CAPE TOWN — A fully autonomous MRI exam showed success in fetal imaging, according to a prospective study shown at the International Society for Magnetic Resonance in Medicine (ISMRM) annual meeting.
The AI-powered self-guided exam pipeline proved successful by showing performance comparable to that of specialists, which could be useful in low-resource settings, said Sara Silva, PhD, from King’s College London in her presentation.
“You can get automatic planning of brain scans, which are very useful for radiological assessment,” Silva told AuntMinnie. “You can get analysis in terms of quantitative sequences and motion correction techniques. We can do all of that in a single exam … with very little operator dependency.”
While MRI can provide valuable insights into fetal anatomy and physiology, its usage is limited due to several factors: fetal motion, variable fetal positioning, operator-dependency for planning, and the need for offline post-processing tools.
AI approaches could help in this area, Silva said, by offering seamless fetal imaging frameworks. She and colleagues in their study explored how an AI pipeline could eliminate manual planning and retrospective post-processing.
They used the framework for image reconstruction environments (FIRE) for real-time exporting and processing of acquired images, using multiple AI pipelines. The pipelines included the following tasks: automatic planning of radiological brain planes, detection and reacquisition of motion-corrupted diffusion MRI, reconstruction of a motion-free fetal brain 3D volume, and volumetric and biometric automated reporting.
Silva noted that all neural networks were trained and validated on over 400 fetal MRI datasets across multiple field strengths, gestational ages, and pathologies.
Sara Silva, PhD, explains the fully automated MRI pipeline that she and colleagues found could have utility in fetal imaging in low-resource areas.
Analysis included a prospective cohort of 54 fetuses with a gestational age range of 25 to 39 weeks who were imaged on a 0.55T scanner. The team performed manual reevaluation of each automated component. And AI-guided planning validation included a subset of 22 fetal MRI datasets, also using 0.55T scanning, and compared manual versus self-prescribed scans.
Silva said the self-guided pipeline was successfully applied in 18 of the fetuses. The four incomplete scans were due to study-specific protocols that were implemented, she added.
Compared to specialists, the self-guided pipeline showed non-inferior results. On a five-point scale, self-planning in the prospective cohort achieved a score of 2.53 compared to 2.6 for expert manual plane prescription. The pipeline also achieved a diagnostic image quality of 3.77 (out of 5).
Brain volumetry and biometry from the self-guided pipeline scored 3.18 and 3.27 (both out of 5), respectively. Silva also reported an average deviation of 3.42 across nine biometric measures.
The pipeline also showed promise in the segmentation of internal uterine structures, as well as brain and placenta T2 values.
|
Performance of fully automated fetal MRI scanning pipeline |
|
|
Measure |
Dice score |
|
Segmentation of internal uterine structures |
0.98 |
|
Average difference in brain T2 value (compared to manual assessment) |
0.96 |
|
Average difference in placenta T2 value (compared to manual assessment) |
0.82 |
Diffusion brain quality control meanwhile achieved a sensitivity of 96.1% and a specificity of 95.7%. Additionally, the complete acquisition lasts about 25 minutes, with processing being around 14 minutes, Silva said.
Meanwhile, six automatic biometric measurements in the comparison cohort showed low deviation from manual measurements. These measurements were performed on automatically and manually planned real-time, 3D brain slice-to-volume reconstruction.
Silva said the pipeline led to three immediate benefits: streamlined acquisition and real-time processing workflows; improved accessibility and operator dependence; and immediate feedback that allows adaptive scanning.
“We want to extend it [pipeline] a bit more by integrating more pathological cases, ensuring that the pipeline also works in those cases,” Silva told AuntMinnie. “Also, we want to disseminate it globally.”
Check out AuntMinnie’s ShowCast for full coverage of ISMRM 2026.