Source-linked AI summary
STORK: Spatio-Temporal Observation of uterine contRactions via neural networKs
Melissa Schween, Tristan Gottwald, Jordina Aviles Verdera, Lisa Story, Mary Rutherford, Jana Hutter
TL;DR
Contraction-affected dynamic fetal MRI is manually identified and discarded, motivating automated detection. STORK learns from series-level labels using motion-aware, factorized spatio-temporal features and mean pooling; it reaches 95.0% AUROC and 94.6% AUPRC at series level.
Problem
Method
STORK learns contraction detection from series-level labels using intensity and Demons displacement inputs, temporally factorized 3D features, and linear mean pooling.
Source: §1 · p. 2, §4.1 · p. 4, §4.2 · p. 6Results
95.0% series-level AUROC and 94.6% AUPRC; STORK outperforms TCN-based baselines and the attention variant at series level while recovering frame-level metrics without frame-level training supervision.
Source: §5.3 · pp. 7–8Takeaways & Limitations
Automated, non-invasive contraction monitoring could support richer phenotyping and, if integrated during acquisition, adaptive temporal resolution or extended scans during live events.
Source: §6 · pp. 9–10Grad-CAM attributions remain qualitative and cannot be validated against placenta-derived labels; quantitative attribution analysis is left for future work.
Source: §5.5 · p. 9
Abstract
from arXiv · showhide
Uterine contractions in fetal MRI are typically identified manually and discarded, limiting insights into contraction dynamics. We formalize Uterine Contractile Activity Detection (UCAD) as a weakly-supervised learning problem and introduce STORK, a multi-instance learning model trained on dynamic MRI series using only coarse, series-level labels. STORK factorizes 3D spatio-temporal convolutions into parallel branches across temporal hyperplanes to capture coherent tissue motion without the cost of full 4D convolutions. Per-frame embeddings, combining intensity and Demons-estimated displacement fields, are aggregated by a linear mean-pooling head. This ensures that frame-level contraction scores can be recovered post-hoc without frame-level training supervision. Evaluated on around 700 multi-vendor dynamic fetal MRI series, STORK achieves a series-level AUROC of 95.0% and AUPRC of 94.6%, substantially outperforming 3D ResNet and ConvNeXt baselines. Grad-CAM analysis suggests that the model draws on predictive features extending beyond the placenta into the uterine tissue, offering an automated tool for richer phenotyping of uterine behavior.
1 Introduction
Contractions complicate dynamic fetal MRI, yet MRI offers whole-uterus spatio-temporal coverage and functional contrasts. STORK is introduced as a weakly supervised approach to detect contractions and recover frame-level scores from series-level labels.
1 Introduction
- Ultrasound has operator-dependent, late-gestation field-of-view limits, while CTG lacks spatial resolution and placental functional insights; dynamic MRI offers whole-uterus coverage but contractions create artifacts.Affected images are currently identified manually and discarded.Source: §1 · p. 1
- Automated contraction detection could support explicit activity quantification, data-driven rejection of affected data, and adaptive imaging protocols.Source: §1 · p. 1
- STORK formalizes Uterine Contractile Activity Detection and introduces weakly supervised, hyperplane-factorized MIL, post-hoc frame scoring, and analysis of discriminative features beyond the placenta.Its linear mean-pooling backbone aggregates per-frame embeddings while preserving recoverable frame scores.Source: §1 · p. 2
2 Related Work
Prior uterine MRI work has characterized contraction-related changes and used placental T2* curves for threshold-based classification. Machine-learning work has commonly treated motion as artifact, whereas this paper uses displacement fields as a discriminative signal and adapts spatio-temporal factorization for detection.
2.1 Clinical Background and Uterine MRI
- Fetal dynamic MRI studies associate contractions with placental T2*, volume, and visual changes, while prior automatic classification used thresholding on placental T2* curves.Other work differentiated contraction types, excluded corrupted data, and examined downstream effects.Source: §2.1 · p. 2
2.2 Machine Learning Background
- Unlike prior fetal MRI machine learning that treated motion as artifact, this work uses displacement fields as a primary signal and adapts hyperplane factorization to encode temporal motion.STORK uses three 3D convolutional hyperplanes that include time, omitting the purely spatial plane.Source: §2.2 · pp. 2–3
3 Uterine Contractile Activity Detection (UCAD)
UCAD treats contraction detection as identifying coordinated tissue change in dynamic MRI despite other sources of motion and similar-looking pathology. The model predicts whether a series contains a contraction, while localization is recovered as a secondary objective.
3 Uterine Contractile Activity Detection (UCAD)
- Because contractions and respiration, peristalsis, fetal motion, or placental pathology can produce similar changes, UCAD learns from observable induced changes rather than explicitly defining contractions.Contractile and non-contractile motion components are not individually observable.Source: §3 · p. 3
- A contraction schema requires sustained contractile change above a threshold in a connected region of sufficient volume, persisting across a minimum time window.The schema is not a closed-form operational definition; the relevant quantities and rejection of non-contractile movement are learned.Source: §3 · p. 4
- The primary mapping predicts a series-level contraction probability; identifying where and when activity occurs is a secondary objective recovered without dedicated supervision.Source: §3 · p. 4
4 Methodology
STORK maps each dynamic MRI series through preprocessing, spatio-temporal feature extraction, and a backbone that aggregates frame embeddings into a series prediction.
4 Methodology
- The methodology maps each MRI series to a series-level classification through preprocessing, feature extraction, and a backbone.Source: §4 · p. 4, §4.2 · p. 5
4.1 Preprocessing
- Preprocessing crops around a placenta-defined bounding box while retaining uterine context, then combines single-echo intensity with clipped Demons displacement fields as four input channels.Channels are normalized per series to preserve temporal intensity changes; frames are resized to a fixed spatial resolution after displacement calculation and normalization.Source: §4.1 · pp. 4–5, Figure 2 · p. 5
4.2 Network Architecture
- Three parallel 3D convolutional branches on temporal hyperplanes encode coherent tissue motion while avoiding the cost of full 4D convolution.The purely spatial hyperplane is omitted; spatial pooling and a linear layer produce contextual per-frame embeddings.Source: §4.2 · p. 5, Figure 2 · p. 5
- A parameter-free temporal mean and linear layer aggregate frame embeddings, and the resulting logit decomposes into per-frame contributions for post-hoc localization.The model is trained end-to-end against series labels using binary cross-entropy, with positive instances weighted for class imbalance.Source: Figure 2 · p. 5, §4.2 · p. 6
5 Experiments
Experiments evaluate STORK and its baselines on multi-vendor dynamic fetal MRI, then examine representation quality and echo choice. STORK achieves strong series-level performance and frame-level localization without frame-level training supervision.
5 Experiments
- The evaluation reports AUROC and AUPRC for series-level classification and frame-level metrics for localized activity, using means and standard deviations across ten random runs.AUPRC is described as more stringent under the dataset’s class imbalance.Source: §5 · p. 6
5.1 Dataset
- 757 data sets from 549 pregnant participants were acquired across 0.55T, 1.5T, and 3T scanners; about 40% of cases were labeled as showing contractions.Labels were based on placental T2* signal, and subjects with manual frame-level labels were reserved for testing.Source: §5.1 · p. 6
5.2 Implementation details
- Frames were resized to 64 × 64 × 32, and Demons displacement fields were clipped to ±20 mm; models used AdamW and early stopping based on validation AUROC.Each configuration was trained with batch size 1 for up to 40 epochs, with the best validation-AUROC checkpoint selected.Source: §5.2 · p. 7
5.3 Main Results
- 95.0% AUROC and 94.6% AUPRC at the series level; STORK outperforms TCN-based baselines and exceeds attention pooling at the series level.The attention variant recovers most performance but remains below STORK at series level and has higher seed variance at frame level; unlike the TCN baselines, STORK provides frame-level localization.Source: §5.3 · pp. 7–8, Table 1 · p. 8
5.4 Latent Space Interpretability
- STORK’s test-set latent-space visualization separates contractile from non-contractile sequences, whereas the ResNet baseline does not.The comparison uses t-SNE projections of sequence embeddings.Source: §5.4 · p. 8, Figure 3 · p. 9
5.5 Ablations
- Echo 1 achieved the best value on two metrics, and its marginal differences from multi-echo input motivated its selection to reduce computational cost.Grad-CAM often highlighted regions outside the placenta, including the uterine wall; these attributions remain qualitative and cannot be validated against placenta-derived labels.Source: §5.5 · pp. 8–9, Table 2 · p. 9
6 Conclusion and Discussion
STORK detects uterine contractile activity in dynamic MRI and recovers frame-level localization without frame-level training. Its labels rely on automated T2* curve proxies, while Grad-CAM suggests predictive features can extend beyond the placenta.
6 Conclusion and Discussion
- STORK detects uterine contractile activity in dynamic MRI and recovers frame-level localization despite training only with series-level labels.The authors describe the approach as a richer phenotype of uterine behavior and suggest it may offer insight into complications such as fetal growth restriction.Source: §6 · p. 9
- Automated T2* curve proxies constrain training-label quality, and the resulting label noise may lead the model to learn signals from regions the proxies do not verify.The paper notes that uterine-wall Grad-CAM attributions remain qualitative and cannot be validated against placenta-derived labels.Source: §6 · p. 9, Figure 4 · p. 10
- Real-time detection during acquisition could allow scanning protocols to adapt temporal resolution or extend acquisitions during live events.The authors also describe the framework as potentially applicable to other transient biomedical events.Source: §6 · p. 10