The inner cell mass (ICM) regions can be used to predict pregnancy with a similar accuracy as the original images of the entire embryo

Abstract Study question Can clinical pregnancy be accurately predicted using only ICM and TE regions in an embryo microscopy image? Summary answer The AI model accurately predicted pregnancy using ICM segmented images with a similar accuracy as using original embryo images. What is known already Previous research suggests that ICM and TE are key factors in determining embryo grade and clinical pregnancy. Grad-CAM images on the CNN models in various literature also highlight the significance of ICM and TE regions. However, there have been no studies to compare and assess the importance of the two regions in pregnancy prediction. This study aims to test the significance of each region by evaluating the performance of models after segmenting ICM and TE regions. Study design, size, duration We performed a retrospective study of single static images of 2,555 Day 5 blastocysts from seven in vitro fertilization (IVF) clinics between June 2011 and May 2022. The images were collected from standard optical light microscopes and matched with metadata such as pregnancy outcomes. ICM and TE regions were manually labeled and examined by 5 embryologists with over 10 years of experience. Participants/materials, setting, methods We defined a positive pregnancy indication as the presence of a gestational sac (G-SAC). We built 4 CNN models using original embryo images (“Original model”), ICM segmented images (“ICM model”), TE segmented images (“TE model”) and ICM & TE segmented images (“ICM & TE model”). The performances were measured through 3-fold cross validation. Main results and the role of chance In 3-fold cross validation, the mean and the standard deviation of AUROCs were 0.753±0.005 for Original model, 0.733±0.003 for ICM model, 0.724±0.016 for ICM & TE model and 0.690±0.011 for TE model, respectively. The DeLong test showed significant differences between the AUCs of Original model and ICM model (p = 0.01), and ICM model and ICM & TE model (p = 0.0003). Positive correlations were found between prediction values with pearson correlation [Confidence interval] for Original image and ICM model (0.426 [0.343, 0.495]), Original model and ICM & TE model (0.345 [0.266, 0.419]) and Original model and TE model (0.223 [0.140, 0.304]). Original model showed the highest accuracy and TE model was the least accurate as expected. However, it is noteworthy that ICM model outperformed ICM & TE model. Although it is well known that TE plays a significant role in implantation, forcing the model to utilize both ICM and TE to predict pregnancy could have muddled the analysis. This study proved that pregnancy can be successfully predicted using ICM regions only as well as original embryo images. Limitations, reasons for caution This study has limitations due to its retrospective nature, using embryo images from seven IVF centers. Further study with a larger dataset is warranted to offset the differences in focus, magnification and color of embryo images. Wider implications of the findings We showed that the AI model using ICM segmented images predicted pregnancy with a similarly high accuracy as using original embryo images. However, the model using TE segmented images showed the lowest performances. Utilizing well-focused ICM images may enhance the performance of pregnancy prediction models. Trial registration number not applicable