A new AI model for reading lower back MRIs is designed to stop radiologists from missing findings that only show up in one scan angle.
Researchers built a system called GateSPINE that generates written reports from lumbar spine MRI scans. Unlike earlier tools that treat an entire MRI exam as one blended volume, GateSPINE keeps scan types distinct at first - it fuses sagittal T1 and T2 images with a training-free operator, then runs that combined sagittal view alongside the separate axial view through two parallel 3D encoders. A gated cross-view fusion module then decides, channel by channel and token by token, which view to trust most before writing the final report. The team tested it on three lumbar MRI datasets, two public benchmarks plus a private set from Phenikaa University Hospital, and it posted the best clinical efficacy F1 score of the group by catching more real abnormalities.
MRI reports are harder to automate than CT ones because a single study can include multiple sequences and imaging planes, each showing something the others do not. Tools that squash all of that into one volume tend to dilute or bury findings that only appear in, say, the axial view - exactly the kind of miss that is most costly in a clinical setting. GateSPINE's bet is that letting the model weigh each view's reliability per feature, instead of blending everything upfront by fixed rule, recovers some of that lost signal.
It is still a research benchmark result, not a hospital rollout - and the authors note the gated fusion component itself was only validated on two of the three datasets, since the third lacked axial scans to fuse in the first place.