Related Experiment Video
Updated: Oct 18, 2026
In Vivo Visualization of Calcium Transients during Fertilization and Early Development in C. elegans
Published on: July 12, 2024
Spatiotemporal deep learning with visualization predicts fertilization for early rescue intracytoplasmic sperm
Masato Saito1, Hirofumi Haraguchi2, Naoko Yamamoto2
1Matsumoto Ladies IVF Clinic, Tokyo, Japan; Department of Integrated Applied Life Science, Integrated Graduate School of Medicine, Engineering, and Agricultural Sciences, University of Yamanashi, Kofu, Yamanashi, Japan.
Objective:
To determine whether a deep learning (DL) model trained on time-lapse videos predicts fertilization following short-term insemination more accurately than senior embryologists, particularly in morphologically difficult cases, to visualize classification regions, and to evaluate whether a shortened observation window could still support accurate classification before early rescue intracytoplasmic sperm injection (r-ICSI).
Design:
Single-center retrospective study.
Subjects:
A total of 1,331 embryos from 250 short-term insemination cycles were identified. After excluding immature and degenerated oocytes, 1,016 embryos (770 fertilized, 246 unfertilized) were analyzed and split at the patient level into training and validation (n = 925) and test (n = 91) sets.
Exposure:
Time-lapse videos recorded from 4.5 to 8 h post-insemination were input into a spatiotemporal convolutional neural network trained with 5-fold cross-validation and compared with independent classification by 5 senior embryologists (>5 years' experience). Holdout embryos were stratified by embryologist accuracy, and classification regions were visualized. Model performance was also assessed using progressively truncated videos.
Main Outcome Measures:
Area under the receiver operating characteristic curve (ROC-AUC), accuracy, recall, specificity, precision, negative predictive value (NPV), and F1 score on the holdout test set, and stratum-specific accuracy.
Results:
The DL model achieved a mean ROC-AUC of 0.938 (ensemble model, 0.968) and significantly higher accuracy (0.884 versus 0.802), recall (0.946 versus 0.837), NPV (0.802 versus 0.557), and F1 score (0.927 versus 0.867) than senior embryologists. In moderate- and low-accuracy cases (≤3 of 5 embryologists correct), the DL model classified fertilized embryos significantly more accurately than embryologists (P < 0.05). Even when trained on videos truncated at 6 h, recall and NPV remained significantly higher than embryologists using the full 4.5-8.0 h videos. Visualization showed cytoplasmic highlighting in fertilized embryos and polar body and perivitelline space highlighting in unfertilized embryos.
Conclusion:
A DL model with region visualization achieved significantly higher predictive accuracy than senior embryologists, including for fertilized embryos in cases prone to human misclassification, and retained higher recall and NPV even when trained on videos truncated at 6 h. This suggests potential utility for early r-ICSI, although multicenter prospective validation is required.
