STORK: Spatio-Temporal Observation of uterine contRactions via neural networKs

Melissa Schween; Tristan Gottwald; Jordina Aviles Verdera; Lisa Story; Mary Rutherford; Jana Hutter

cs.CVOct 7, 2026

STORK: Spatio-Temporal Observation of uterine contRactions via neural networKs

Authors: Melissa Schween, Tristan Gottwald, Jordina Aviles Verdera, Lisa Story, Mary Rutherford, Jana Hutter

Organizations: CAIMED, Leibniz University Hannover, Hannover, Germany · L3S, Leibniz University Hannover, Hannover, Germany · Institute of Information Processing, Leibniz University Hannover, Germany · Department for Biomedical Engineering, King’s College London, London, UK

Abstract

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.

Figures & tables

Figure 1 : Illustration of contractile activity captured during a dynamic Multi-Echo Gradient Echo MRI scan, showing quantitative mean T2* and volume (left), acquired data at the first echo time (center) and quantitative T2* maps (right).

Figure 2 : STORK network architecture composed of a preprocessing stage (green), a feature extractor (blue) and backbone (pink). Preprocessing stage transforms each series into a 4-channel input combining magnitude images with Demons displacement fields. The feature extractor processes the 4D volume by leveraging 3D convolutions on hyperplanes always involving the temporal dimension. The backbone aggregates information over the time dimension by average pooling combined with a linear projection.

	Metrics (mean $\pm$ std)
Approach	Series AUROC $\uparrow$	Series AUPRC $\uparrow$	Frame AUROC $\uparrow$	Frame AUPRC $\uparrow$
STORK (ours)	0.950 $\pm$ 0.011	0.946 $\pm$ 0.016	0.797 $\pm$ 0.015	0.674 $\pm$ 0.026
STORK $+$ Attention	0.915 $\pm$ 0.022	0.899 $\pm$ 0.036	0.763 $\pm$ 0.057	0.657 $\pm$ 0.091
ResNet Baseline	0.556 $\pm$ 0.096	0.520 $\pm$ 0.096	n/a	n/a
ConvNeXt Baseline	0.610 $\pm$ 0.064	0.558 $\pm$ 0.073	n/a	n/a

Table 1 : Test-set results over seeds. Best values in bold .

Panel of Figure 3 — Figure 3 : t-SNE plot of the latent spaces for ResNet baseline (a) and STORK (b), showing separation of non-contractile (blue) and contractile (red) cases.

	Metrics (mean $\pm$ std)
Preprocessing	Series AUROC $\uparrow$	Series AUPRC $\uparrow$	Frame AUROC $\uparrow$	Frame AUPRC $\uparrow$
Echo 0	0.937 $\pm$ 0.014	0.934 $\pm$ 0.017	0.784 $\pm$ 0.025	0.654 $\pm$ 0.041
Echo 1	0.950 $\pm$ 0.011	0.946 $\pm$ 0.016	0.797 $\pm$ 0.015	0.674 $\pm$ 0.026
Echo 2	0.950 $\pm$ 0.011	0.943 $\pm$ 0.020	0.795 $\pm$ 0.016	0.658 $\pm$ 0.024
Echo 0, 1, 2	0.948 $\pm$ 0.010	0.948 $\pm$ 0.015	0.799 $\pm$ 0.019	0.664 $\pm$ 0.043

Table 2 : Echo ablation comparing single and the best combined echoes at the series and frame level. Best values in bold . Echoes are indexed from 0.

Figure 4 : Frame-wise Grad-CAM. Top: reference intensity with placenta contour (cyan); bottom: same frames with the Grad-CAM overlay for the series logit.

Explore similar work

Unsupervised Adversarial Domain Adaptation for Uterine layer Segmentation: From Labeled Cine to Unlabeled Dynamic EPI MRI
Aug 4, 2026Smiti Tripathy, Milauni Desai, Jordina Aviles Verdera +1Magnetic Resonance Imaging U-Net
Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany · CAIMED, L3S, Hannover, Germany · Institut für Informationsverarbeitung, Leibniz University Hannover, Germany
Female-RHINO: A Real-Time Scanner-Integrated Framework for Automated Quantitative Uterine MRI Analysis and Structured Reporting
Jun 23, 2026Deepak Bhatia, Saad Ahmad, Smiti Tripathy +7Real-World Clinical Workflows One-Shot Medical Landmark Detection
Radiologisches Institut, Universitätsklinikum Erlangen, Erlangen, Germany · School of Biomedical Engineering and Imaging, KCL, London, UK · Institute for Information Processing, Leibniz University Hannover, Germany and CAIMed, L3S research institute, Hannover, Germany +1
Fetal Brain Imaging: A Composite Neural Network Approach for Keyframe Detection in Ultrasound Videos
May 10, 2026Aleksander Zamojski, Kacper Jarczak, Radoslaw RoszczykUltrasound Convolutional Neural Networks
Warsaw Universitu of Technology Warsaw, Poland · Warsaw University of Technology Warsaw, Poland

Aug 4, 2026, eess.IVJ/K move · Enter open · S save

Unsupervised Adversarial Domain Adaptation for Uterine layer Segmentation: From Labeled Cine to Unlabeled Dynamic EPI MRI

Smiti Tripathy, Milauni Desai, Jordina Aviles Verdera, Jana Hutter

Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany · CAIMED, L3S, Hannover, Germany · Institut für Informationsverarbeitung, Leibniz University Hannover, Germany

Magnetic Resonance Imaging U-Net Domain Adaptation Tissue Films

Open paper AskPDF arXiv

Uterine peristalsis is a key physiological phenomenon responsible for various functions across the menstrual cycle, intimately linked to uterine wall microstructure. Alterations in uterine motion and tissue properties are implicated in the etiology of gynecological diseases, yet these processes have been studied in isolation. We introduce a dynamic multi-echo gradient echo EPI framework for simultaneous characterization and correlation of uterine peristaltic activity and time-resolved T2* changes at 0.55T. Inherent susceptibility artifacts, reduced resolution, and burden of manual uterine layer annotation are addressed by an unsupervised adversarial domain adaptation framework, transferring segmentation knowledge from labeled cine MRI to unlabeled dynamic EPI. We implemented Unet-LSTM with multi-scale domain discriminators that exploits temporal layer dynamics. A Dice score of 0.88 and Jaccard index of 0.80 was achieved. Mean T2* values were 108ms, 76ms, and 124ms for the myometrium, junctional zone, and endometrium. A negative correlation between junctional zone area and T2* was observed in 14/39 cases, providing first insights into oxygenation patterns associated with junctional zone contraction and motion, demonstrating feasibility of assessing the interplay between contractility and dynamic T2* changes.

Abstract

Explore similar work

Unsupervised Adversarial Domain Adaptation for Uterine layer Segmentation: From Labeled Cine to Unlabeled Dynamic EPI MRI

Female-RHINO: A Real-Time Scanner-Integrated Framework for Automated Quantitative Uterine MRI Analysis and Structured Reporting

Fetal Brain Imaging: A Composite Neural Network Approach for Keyframe Detection in Ultrasound Videos