MIT-led team builds head symmetry into a network to track a moving fetus in MRI

Researchers at MIT, Boston Children's Hospital and Inria, with MIT PhD student Ramya Muthukrishnan as first author and Elfar Adalsteinsson, Ellen Grant and Polina Golland as senior authors, have described E(3)-Pose, a method for estimating where a fetus's head is pointing during an MRI scan, with the left-right symmetry of the head built into the neural network. MIT News reported the work on 19 October 2026; the paper, "Equivariant symmetry-aware head pose estimation for fetal MRI", was presented at the European Conference on Computer Vision (ECCV 2026), and the authors report lower average errors than six existing methods on their clinical-style test sets in most comparisons.

The problem it targets

Fetal brain MRI is built from stacks of 2D slices. The paper says that, because of fetal motion between slices, "oblique slices and spatial coverage gaps in the acquired stack present challenges for radiological assessment and often necessitate re-acquisition of the entire stack".

The approach is to take a rapid, low-resolution 3D "navigator" scan before each slice, estimate the head's pose from it, and tilt the next slice to match. That needs an estimate in under a second, because "the time interval between every navigator and the next slice is 1s". E(3)-Pose runs in 0.3 seconds; the FireANTs registration tool takes 1.8 seconds, which the authors say "prohibits clinical deployment".

What "equivariance" means here

The paper defines it this way: "A function is equivariant under a group of transforms if it commutes with every transform in the group." Put plainly, if you rotate the input scan, the network's answer rotates by the same amount, because the network is built that way rather than having to learn it from rotated training examples.

The second ingredient is symmetry. At low resolution a fetal head looks almost the same mirrored left to right, which the paper says introduces "pose ambiguities". E(3)-Pose represents the left-right axis with a "pseudovector", a quantity that does not change under that mirror flip, so the network is not penalised for the ambiguity. The authors say they are "the first to use them to address object symmetries"; we have not checked that claim.

How it was tested

• Training data: separate models on a Boston Children's research set (153 pregnant volunteers) and the public dHCP set (245).
• Test data: Clinical-Young, 60 clinical scans of fetuses at 18 to 23 weeks; and Navigators, "47 time-series (1210 total volumes)" from 9 volunteers, which the paper calls the data closest to the intended use.
• Comparison: the main results table has 12 method rows (two untrained methods, plus five trained methods per training set) and eight number columns.

On Navigators, the model trained on the research set averaged a 9.4° rotation error; the next best was RbR at 22.6°, and the others ranged from 38.8° to 65.0°. Trained on dHCP, E(3)-Pose averaged 13.9°, against 59.0° to 93.9° for the other trained methods. On Clinical-Young it averaged 9.1°, ahead of Fetal-Align (10.3°) and RbR (11.2°).

It did not win everywhere. Trained on dHCP, it averaged 12.1° on Clinical-Young, behind Fetal-Align (10.3°), which needs no pose training; the paper ranks methods within each training setting. On the research test set Fetal-Align matched or beat it (5.0°, against 5.1° and 5.7° for the two E(3)-Pose models). On the high-quality dHCP test set, FireANTs (0.4°), and EquiTrack trained on the research set (7.1° against 7.4°), were ahead. The authors attribute that to "the (unfair) information advantage of a subject-specific template already aligned to the GT pose". An appendix adds that the design "does not improve performance on in-distribution, high-SNR volumes".

Limitations the authors state

• The Navigators analysis "is limited to 9 pregnant participants who were recruited at the same institution and imaged on the same scanner", and it lacks "sufficient representation of second-trimester fetuses".
• The symmetry assumption "may not hold" under "large lateralized pathologies". In simulation it held for lesions comparable to ventriculomegaly but "breaks down" for tumour-sized asymmetries of 40 to 70mm; real pathological cases are left "to future work".
• It "does not explicitly model motion history or pose uncertainty".
• The slice-tilting benefit was shown "in our simulations" (168 simulated stacks), not in patients.

Is it in clinical use?

The paper does not describe clinical use; it frames its results as "supporting the potential for future clinical deployment". The authors say they ran the system on a 3T Siemens scanner and "have since used the full pose returned by E(3)-Pose to automatically prescribe axial, sagittal, and coronal slice stacks in utero", but that "Future work will evaluate the diagnostic potential of prescribed slice stacks on a larger cohort of pregnant volunteers". The paper's scope "is the technical development of our method". MIT News quotes co-author Ellen Grant as saying the work "will automate fetal brain MRI"; that is a forecast, not a reported result.

Code and weights

The code is on GitHub (MedicalVisionGroup/E3-Pose), and the README links trained weights on Google Drive. The GitHub API reports no licence for the repository, and the README states none for the weights.

Why it matters

On our reading, the interest is the design choice: building known physical properties into a network instead of learning them. In this paper, that made the biggest difference where training and real-world data differed most. With the closest-to-clinical test covering nine volunteers at one hospital, the planned larger cohort is the next test.