USC researchers create automated MRI system to advance stroke treatment research

A new MRI-based system could change how experimental stroke treatments are evaluated, allowing researchers to compare brain injury across thousands of laboratory animals with far greater consistency than traditional methods. Developed by scientists at the Mark and Mary Stevens Neuroimaging and Informatics Institute at the Keck School of Medicine of USC, the open-source pipeline automatically processes brain scans collected across multiple research centers, reducing the influence of different scanners, imaging settings, and human judgments.

The technology was created for the National Institutes of Health-sponsored Stroke Preclinical Assessment Network, or SPAN, a large collaborative effort designed to test potential therapies for acute ischemic stroke before they move toward clinical trials. In a study published in Imaging Neuroscience, the researchers evaluated the system using MRI data from 2,442 mice and rats scanned at six academic research centers. The scale of the dataset makes the work one of the most extensive demonstrations of automated, harmonized MRI analysis in preclinical stroke research.

Ischemic stroke occurs when a blood clot or other obstruction blocks blood flow to part of the brain, depriving tissue of oxygen. Although numerous treatments have appeared promising in animal experiments, only a small proportion have ultimately benefited patients. One reason is that preclinical studies can vary substantially in how strokes are induced, how animals are monitored, how images are collected, and how damage is measured. Even small differences in methodology can make it difficult to determine whether an apparent treatment effect is real or simply the result of inconsistent analysis.

For decades, researchers have often assessed experimental stroke damage by removing an animal’s brain, slicing it into sections, applying chemical stains, and manually tracing the injured regions. Histological analysis can provide valuable information, but it may distort tissue and generally captures only one time point. Manual measurements can also vary between observers, particularly when the boundaries between healthy, swollen, and irreversibly damaged tissue are unclear. MRI offers a noninvasive alternative, enabling scientists to scan the same animal repeatedly and observe how injury develops from the early stages of swelling to later tissue loss.

The challenge is that large multicenter studies generate immense numbers of images under different technical conditions. MRI scanners may operate at different magnetic field strengths, use different acquisition protocols, or produce images with varying contrast and quality. The new pipeline addresses these sources of variation through a sequence of automated processing steps. It first evaluates image quality, then harmonizes scans to reduce scanner-related differences, identifies the brain, and measures several outcomes associated with stroke, including injured tissue, swelling, displacement, and longer-term loss of brain volume.

A key component is a deep-learning model trained to distinguish the brain from surrounding bone, muscle, and other tissues. This process, known as brain extraction or skull stripping, is essential because inaccurate identification of the brain can distort every measurement that follows. However, the researchers did not rely entirely on an opaque artificial-intelligence system to define stroke damage. After the brain was isolated, the pipeline used transparent, rule-based image-processing methods to identify and quantify abnormal tissue. This combination was intended to provide the speed and resilience of machine learning while preserving enough interpretability for scientists to understand how each result was produced.

The automated measurements closely matched injury outlines made by imaging specialists. According to the study, the pipeline performed about as consistently relative to a human reviewer as two human reviewers performed relative to one another. That comparison is important because agreement between experts is often treated as a practical benchmark for evaluating automated medical-image analysis. The software also processed the vast majority of scans despite differences in species, scanners, magnetic field strengths, acquisition environments, and image quality, demonstrating that the method was designed for real-world network conditions rather than a single highly controlled laboratory.

The system could be particularly valuable for detecting treatment effects that are too subtle or inconsistent to identify reliably through conventional methods. By applying the same computational rules to every scan, researchers can reduce the possibility that one laboratory will report a different result simply because it uses another scanner or a different style of manual analysis. Repeated MRI examinations may also reveal when a therapy changes the progression of injury rather than merely altering the final amount of tissue loss. That temporal information could help scientists distinguish treatments that protect brain tissue early from those that influence recovery later.

The researchers have released the MRI data and processing software publicly, creating a resource that other laboratories can inspect, reproduce, and adapt. The pipeline was developed for standardized animal models and cannot directly solve the much greater biological complexity of human stroke, where injuries differ widely in location, timing, severity, and cause. Nevertheless, its framework could be expanded to incorporate additional MRI techniques and other measures of brain outcome. By combining automated quality control, image harmonization, deep learning, transparent analysis rules, and high-performance computing, the SPAN platform offers a reproducible foundation for deciding which experimental stroke treatments deserve the next step toward clinical testing.

News Publication Date: 7-Aug-2026

Web References: https://ini.usc.edu/ ; https://github.com/cabeen/span-mri ; https://direct.mit.edu/imag/article/doi/10.1162/IMAG.a.1328/137726

References: Imaging Neuroscience, DOI: 10.1162/IMAG.a.1328