In an article recently published in the journal Nature Communications, researchers introduced the Blastocyst Evaluation Learning Algorithm (BELA), a novel machine-learning model for advancing in vitro fertilization (IVF). This model aims to enhance the accuracy and efficiency of embryo selection by predicting ploidy status and assessing embryo quality through time-lapse imaging, ultimately improving IVF outcomes. The study addresses significant challenges in embryo selection, emphasizing the need for non-invasive techniques to assess embryo viability using artificial intelligence (AI).
Advancement in Embryo Assessment
Since its introduction in 1978, IVF has been a crucial option for individuals facing infertility, resulting in over 8 million births worldwide. A key factor of IVF success is selecting high-quality embryos, primarily determined by their ploidy status. Ploidy status represents the chromosomal composition of an embryo, which significantly influences pregnancy outcomes.
Euploid embryos, with a normal chromosome count, are more likely to lead to successful pregnancies. In contrast, aneuploid embryos, which have chromosomal abnormalities, are linked to miscarriages and genetic disorders. Traditional methods for assessing ploidy, such as preimplantation genetic testing for aneuploidy (PGT-A), are invasive, costly, and time-consuming, affecting embryo viability. However, recent advancements in AI and computer vision offer automated non-invasive embryo assessment using time-lapse imaging.
BELA: An Innovative AI Model
In this study, the authors presented BELA, an AI-based model designed to predict embryo ploidy status using time-lapse sequences of embryo development. They employed deep learning techniques to analyze these sequences across various stages of growth.
BELA operates in two main steps. First, it predicts the blastocyst score (BS), an intermediary measure of embryo quality, using processed day-5 time-lapse videos to capture key developmental stages. For feature extraction, the model utilizes a pre-trained visual geometry group 16-layer (VGG16) convolutional neural network (CNN) to capture spatial and temporal details from the images.
Then, a bidirectional long-short-term memory (BiLSTM) architecture is used to predict various morphological scores, including inner cell mass (ICM), trophectoderm (TE), and expansion scores. These scores are combined to derive a model-based blastocyst score (MDBS). In the second step, BELA uses the MDBS, along with maternal age, to predict embryo ploidy status through logistic regression.
Testing and Performance Evaluation
The study utilized two datasets from Weill Cornell Medicine’s Center for Reproductive Medicine, containing time-lapse sequences and PGT-A results for thousands of embryos. Datasets from clinics in Spain and Florida were also used to train and validate the model. Furthermore, performance was evaluated using accuracy, area under the receiver operating characteristic curve (AUC), precision, and recall.
Key Findings and Insights
The outcomes demonstrated that BELA outperformed accurately in predicting embryo ploidy status. It achieved an AUC of 0.66 when classifying embryos as euploid or aneuploid. This improved to 0.76 with the addition of maternal age, matching the performance of models based on embryologists' manual assessments.
BELA also showed promising results in identifying complex aneuploid embryos, reaching an AUC of 0.826 when maternal age was included. This suggests that BELA effectively detects embryos with multiple chromosomal abnormalities, which are often more challenging to identify. The study highlighted maternal age as a key predictor of ploidy, with lower maternal age correlating with higher rates of euploid predictions.
Compared to traditional models that rely solely on time-lapse imaging or manual annotations, BELA demonstrated superior performance. Its robustness across datasets from various clinics with different practices underscores its adaptability.
By predicting BS and ploidy status without requiring manual input from embryologists, BELA can potentially streamline the embryo evaluation process. The model’s consistent performance across maternal age groups indicates its broad applicability. Additionally, using time-lapse sequences from 96 to 112 hours post-insemination (HPI) and maternal age as inputs allows seamless integration into diverse clinical workflows.
Applications
This research proposed a non-invasive, accurate, and efficient method for predicting embryo ploidy status, enhancing the efficiency and accuracy of embryo selection in IVF. BELA has the potential to improve pregnancy outcomes and reduce the risks associated with invasive procedures like PGT-A. Its ability to provide explainable predictions through model-derived blastocyst scores can assist embryologists in making informed decisions.
Furthermore, the authors developed a web-based application called STORK-V to facilitate the clinical use of BELA, making it accessible to IVF clinics and embryologists. This platform not only predicts ploidy status but also offers intermediary quality scores, enriching the analytical tools available to medical professionals.
Conclusion and Future Directions
In summary, BELA represents a significant advancement in automating embryo assessment in IVF and could revolutionize the field of reproductive medicine. Its performance across multiple datasets and adaptability to different clinical workflows highlight its potential to improve the embryo selection process. While BELA is not intended to replace PGT-A, it provides valuable supplementary information that can support embryologists in their decision-making, ultimately enhancing IVF success rates.
Future work should focus on refining the model by incorporating additional maternal features and exploring advanced video classification architectures. Expanding the training dataset to include more diverse clinical settings could enhance the model's generalizability.
Journal Reference
Rajendran, S., Brendel, M., Barnes, J. et al. Automatic ploidy prediction and quality assessment of human blastocysts using time-lapse imaging. Nat Commun 15, 7756 (2024). DOI: 10.1038/s41467-024-51823-7, https://www.nature.com/articles/s41467-024-51823-7
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.