermore, canonical correlation analysis revealed significant relationships between the radiomic signature and clinical traits such as age, sex, hypertension, and smoking history. These findings suggest that radiomics can capture subvisible microstructural damage on routine clinical scans, offering a potential tool for evaluating brain health and predicting disease progression.
Introduction
Small Vessel Disease (SVD) is a primary driver of stroke and cognitive decline. In clinical practice, SVD is typically monitored via White Matter Hyperintensities (WMH) visible on MRI. However, these visible lesions represent macrostructural damage, often appearing only after significant microstructural degradation has occurred. While advanced techniques like Diffusion Tensor Imaging (DTI) can detect earlier changes, their high cost and long acquisition times limit clinical utility.
Radiomics—the high-throughput extraction of quantitative features from medical images—offers a solution. By describing the texture, shape, and intensity distributions of images beyond human perception, radiomics may identify "subvisible" markers of brain health on standard T2-FLAIR scans. This study aims to establish a radiomic signature of WMH and interpret its relationship with common cardiovascular risk factors in the largest such study to date.
Methods
Study Population and Data
The study utilized the MRI-GENIE (MRI-GENetics Interface Exploration) dataset, comprising 4,163 AIS patients from 17 international sites. This diversity ensures the robustness of the findings across different clinical MRI scanners.
Computational Pipeline
- Segmentation: Deep-learning models automatically segmented the brain and visible WMH.
- Feature Extraction: Using the
PyRadiomicstoolbox, 763 rotation-invariant features were extracted from "normal-appearing" brain tissue (total brain mask minus the WMH mask). These features included intensity histograms and complex textural descriptors. - Predictive Modeling: ElasticNet linear regression was employed to predict personalized WMH burden based on these textural features.
- Clinical Mapping: Canonical Correlation Analysis (CCA) was used to relate the radiomic signature to clinical variables: age, sex, hypertension (HTN), atrial fibrillation (AF), diabetes mellitus (DM), coronary artery disease (CAD), and smoking history.
Results
Prediction of WMH Burden
Radiomic features demonstrated exceptional predictive power for WMH burden, achieving an $R^2$ of $0.855 \pm 0.011$. Notably, when the model relied solely on brain shape and size, the predictive power dropped significantly ($R^2 = 0.41$), highlighting that the critical information resides in the subvisible voxel intensities and textures.
Clinical Phenotype Associations
Seven pairs of canonical variates (CV) significantly correlated the radiomic signature with clinical traits. The strongest correlation was found in CV1 ($r = 0.81$), primarily driven by age. CV2 ($r = 0.65$) was influenced by sex.
Specific risk factors were associated with distinct "textural fingerprints":
- Hypertension: Linked to "patchy" textures (inhomogeneity at lower spatial frequencies).
- Smoking and Diabetes: Associated with fine-grained heterogeneity (high spatial frequency changes).
- Atrial Fibrillation: Correlated with "spots" of subvisible damage, potentially reflecting embolic origins.
Discussion
The ability of radiomics to predict visible damage from "normal-appearing" tissue supports the "WMH penumbra" hypothesis—that brain degradation is a continuum. Radiomics allows clinicians to "see" the damage before it manifests as a visible lesion.
Implications
This approach "democratizes" advanced diagnostics. Because it utilizes standard T2-FLAIR images, it could be implemented in existing clinical workflows without new hardware. It provides a novel solution for assessing brain structural integrity, which is crucial for treatment selection, prognosis, and recovery prediction in stroke and other neurological diseases.
Limitations and Future Directions
The study did not explicitly mask acute stroke lesions, which might introduce minor bias. Additionally, features were extracted from the whole brain rather than isolated white matter. Future research should benchmark this radiomics approach against DTI metrics to confirm its role as a cost-effective surrogate for predicting cognitive and functional outcomes.
Conclusion
Radiomics extracted from routine T2-FLAIR images of AIS patients capture microstructural damage and correlate strongly with clinical phenotypes. This method offers a scalable, high-resolution window into brain health, paving the way for better monitoring of small vessel disease progression.
Summary updated