DSA-NRP: No-Reflow Prediction from Angiographic Perfusion Dynamics in Stroke EVT

Technical Summary: DSA-NRP

Problem Statement
Following successful endovascular thrombectomy (EVT) for acute ischemic stroke (AIS), a significant complication known as "no-reflow" can occur. This phenomenon is defined by persistent microvascular hypoperfusion despite the restoration of large-vessel patency (macrovascular recanalization), leading to poor tissue recovery and worse clinical outcomes. Currently, the identification of no-reflow relies on follow-up perfusion magnetic resonance imaging (MRI) performed within 24 hours post-procedure. This delay prevents immediate intervention. While digital subtraction angiography (DSA) is routinely performed during EVT, its potential to predict no-reflow in real-time using intra-procedural perfusion dynamics has remained unexplored.

Methodology
The authors propose DSA-NRP, a machine learning (ML) framework designed to predict no-reflow immediately after EVT by integrating clinical variables with intra-procedural DSA sequences.

• Dataset: The study utilized a retrospective cohort of 39 AIS patients treated at UCLA Medical Center (2011–2024) with large-vessel occlusions (M1 or ICA) who achieved successful recanalization (mTICI scores of 2c or 3). All patients had paired pre- and post-procedure MRI scans.
• Ground Truth Labeling: No-reflow was defined quantitatively as a >15% reduction in relative cerebral blood volume (rCBV) or flow (rCBF) within the infarct core compared to the contralateral hemisphere, derived from post-procedure MRI perfusion maps. Infarct regions of interest (ROIs) were generated using a U-Net architecture trained on BRATS21 and ISLES 2022 datasets.
• DSA Pre-processing:
  • DSA sequences (anteroposterior and lateral views) were manually annotated to identify the Target Downstream Territory (TDT).
  • A pixel-wise minimum intensity projection (minProjection) was used to delineate the TDT.
  • Mean pixel intensity within the TDT was computed for each frame, converting 3D spatio-temporal data into 1D time-series signals.
  • Signals were inverted so that brighter pixels represented higher tracer concentrations, synchronized to contrast onset, and truncated to the first 5 seconds (arterial/capillary phases).
• Feature Extraction: Three categories of features were extracted from the 1D time-series signals:
  1. PEAK: Dynamics of the perfusion peak (height, width, slope, ratio).
  2. SIPS (Statistical Intensity Profiling): Statistical properties including mean, standard deviation, skewness, kurtosis, and minimum intensity.
  3. FLOW: Temporal descriptors including time-to-peak, decay time, plateau duration, and signal correlation.
  • Additionally, RAW features consisted of fixed-length (15-frame) 1D time-series vectors.
  • A Clinical (CLN) baseline included age, sex, and final mTICI score.
• Model Training: Due to the small cohort size, a Leave-One-Out Cross-Validation (LOOCV) strategy was employed. A Gradient Boosting Classifier was selected as the primary model. Ablation studies were conducted to evaluate individual feature sets (PEAK, SIPS, FLOW, RAW) and their combinations against the clinical baseline.

Key Contributions

1. First ML Framework for Intra-procedural Prediction: This work presents the first machine learning framework to predict post-thrombectomy no-reflow using structured, interpretable features derived from intra-procedural DSA sequences.
2. Superiority of DSA-Derived Features: The study demonstrates that DSA-derived features (specifically statistical and time-series descriptors) significantly outperform traditional clinical baselines in predicting no-reflow.
3. Physiological Interpretability: Through feature importance analysis, the authors identify physiologically meaningful predictors, such as peak tracer intensity, temporal variance, and flow decay, linking them to downstream perfusion failure and microvascular dysfunction.

Results

• Performance: The combined ensemble model (CLN + TSF + RAW) achieved an AUROC of 0.9330 (p = 0.006) and an accuracy of 0.9231. This significantly outperformed the clinical baseline (AUROC: 0.7768).
• Feature Analysis:
  • Models using only DSA-derived features (RAW or TSF) achieved high performance (AUROC ~0.90–0.91), suggesting that baseline clinical variables contribute limited additional information beyond the angiographic data.
  • The FLOW and SIPS feature sets were the most predictive among statistical features, particularly when derived from post-procedure signals.
  • Feature importance analysis revealed that early arterial-phase dynamics (RAW features) and post-peak washout characteristics (plateau duration, skewness) were the dominant predictors.
• Subgroup Analysis: The model maintained consistent performance across subgroups stratified by mTICI score (2c vs. 3), sex, race, age, and baseline stroke severity (NIHSS), with no statistically significant differences in model behavior (all p > 0.1).
• Recall: The best-performing models achieved perfect recall (1.0000), identifying 100% of no-reflow cases in this cohort.

Significance and Claims
The paper claims that DSA-NRP establishes a preliminary foundation for immediate, accurate no-reflow prediction without reliance on delayed imaging. By leveraging intra-procedural DSA perfusion dynamics, the approach enables clinicians to proactively manage high-risk patients through targeted interventions (e.g., neurocritical care, blood pressure management, or intra-arterial thrombolysis) while the patient is still in the angiography suite.

The authors emphasize that while the results are promising, they are preliminary due to the small sample size (n=39) and single-center nature of the study. They explicitly state that the findings require validation in larger, independent, multi-center cohorts before clinical deployment. The work serves as a proof-of-concept that microvascular integrity information is encoded in standard DSA sequences and can be extracted via machine learning to improve stroke care workflows.