Researchers at the Keck School of Medicine of USC have developed an open-source, AI-assisted MRI analysis pipeline capable of processing thousands of preclinical stroke scans with expert-level accuracy, addressing a critical reproducibility challenge in the translation of experimental stroke therapies towards clinical trials.
Researchers at the Mark and Mary Stevens Neuroimaging and Informatics Institute at the Keck School of Medicine of USC have developed an automated system that uses MRI scans to measure brain tissue damage in animal models of stroke.
The open-source pipeline, described in a study published in Imaging Neuroscience, was tested on scans from 2,442 mice and rats across six academic research centres. Its measurements closely matched those made by human imaging experts while reducing variation caused by differences between MRI scanners and imaging environments.
The system was developed for the National Institutes of Health-sponsored Stroke Preclinical Assessment Network (SPAN), which evaluates potential treatments for acute ischaemic stroke before the most promising candidates progress towards clinical trials.
“Before a potential treatment can be tested in people, researchers need confidence that its effects have been measured rigorously and consistently,” said Kirsten Lynch, Assistant Professor of Research Neurology and a co-first author of the study. “This pipeline gives us an objective and scalable way to assess brain injury across a large research network.”
Tackling a major challenge in stroke research
Ischaemic strokes are caused by blocked blood vessels that stop blood and oxygen from reaching part of the brain. While numerous experimental treatments have shown promise in laboratory studies, relatively few have translated into effective therapies for patients.
One challenge is the variability between preclinical studies. Researchers have traditionally assessed stroke damage by removing an animal’s brain, staining tissue sections and manually outlining injured areas. The process can distort tissue and relies partly on individual judgement.
MRI offers a less invasive alternative, allowing researchers to scan the same animal repeatedly and monitor changes over time. However, analysing thousands of scans collected using different equipment presents its own difficulties.
“The scale of SPAN made automation essential,” said Ryan Cabeen, a computational scientist at the Stevens INI who led development of the imaging biomarker platform and who was co-first author of the study. “We needed a method that could process thousands of scans while applying the same rules to every image, regardless of where the data were collected.”
Combining AI with transparent analysis
The pipeline performs several stages of image analysis, including checking image quality, correcting for scanner differences, identifying the brain and measuring stroke-related changes such as tissue injury, swelling, displacement and longer-term tissue loss.
A deep-learning model is used to distinguish the brain from surrounding bone, muscle and other tissue. However, the researchers deliberately avoided making the entire process dependent on an artificial intelligence system.
Instead, transparent, rule-based image-processing methods are used to identify and measure stroke damage.
“We wanted researchers to understand how the results were produced,” Lynch said. “A method can be highly automated without becoming a black box. The combination of deep learning and transparent image-processing rules gave us both robustness and interpretability.”
Matching human experts
More than 2,200 animals underwent MRI scans two days after an experimentally induced stroke, with 1,750 receiving a second scan around one month later.
When the automated measurements were compared with damage manually outlined by imaging experts, the results showed extremely close agreement. The pipeline performed at approximately the same level of consistency as two human reviewers assessing the same scans.
It also reduced variation associated with different scanners and imaging environments, allowing researchers at all six centres to use the same analysis approach.
The software processed the vast majority of scans despite differences in animal species, equipment, magnetic field strength and image quality.
An open tool for future studies
The researchers have made both the software and MRI data publicly available, allowing other groups to reproduce the findings and adapt the pipeline for future preclinical studies.
The system was designed for standardised animal models and is not intended to analyse the highly variable stroke injuries found in patients. However, its framework could potentially be expanded to incorporate additional imaging techniques and measures of brain tissue outcomes.
“Large, collaborative studies require tools that produce reliable results across institutions,” said Arthur Toga, Director of the Stevens INI and a co-author of the study. “By combining advanced imaging, artificial intelligence, data harmonisation, and high-performance computing, this work provides a reproducible foundation for evaluating which experimental stroke treatments have the greatest potential to move toward clinical testing.”
By reducing differences in how preclinical brain injuries are measured, the system could help researchers identify the most promising stroke treatments earlier and provide more consistent evidence before candidates enter human trials.
No comments yet