2025.04.17
Masato Saito, Hirofumi Haraguchi, Ikumi Nakajima, Shinya Fukuda, Chenghua Zhu, Norio Masuya, Kazunori Matsumoto, Yuya Yoshikawa, Tomoki Tanaka, Satoshi Kishigami, Leona Matsumoto
http://dx.doi.org/10.1002/rmb2.12649
This study established a machine learning model (MLM) trained on embryo images to predict fertilization following short-term insemination for early rescue ICSI and compared its predictive performance with the embryologist’s manual classification.
Embryo images at 4.5 and 8 h post-insemination were preprocessed into vectors using ResNet50. The Light Gradient Boosting Machine (Light GBM) was employed for training vectors. Fertilization in the test dataset was assessed by MLM, with seven senior and 11 junior embryologists. Predictive metrics were analyzed using repeated measures ANOVA and paired t-tests.
Comparing MLM, senior embryologists, and junior embryologists, significant differences were observed in accuracy (0.71 ± 0.01, 0.75 ± 0.05, 0.61 ± 0.05), recall (0.84 ± 0.02, 0.84 ± 0.10, 0.61 ± 0.07), F1-score (0.78 ± 0.01, 0.81 ± 0.04, 0.66 ± 0.04), and area under the curve (0.73 ± 0.0 3, 0.73 ± 0.06, 0.61 ± 0.07), the MLM outperforming junior embryologists with <1 year of experience. No significant differences were observed between the MLM and senior embryologists with over 5 years of experience.
MLM can effectively predict fertilization following short-term insemination by analyzing cytoplasmic changes in images. These results underscore the potential to enhance clinical decision-making and improve patient outcomes.