EUVIMEDEUVIMED entdecken
Crossref

L26/P-209 Embryo Re-Identification and Traceability through Graph-Based Neural Network Mirroring

S. Muñoz-Herrera, O. Paredes, G. Palmer, I. Aguilar, A. Drakeley, A. Chavez-Badiola, J. Cohen

Human Reproduction · 01.07.2026 · 41 (Supplement_1)

DOI: 10.1093/humrep/deag083.545 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von S. Muñoz-Herrera und 6 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study question Can graph-based embryo morphology combined with neural network mirroring (siamese neural networks) enable robust re-identification and embryo traceability across time-lapse sequences despite ongoing developmental changes? Summary answer A Siamese-trained graph neural network reliably re-identifies the same embryo through time-lapse observations (Top-1=71.8%, AUC=0.922), demonstrating graph-based representations preserve identity despite morphological changes during development. What is known already Time-lapse systems generate large volumes of images, yet embryo traceability depends largely on manual visual inspection. While recent studies using convolutional neural networks and vision transformers demonstrate that embryo re-identification is possible under controlled conditions, performance often declines across imaging devices or developmental stages. These approaches treat embryos as pixel arrays, potentially overlooking relational and structural information. Here, we adopt graph-based representation that models embryos as interconnected image regions. Combined with Siamese neural network architecture, where two identical networks learn to compare inputs, our method focuses on learning embryo identity and similarity across developmental time rather than simple classification. Study design, size, duration Retrospective single-centre study using 221 embryos from a private assisted reproduction laboratory. Each embryo time-lapse sequence contained 55-72 frames captured at 2-hour intervals. Graph representations were generated for each frame using patch-based nodes with k-nearest neighbour connectivity. The Siamese training followed a three-phase curriculum (A→B→C), progressively incorporating temporal quartiles (early→intermediate→late developmental stages) to handle increasing morphological variation. Participants/materials, setting, methods We included embryos with complete time-lapse sequences reaching blastocyst stage. A graph was made from each frame (nodes=image patches, edges=k-NN connections). A mirror (Siamese) architecture with twin GCN encoders learnt representations preserving identity across time. Training used contrastive loss (InfoNCE) with hard triplet mining. Evaluation employed two complementary approaches: (1) retrieval scenario measuring Top-k accuracy, matching late-stage queries against early-stage gallery, and (2) binary verification (same vs. different embryo) using cosine similarity with ROC/PR curves. Main results and the role of chance The “mirror model” demonstrated robust re-identification across developmental stages. In the retrieval scenario (TEST set), the model achieved Top-1 accuracy of 71.8%, with performance improving to Top-5=74.5%, Top-10=75.3%, and Top-20=80.3%. This indicates approximately 72% of queries, the correct embryo appeared as the first-ranked match when comparing late developmental stages against earlier stages from the same embryo. For binary verification (same embryo vs. different embryo), the model achieved AUC=0.922 and Average Precision (AP)=0.869, confirming strong separability between positive and negative pairs. Operating at the F1-maximising threshold (cosine similarity=0.777), the system showed precision=0.856, recall=0.701, and specificity=0.963. The model exhibited notably conservative behaviour patterns: when pairs were truly from different embryos, the system correctly classified 96.3% of cases (minimising false positives that would incorrectly match different embryos). However, for pairs from the same embryo, recall was 70.1%, indicating 29.9% false negatives where the system failed to recognise matches. This conservative pattern prioritises avoiding embryo mix-ups over maximising all correct matches. Similarity score distributions showed clear separation: positive pairs concentrated near cosine similarity=1.0, while negative pairs clustered around 0 or below. Overlap occurred in an intermediate zone (approximately 0.3-0.6), concentrating challenging cases including morphologically similar embryos and extreme developmental transitions within the same embryo. Limitations, reasons for caution The study used single-centre data with one time-lapse platform. Graph construction depends on segmentation quality and hyperparameter selection (patch size, k-neighbours). The conservative threshold reduces false matches but increases false negatives. Cross-device validation and prospective evaluation are needed to confirm generalisability across different acquisition systems and clinical settings. Wider implications of the findings Graph-based representations offer novel approaches to embryo traceability, capturing morphological relationships rather than pixel patterns. This framework complements metadata-based quality control systems, providing intrinsic morphological verification. The conservative behaviour pattern may be clinically appropriate, prioritising patient safety by avoiding embryo mix-ups. Future applications could include cross-device identification and post-cryopreservation matching. Trial registration number No

Zitieren

APA

Muñoz-Herrera, S., Paredes, O., Palmer, G., Aguilar, I., Drakeley, A., Chavez-Badiola, A., & Cohen, J. (2026). L26/P-209 Embryo Re-Identification and Traceability through Graph-Based Neural Network Mirroring. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.545

Vancouver

Muñoz-Herrera S, Paredes O, Palmer G, Aguilar I, Drakeley A, Chavez-Badiola A, et al. L26/P-209 Embryo Re-Identification and Traceability through Graph-Based Neural Network Mirroring. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.545

BibTeX
@article{MuozHerrera2026,
  author = {Muñoz-Herrera, S. and Paredes, O. and Palmer, G. and Aguilar, I. and Drakeley, A. and Chavez-Badiola, A. and Cohen, J.},
  title = {L26/P-209 Embryo Re-Identification and Traceability through Graph-Based Neural Network Mirroring},
  journal = {Human Reproduction},
  year = {2026},
  volume = {41},
  number = {Supplement_1},
  doi = {10.1093/humrep/deag083.545}
}
Kostenlos starten →