Source-linked AI summary
Spatiotemporal Distillation via Recurrent Bottlenecks for Aortic Tracking
Dexter Wen Jie Teo, Nairouz Shehata, Herve Lombaert
TL;DR
Independent 2D cine-MRI segmentation can lose aorta boundaries across cardiac cycles, while expert annotations for spatiotemporal training are scarce. The paper distills a corrected spatial teacher into a recurrent spatiotemporal student with latent interception and smoothness regularization. The student reduces structural anomalies relative to 2D baselines while retaining high surface-tracking accuracy.
Problem
Independent-frame 2D segmentation suffers temporal tracking dropouts and boundary inconsistency when systolic flow obscures aortic boundaries, while spatiotemporal models face scarce expert annotations.
Method
A semi-supervised framework corrects spatial-teacher pseudo-labels and distills them into a spatiotemporal student using a recurrent latent bottleneck and temporal smoothness objective.
Results
451.2±67.8 structural anomalies per fold were achieved by the spatiotemporal student, compared with 1,036.8 for the nnU-Net baseline, while NSD@1mm reached 92.3% ± 0.2%.
Takeaways & Limitations
The framework completely eliminates frame-to-frame tracking dropouts while maintaining high surface accuracy across 17,539 frames.
Takeaways & Limitations
The correction assumes two isolated circular cross-sections and bidirectional ConvLSTM buffering of an 11-frame window, limiting complex-topology and real-time use.
Abstract
from arXiv · showhide
Cardiac cine-MRI serves as a direct visual indicator of cardiovascular hemodynamics by capturing the continuous wall motion of the aorta. Quantifying these dynamic structural changes across the cardiac cycle is essential for measuring aortic distensibility, a primary marker of arterial stiffness. However, standard 2D segmentation networks focus on each frame independently. Consequently, when rapid systolic flow temporarily obscures the aorta's boundaries, this lack of continuous context results in frame-to-frame tracking dropouts and boundary inconsistencies. Spatiotemporal ($2\text{D}+t$) networks can enforce temporal consistency across the sequence but suffer from a scarcity of expert annotations. To address this, we present a semi-supervised spatiotemporal ($2\text{D}$ to $2\text{D}+t$) knowledge distillation framework exploiting the cardiac cycle. The framework distills a spatial teacher's expertise into a spatiotemporal student network by executing a dynamic latent interception, pairing a recurrent spatiotemporal bottleneck with a residual spatial bypass. Our model selection strategy applies a baseline validation threshold ($\text{DSC} \ge 0.50$) prior to selecting the epoch that maximizes anatomical consistency. This strategy enables the spatiotemporal student model to achieve superior surface tracking accuracy ($\text{NSD@1mm} = 92.3\% \pm 0.2\%$) and high structural reliability ($\text{Frac}_{2\text{CC}} = 99.2\% \pm 0.6\%$), reducing population-wide structural anomalies by over 56\% compared to a 2D nnU-Net baseline.
1 Introduction
The paper targets label-efficient aortic tracking across cine-MRI cardiac cycles, where independent-frame segmentation and limited annotations undermine temporal and anatomical consistency. It introduces semi-supervised distillation combining spatial teacher predictions, structural correction, recurrent temporal modeling, and smoothness regularization.
- Continuous aortic tracking supports cardiovascular-function assessment through dynamic measurements such as aortic distensibility and pulse wave velocity.
- Manual annotation across many cine-MRI frames is labor-intensive, observer-variable, and compromises reproducibility of consecutive-frame boundaries.
- Independent 2D networks ignore temporal dynamics, causing topological inconsistencies, obscured-aorta dropouts, and mistaken segmentation of adjacent bright vessels.
- Standard knowledge distillation can propagate artifact-induced hallucinations from 2D teachers into pseudo-labels, destabilizing student training.
- The proposed framework sanitizes static-teacher pseudo-labels before distilling them into a spatiotemporal student that uses adjacent-frame context to preserve boundary continuity.
- The student combines a parameter-efficient latent recurrent bottleneck with temporal smoothness regularization that penalizes tracking flicker and boundary jitter while allowing physiological motion.
2 Methodology
The method separates supervised spatial learning and pseudo-label correction from sequence-wide student distillation. Temporal features are injected through a bidirectional recurrent bottleneck and optimized with a smoothness objective to maintain anatomically valid tracking.
- Problem formulation: The framework maps cine-MRI sequences to binary masks while requiring exactly two connected components for the ascending and descending aorta.
- Training strategy: Three data pools support decoupled training: labeled end-diastolic data trains the teacher, unannotated data receives corrected pseudo-labels, and a separate set provides baseline testing.
- Phase I: Teacher optimization: The spatial teacher uses a joint Dice and focal objective, while checkpoint selection first requires DSC ≥τdice and then maximizes anatomical feasibility rate.
- Phase I: Geometric label correction: Teacher predictions are binarized and corrected offline with deterministic verification, pruning, and bidirectional circular stitching under an 8-connectivity rule.
- Phase II: Student architecture: A dynamic hook intercepts deepest-layer features, compresses them, processes them bidirectionally with ConvLSTM, and merges the temporal correction through a residual branch.
- Phase II: Student architecture: The student produces sequence tracking probabilities after enhanced bottleneck features are re-folded through standard upsampling pathways.
- Phase II: Student objectives: The student objective balances frame-wise spatial loss with temporal smoothness, including amplified weighting for the clip-center anchor frame.
- Phase II: Student objectives: The differentiable smoothness loss uses the mean L1 norm of a discrete second derivative, vanishing under uniform velocity while penalizing sudden tracking anomalies.
3 Results
The evaluation combines spatial accuracy, anatomical feasibility, and population-wide structural audits across independent test data and unannotated frames. The spatiotemporal student preserves topology and reduces structural anomalies, with temporal regularization requiring a balanced setting.
- Experimental setup: 437 scans were partitioned into a 412-patient distillation pool and an unseen 25-patient test set, with development confined to 5-fold cross-validation within the distillation pool.The test set was reserved exclusively for final evaluation.
- Evaluation metrics: The evaluation reports DSC, NSD@1mm, HD95, ASSD, and the Anatomical Feasibility Rate, which measures frames retaining exactly two foreground components.These metrics jointly assess spatial accuracy and sequence-wide topology.
- Qualitative tracking: The student maintained anatomical plausibility during fading aortic visibility, whereas the 2D nnU-Net baseline selected an erroneous nearby vessel.The comparison spans consecutive frames t − 1, t, and t + 1.
- Population-wide consistency: The student reduced structural anomalies to 451.2±67.8 per fold across 17,539 frames, compared with 1,036.8 for nnU-Net.The student also completely eliminated low-confidence tracking failures.
- Hyperparameter sensitivity: β = 0.10 produced the population-wide performance peak; β = 0.00 inflated tracking errors, while β = 0.50 induced tracking lag.The sweep evaluated the balance between voxel overlap and sequence-wide boundary acceleration constraints.
4 Conclusion
The semi-supervised distillation framework enforces sequence-wide consistency in 2D aortic cine-MRI tracking without extra annotations, while achieving high surface accuracy. Its applicability is bounded by assumptions about topology and temporal buffering.
- 17,539 frames validated the framework, which eliminated frame-to-frame tracking dropouts while achieving NSD@1mm = 92.3% ± 0.2%.
- The framework enforces sequence-wide consistency without requiring extra annotations through semi-supervised distillation.
- Geometric mask correction assumes two isolated circular cross-sections (N = 2), limiting use in branching regions and split-vessel pathologies.
- Bidirectional ConvLSTM buffering of an 11-frame window (T = 11) makes the framework suitable for retrospective review rather than real-time streaming.
Disclosure of Interests
The authors declare no relevant competing interests.
- The authors declare no competing interests relevant to the article's content.