近幾年來，胚胎時差培養技術（time-lapse incubation）在輔助生殖治療領域取得了重大突破，改變了常規胚胎培養模式。已經有大量研究文獻證明了EmbryoScope時差培養箱在臨床上的安全性和有效性，但是目前需要利用輔助注釋（Guided Annotation）工具自動地提示所選的變量，并由胚胎學家進行確認，使用KIDScore決策支持工具為胚胎評估預測胎心妊娠的可能性，選擇出著床潛能最高的胚胎進行移植或者冷凍。
深度學習模型能否從time-lapse視頻中預測胎心（FH）妊娠的可能性？
我們創建了一個名為IVY的深度學習模型，該模型是一種客觀且完全自動化的系統，可以從原始time-lapse視頻中直接預測胎心妊娠的可能性，無需任何人工形態動力學注釋或囊胚形態學評估。實驗設計：回顧性分析了2014年1月至2018年12月期間，8個來自四個不同國家（地區）的IVF中心， 總計10 638個胚胎的實時成像和臨床結果，使用具有已知胎心FH妊娠結果的實時成像視頻對深度學習模型進行了模擬訓練，在給定延時視頻的情況下，執行預測FH懷孕可能性的二進制分類任務。深度學習模型能夠根據實時成像視頻的5層分層交叉驗證中的AUC為0.93 [95％CI 0.92-0.94]來預測胎心FH妊娠。在八個不同的實驗室進行的驗證測試表明AUC具有可重現性，AUC范圍為0.95至0.90。深度學習模型對胚胎植入具有很高的預測價值，可能會提高通過實時成像系統胚胎選擇的方法的有效性， 可以改善單次胚胎移植的結果。 深度學習模型也可能被證明可為隨后低溫保存的胚胎的轉移提供最佳順序。
STUDY QUESTION: Can a deep learning model predict the probability of pregnancy with fetal heart (FH) from time-lapse videos?
SUMMARY ANSWER: We created a deep learning model named IVY, which was an objective and fully automated system that predicts the probability of FH pregnancy directly from raw time-lapse videos without the need for any manual morphokinetic annotation or blastocyst morphology assessment.
WHAT IS KNOWN ALREADY: The contribution of time-lapse imaging in effective embryo selection is promising. Existing algorithms for the analysis of time-lapse imaging are based on morphology and morphokinetic parameters that require subjective human annotation and thus have intrinsic inter-reader and intra-reader variability. Deep learning offers promise for the automation and standardization of embryo selection.
STUDY DESIGN, SIZE, DURATION: A retrospective analysis of time-lapse videos and clinical outcomes of 10 638 embryos from eight different IVF clinics, across four different countries, between January 2014 and December 2018.
PARTICIPANTS/MATERIALS, SETTING, METHODS: The deep learning model was trained using time-lapse videos with known FH pregnancy outcome to perform a binary classification task of predicting the probability of pregnancy with FH given time-lapse video sequence.The predictive power of the model was measured using the average area under the curve (AUC) of the receiver operating characteristic curve over 5-fold stratified cross-validation.
MAIN RESULTS AND THE ROLE OF CHANCE: The deep learning model was able to predict FH pregnancy from time-lapse videos with an AUC of 0.93 [95% CI 0.92–0.94] in 5-fold stratified cross-validation. A hold-out validation test across eight laboratories showed that the AUC was reproducible, ranging from 0.95 to 0.90 across different laboratories with different culture and laboratory processes.
LIMITATIONS, REASONS FOR CAUTION: This study is a retrospective analysis demonstrating that the deep learning model has a high level of predictability of the likelihood that an embryo will implant. The clinical impacts of these findings are still uncertain. Further studies, including prospective randomized controlled trials, are required to evaluate the clinical significance of this deep learning model. The time-lapse videos collected for training and validation are Day 5 embryos; hence, additional adjustment would need to be made for the model to be used in the context of Day 3 transfer.
WIDER IMPLICATIONS OF THE FINDINGS: The high predictive value for embryo implantation obtained by the deep learning model may improve the effectiveness of previous approaches used for time-lapse imaging in embryo selection. This may improve the prioritization of the most viable embryo for a single embryo transfer. The deep learning model may also prove to be useful in providing the optimal order for subsequent transfers of cryopreserved embryos.
STUDY FUNDING/COMPETING INTEREST(S): D.T. is the co-owner of Harrison AI that has patented this methodology in association with Virtus Health. P.I. is a shareholder in Virtus Health. S.C., P.I. and D.G. are all either employees or contracted with Virtus Health. D.G. has received grant support from Vitrolife, the manufacturer of the Embryoscope time-lapse imaging used in this study. The equipment and time for this study have been jointly provided by Harrison AI and Virtus Health.