Faculty Bibliography

BACKGROUND AND PURPOSE/OBJECTIVE: MATERIALS AND METHODS/METHODS: RESULTS: CONCLUSIONS:The dynamics of the CVR response can reveal differential and previously occult temporal signatures in the pharmacologic response of white matter to ACZ, suggesting potential interactions between macrovascular and microvascular hemodynamic impairment.

BACKGROUND:Posterior circulation large-vessel occlusion strokes have significant morbidity and mortality, but patient selection for acute interventions remains understudied. Multiple computed tomography (CT)-based scores exist, including the CT perfusion-based critical area perfusion score (CAPS), CT angiography-based basilar artery treatment management and posterior circulation CT angiography scores, and CT angiography source image or noncontrast CT-based posterior circulation Alberta Stroke Program Early CT Score, but their predictive values for long-term outcomes after thrombectomy have not been directly compared. METHODS:We conducted a retrospective multicenter cohort study of patients with basilar artery occlusions treated with endovascular thrombectomy. Four CT-based scores were assessed: posterior circulation Alberta Stroke Program Early CT Score, basilar artery treatment management score, posterior circulation CT angiography score, and CAPS. The primary outcome of interest for the study was a favorable functional outcome at 3 months (modified Rankin Scale score, 0-3). We calculated sensitivity, specificity, positive and negative predictive values, and generated receiver operating characteristic curves measuring area under the curve (AUC), which were compared with nonparametric methods. RESULTS:=0.23, respectively). CAPS ≤3 demonstrated 100% sensitivity and negative predictive value (95% CI, 92.6-100.0%; 95% CI, 75.7-100.0%, respectively) for good outcome but had low specificity (24.0% [95% CI, 14.3-37.4]). CONCLUSIONS:The CT perfusion-based CAPS had the best performance for predicting good functional outcome at 3 months and had high sensitivity and negative predictive values though our finding is limited by the small number of patients who had CAPS >3. The CAPS should be validated in prospective studies.

BACKGROUND:Assessment of cerebrovascular reactivity (CVR) has been reported using acetazolamide-augmented blood oxygenation level-dependent (BOLD) MRI; however, wide intersubject baseline variability can complicate interpretation. PURPOSE/OBJECTIVE:To develop an integrated normalization framework combining machine learning-guided identification of healthy voxels for subject-specific baseline rescaling with atlas-based assessment. STUDY TYPE/METHODS:Retrospective. POPULATION/METHODS:26 subjects (age 51 ± 13 at first scan, 50% women, 34 exams) with unilateral steno-occlusive disease (SOD) and 9 subjects with bilateral disease (age 55 ± 12, 4/9 women) underwent acetazolamide-augmented BOLD-MRI. FIELD STRENGTH/SEQUENCE/UNASSIGNED:T1-weighted MPRAGE (1 mm isotropic) and gradient-echo EPI at 3 T. ASSESSMENT/RESULTS:Machine learning models (random forest, XGBoost, LightGBM, and neural networks) were trained to identify healthy candidate voxels using a previously reported pipeline based on baseline resting-state BOLD temporal shift features derived from 32 arterial territories. A healthy reference CVR atlas was constructed from unaffected hemispheres. Individual CVR maps were rescaled using predicted healthy voxels and converted to Z-scores and risk index maps using the reference atlas. STATISTICAL TESTS/METHODS:Coefficient of variation (CV) within normal hemispheres, lesion contrast-to-noise ratio (CNR), Lesion ROC AUC, asymmetry index (AI) in unilateral patients, and chi-square distance between thresholded normal and abnormal CVR distributions across all patients were compared between atlas-based group normalization alone and the proposed integrated normalization using paired t-tests. (cutoff p = 0.05). RESULTS:CV within normal hemispheres significantly decreased (1.07 ± 0.15 vs. 0.45 ± 0.09), indicating improved inter-subject stability. Lesion CNR and ROC AUC improved across all impairment thresholds (CNR: 0.73-1.12 vs. 0.79-1.19; AUC: 0.76-0.89 vs. 0.81-0.90). AI increased significantly at higher risk thresholds (≥ 0.78), with no significant difference at the lowest threshold (p = 0.181). Chi-square analysis demonstrated significantly increased separation between CVR distributions at intermediate thresholds. DATA CONCLUSION/CONCLUSIONS:The proposed integrated normalization improves the stability, discriminability, and interpretability of acetazolamide-augmented BOLD-CVR for detection of cerebrovascular impairment. EVIDENCE LEVEL/METHODS:3. TECHNICAL EFFICACY/UNASSIGNED:Stage 2.

Head computed tomography (CT) imaging is a widely used imaging modality with multitudes of medical indications, particularly in assessing pathology of the brain, skull and cerebrovascular system. It is commonly used as the first-line imaging in neurologic emergencies given its rapidity of image acquisition, safety, cost and ubiquity. Deep learning models may facilitate detection of a wide range of diseases. However, the scarcity of high-quality labels and annotations, particularly among less common conditions, substantially hinders the development of powerful models. To address this challenge, we introduce FM-HCT, a Foundation Model for Head CT for generalizable disease detection, trained using self-supervised learning. Our approach pretrains a deep learning model on a large, diverse dataset of 361,663 non-contrast 3D head CT scans without the need for manual annotations, enabling the model to learn robust, generalizable features. Our results demonstrate that the self-supervised foundation model substantially improves performance on downstream diagnostic tasks compared to models trained from scratch and previous 3D CT foundation models trained on scarce annotated datasets.

BACKGROUND AND PURPOSE/OBJECTIVE:CVR values. MATERIALS AND METHODS/METHODS:. 32 distinct time-delays were computed voxelwise, with 32 maximum correlation values constrained to each of 32 paired arterial territories. Performance in prediction of ground-truth reference CVR was computed and compared. RESULTS:The ensembled model achieved AUC of 0.81 in predicting candidate unaffected voxels, demonstrating strong performance in estimation of normal-hemisphere CVR (median absolute percent error [95%CIs] 7.28[3.48-10.34] and 5.61[2.90-9.86] to predict median and mean reference CVR), exhibiting significant improvements over naïve whole-brain voxel selection (P=0.005 and P=0.048, respectively) or conventional cerebellar normalization (26.4[10.1-40.3] median and 27.6[23.7-33.2] mean). In nine bilateral cases assessed to illustrate potential use, the proportion of candidate voxels and corresponding volumes predicted by the ensembled model was significantly lower than in most healthy hemispheres, but yielded subjectively improved delineation of putatively abnormal regions. CONCLUSIONS:We demonstrate feasibility of learning unaffected reference voxel CVR signatures for BOLD-CVR MRI. The approach facilitates extension of brain CVR beyond existing constraints in subjects with bilateral disease.

OBJECTIVES/OBJECTIVE:In mild traumatic brain injury, imaging biomarkers are needed to support clinical management. In four antecedent publications, we used two new (sodium and fingerprinting) and two established (spectroscopy and diffusion) MR techniques in a longitudinally followed patient cohort. Here we report final results and combine all data to determine which marker(s) from the four modalities offer the greatest utility for detecting injury and predicting outcomes. We also leverage the independent specificities offered by each modality to explore injury mechanisms. MATERIALS AND METHODS/METHODS:The longitudinal spectroscopy data were analysed to complete a full data set of proton (spectroscopy, fingerprinting, diffusion) and sodium MRI, acquired alongside symptomatic, cognitive, and functional assessments in 27 patients at 1, 3, and 12 months following injury. Twenty-three matched controls were scanned once. Testing for associations between nine MR markers and three outcome measures was standardized across the entire data set, and performed using Spearman correlations and logistic regression. RESULTS:from fingerprinting (marker of the cellular microenvironment). CONCLUSIONS:We identified independent, dynamic, metabolic and ionic changes, with choline and creatine from spectroscopy fulfilling the most criteria for a clinical biomarker.

BACKGROUND/UNASSIGNED:Durable occlusion after endovascular coiling can be compromised by recanalization, underscoring the need for accurate cerebral aneurysm assessment. Precise volume measurement not only informs treatment decisions and detects subtle aneurysm growth but also refines calculations of packing density, historically linked to occlusion success. This study compares three volume-measurement approaches-traditional two-dimensional (2D) estimation, a semi-automated three-dimensional (3D) technique, and an artificial intelligence (AI)-based 3D method. METHODS/UNASSIGNED:In this retrospective analysis, 24 aneurysms were assessed using 3D rotational angiography. Manual segmentation by three specialists using ITK-SNAP or mimics served as the reference standard. These results were compared with volumes from a semi-automated 3D platform (Philips Advanced Visualization Workspace), an AI-based tool (RapidAI for Aneurysm), and traditional 2D estimations. Agreement with the reference standard was quantified through Passing-Bablok regression slopes and mean biases. RESULTS/UNASSIGNED:(RapidAI). RapidAI demonstrated the strongest correlation with the reference standard, whereas 2D estimations showed the largest discrepancy. The semi-automated 3D method exhibited intermediate accuracy, potentially influenced by the clinician input required for segmentation. CONCLUSION/UNASSIGNED:All methods underestimated aneurysm volumes compared to the reference standard, suggesting that inaccurate volume measurements may mask early aneurysm growth. Among the techniques assessed, the AI-based approach provided the closest agreement with the reference, indicating that improved volumetric methods-particularly AI-driven ones-can enhance early detection of aneurysm expansion, guide treatment decisions, and help establish more reliable follow-up strategies for both treated and conservatively managed aneurysms.

BACKGROUND AND PURPOSE/OBJECTIVE:Cerebrovascular reactivity (CVR) is a widely studied biomarker of cerebral hemodynamics, commonly used in risk stratification and treatment planning in patients with steno-occlusive disease (SOD). Conventional use relies on normalization of estimates to contralateral hemisphere reference values, which is unsuitable for bilateral or indeterminate distributions of disease. We report upon a custom data-driven approach leveraging random forest classifiers (RFc) to identify candidate voxels for normalization in order to facilitate interrogation outside conditions of known unilateral SOD MATERIALS AND METHODS: We retrospectively analyzed 16 patients with unilateral SOD who underwent acetazolamide-augmented BOLD-MRI and DSC perfusion. Three RFc models were trained using leave-one-out cross-validation (LOOCV) to identify candidate voxels brain-wide whose CVR were within 10% of the normal hemispheric median: i. all voxels; ii. gray matter only; and iii. white matter only. Model input features included time-to-maximum (Tmax), mean transit time (MTT), cerebral blood flow (CBF), and cerebral blood volume (CBV) from contemporaneous DSC. The median model-predicted reference CVR (CVRref) was compared to ground-truth medians in LOOCV, and its impact on threshold-based volumetric classification of CVR reduction assessed. RESULTS:RFc models effectively predicted ground-truth CVR voxels, achieving median absolute percent differences of 12.8% (IQR: 5.0%-18.9%) using all voxels, 11.3% (IQR: 9.3%-16.1%) for gray matter, and 9.8% (IQR: 4.4%-16.9%) for white matter. Volumetric estimates of CVR reduction across thresholds for the models revealed excellent agreement between ground-truth and model estimates without statistically significant differences (p>0.01), excepting lowest white matter CVR thresholds. Model use in a small pilot deployment of bilateral SOD cases demonstrated the potential utility, enabling voxel-wise CVR assessment without reliance on contralateral reference. CONCLUSIONS:We present a novel data-driven approach for normalizing CVR maps in patients with bilateral or indeterminate SOD. Using an RFc, our method provides an individualized, brain-wide reference CVR, expanding the utility of CVR estimates beyond the typical constraints of unilateral disease, and with potential application to other, similarly constrained scenarios such as for SPECT or PET hemodynamic studies. ABBREVIATIONS/BACKGROUND:CVR = cerebrovascular reactivity; RFc = random forest classifier; SOD = steno-occlusive disease.

BACKGROUND:Early prediction of hematoma expansion (HE) following nontraumatic intracerebral hemorrhage (ICH) may inform preemptive therapeutic interventions. We sought to identify how accurately machine learning (ML) radiomics models predict HE compared with expert clinicians using head computed tomography (HCT). METHODS:We used data from 900 study participants with ICH enrolled in the Antihypertensive Treatment of Acute Cerebral Hemorrhage 2 Study. ML models were developed using baseline HCT images, as well as admission clinical data in a training cohort (n = 621), and their performance was evaluated in an independent test cohort (n = 279) to predict HE (defined as HE by 33% or > 6 mL at 24 h). We simultaneously surveyed expert clinicians and asked them to predict HE using the same initial HCT images and clinical data. Area under the receiver operating characteristic curve (AUC) were compared between clinician predictions, ML models using radiomic data only (a random forest classifier and a deep learning imaging model) and ML models using both radiomic and clinical data (three random forest classifier models using different feature combinations). Kappa values comparing interrater reliability among expert clinicians were calculated. The best performing model was compared with clinical predication. RESULTS:The AUC for expert clinician prediction of HE was 0.591, with a kappa of 0.156 for interrater variability, compared with ML models using radiomic data only (a deep learning model using image input, AUC 0.680) and using both radiomic and clinical data (a random forest model, AUC 0.677). The intraclass correlation coefficient for clinical judgment and the best performing ML model was 0.47 (95% confidence interval 0.23-0.75). CONCLUSIONS:We introduced supervised ML algorithms demonstrating that HE prediction may outperform practicing clinicians. Despite overall moderate AUCs, our results set a new relative benchmark for performance in these tasks that even expert clinicians find challenging. These results emphasize the need for continued improvements and further enhanced clinical decision support to optimally manage patients with ICH.