Logo image
Back
Abstract A028: A Robust and Reproducible Pipeline for Quantifying Stroke Injury in a Heterogeneous, Multi-Site Preclinical Network
Abstract   Peer reviewed

Abstract A028: A Robust and Reproducible Pipeline for Quantifying Stroke Injury in a Heterogeneous, Multi-Site Preclinical Network

Kirsten Lynch, Ryan Cabeen, Joseph Mandeville, Basavaraju Sanganahalli, Fahmeed Hyder, Daniel Thedens, Ali Arbab, Shuning Huang, Adnan Bibic, Erendiz Tarakci, …
Stroke (1970), Vol.57(Suppl_1), A028
02/2026
DOI: 10.1161/str.57.suppl_1.A028

View Online

Abstract

Introduction: The failure to translate promising preclinical stroke therapies into clinical success is largely attributed to a lack of rigorous, reproducible outcome measures. While magnetic resonance imaging (MRI) offers a translational alternative to traditional histology, its use in large, multi-site trials is challenged by data heterogeneity and the need for scalable analysis. To address this, we developed and validated a fully automated, open-source image analysis pipeline for the Stroke Preclinical Assessment Network (SPAN), a six-center preclinical trial. Methods: T2 and ADC MRI scans were acquired from 2443 mice and rats (including aged and obese cohorts) at 6 centers 2 and 30 days after middle cerebral artery occlusion (MCAO). Our open-source pipeline performs a complete workflow (Figure 1): 1) preprocessing (image reconstruction, denoising, parameter estimation and quality assessment measures); 2) intensity harmonization to reduce inter-site variability; 3) brain extraction using either traditional rule-based segmentation (rats) or U-net deep learning model (mice); 4) rule-based lesion segmentation via thresholding of harmonized T2 and ADC maps (Figure 2A-G); and 5) quantification of midline shift as a proxy for swelling and atrophy (Figure 3A). Validation was performed against expert manual tracing on both MRI and 2,3,5-Triphenyltetrazolium chloride (TTC)-stained tissue. Results: The pipeline successfully processed thousands of scans from a heterogeneous collection of scanners. The U-net brain extraction model was highly accurate (Dice score=0.964) and successfully segmented cases where traditional methods failed. Automated lesion volumes correlated strongly with manual expert MRI tracing (R=0.957) and with TTC staining in optimal preparations (R=0.86; Figure 2H). Harmonization significantly reduced site-specific differences in MRI values. Our geometric midline estimation consistently demonstrated a midline shift towards the contralesional side indicative of swelling on Day 2 and a midline shift towards the ipsilesional side indicative of atrophy on Day 30 (Figure 3B). Conclusion: We have developed a validated, end-to-end automated pipeline for quantifying stroke injury in large, multi-site preclinical trials. This work delivers a scalable, objective, and reproducible framework as a shareable, open-source tool that enhances the rigor of preclinical research to help bridge the translational gap in stroke.
 Stroke Computational modeling Imaging

Details

Metrics

3 Record Views
Logo image