Short-axis cardiac cine-MRI hangs the ascending and descending aorta as two circular cross-sections. This method writes frame-wise binary overlays that hold through the cardiac cycle, so you can still read aortic distensibility and arterial-stiffness markers when systolic flow washes the vessel walls out. Dexter Wen Jie Teo (NTU Singapore and Polytechnique Montreal), Nairouz Shehata, and Herve Lombaert submitted arXiv:2608.23879 on 24 August 2026. The paper is Spatiotemporal Distillation via Recurrent Bottlenecks for Aortic Tracking, accepted to the STACOM workshop at MICCAI 2026. Code is public at github.com/dexterteo4/aortic-temporal-distillation.
Plain 2D nets score each frame alone. Systolic acceleration fades boundaries. Start and end frames sit in low-contrast fade-out. You get holes, fragments, a missed ascending aorta, or a grab on an adjacent brighter vessel. Contours flicker on retrospective review, and the area curve for distensibility is garbage.
Teacher on end-diastole, then scrub the topology
Phase I trains a spatial teacher (attention U-Net style) on expert end-diastole labels for 198 patients inside the distillation pool. Loss is Dice plus λ Focal with λ = 0.50. Checkpoint selection clears DSC ≥ 0.50, then picks the epoch that maximizes Frac2CC (fraction of frames with exactly two connected components).
That teacher writes soft maps on the unlabeled pool. Softmaps are binarized at τbin = 0.50, then scrubbed offline with a three-pass topological sanitizer under 8-connectivity. Keep N = 2. If N > 2, prune to the two largest blobs. If N < 2, circular-stitch from the nearest valid anchor: copy the missing blob when one side survives, or copy the whole mask when the frame is empty. The scrub runs once before distillation. It does not run at test time.
BiConvLSTM at the bottleneck, residual α = 0.10
Phase II warm-starts a student with the same 2D attention U-Net. A batch of T frames is folded through the 2D backbone. A dynamic hook at the deepest bottleneck unfolds features to B × T × C × H × W, projects to Ctemp = 128 with a 1 × 1, runs a bidirectional ConvLSTM, and residual-adds the temporal correction with α = 0.10 (temporal branch zero-initialized). Spatial capacity stays with the teacher weights. Sequence context rides the bottleneck.
Student loss is Lspatial (Dice:BCE = 1.0:0.5, center frame weight 5, others 1) plus β Lsmooth with β = 0.10. Lsmooth is the mean L1 of the discrete second derivative of soft probabilities. Uniform motion costs nothing. Flicker and edge stutter get penalized. Sliding window is T = 11, stride 1, circular wrap. Inputs are 320 × 320. AdamW for 70 epochs, ηenc = 1 × 10-5, ηtemp = 1 × 10-4, BN frozen in eval mode.
β peaks at 0.10 on the population audit. β = 0 over-errors. β = 0.50 induces tracking lag.
437 cine studies, 25-patient held-out test
Data is 437 cardiac cine-MRI scans on a 1.5T Siemens Magnetom Aera (tfi3D-fs, 2.0 mm pixel spacing, 1.6 mm slice thickness, Z = 1). Patient-level split: 412 distillation (DD) / 25 untouched test (DT), stratified by age, BSA, and sex. Within DD, 5-fold CV. Teacher DL is the 198 ED-annotated patients. Test DT is 948 continuous frames, nnU-Net-initialized then manually edited and audited. Population audit pool is 17,539 unannotated frames.
On the 25-patient / 948-frame test (mean ± std over 5 folds), the spatiotemporal student reaches Frac2CC 99.2% ± 0.6 and NSD@1mm 92.3% ± 0.2 (DSC 90.9% ± 0.2). Static student is Frac2CC 98.7% ± 0.7. Static 2D teacher is Frac2CC 93.0% ± 4.4. A 2D nnU-Net baseline still wins raw DSC (92.9% ± 1.1), HD95, and ASSD. Their win is topology and surface NSD@1mm. Across the 17,539-frame audit the paper reports more than a 56% cut in population-wide structural anomalies versus that nnU-Net baseline (451.2 vs 1,036.8 anomalies per fold).
It fails when the short-axis view no longer shows exactly two components: branching near the aortic arch, or pathology that splits the vessel. The bidirectional T = 11 window also makes this retrospective only, not real-time streaming.
How this lands in a viewer
If you already hang cine in a DICOM viewer, treat this as a short-axis aorta overlay for retrospective review. Load the cine, hang ascending and descending contours, scrub through systole, and keep the area curve for distensibility. Do not ship it as a live streaming contour. Branching arch slices and split-vessel pathology should fail closed or hand off to a human edit. Frac2CC 99.2% on 948 audited frames is a tracking claim, not a device claim.
MedPixel is phrase or loose box on a hung slice. CoInS-Net is two endpoints and an in-between plane. SAT3D is tumour plus uncertainty in Slicer. This paper is short-axis cine in, stable ascending and descending aorta overlays out.
Rebuild from the public repo and arXiv:2608.23879. Annotations credit Amr Elsawy and Mohamed Nagy at the Aswan Heart Centre, Egypt. The preprint plus STACOM acceptance is the source of record until a camera-ready venue version exists.
Sources
- Teo, D.W.J., Shehata, N., Lombaert, H. Spatiotemporal Distillation via Recurrent Bottlenecks for Aortic Tracking. arXiv:2608.23879. Submitted 24 August 2026. Accepted STACOM workshop, MICCAI 2026. https://arxiv.org/abs/2608.23879. Code: github.com/dexterteo4/aortic-temporal-distillation
- Oktay, O., et al. Attention U-Net: Learning Where to Look for the Pancreas. MIDL 2018.
- Isensee, F., et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods 2021. DOI: 10.1038/s41592-020-01008-z
- Bai, W., et al. Recurrent neural networks for aortic image sequence segmentation with sparse annotations. MICCAI 2018.
- Cecelja, M., et al. Aortic distensibility measured by automated analysis of magnetic resonance imaging predicts adverse cardiovascular events in UK Biobank. J. Am. Heart Assoc. 2022.
- Storey, P., et al. Flow artifacts in steady-state free precession cine imaging. Magn. Reson. Med. 2004.