A classification system of day 3 human embryos using deep learning. (September 2021)

• Record Type: Journal Article
• Authors: Wu, Chongwei; Yan, Wei; Li, Hongtu; Li, Jiaxin; Wang, Hongkai; Chang, Shijie; Yu, Tao; Jin, Ying; Ma, Chao; Luo, Yahong; Yi, Dongxu; Jiang, Xiran
• Is Part Of: Biomedical signal processing and control. Volume 70(2021)
• Publication Date: 2021-09
• Subjects: ART Assisted reproductive technologies -- IVF In vitro fertilization -- ICSI Intracytoplasmic sperm injection -- CNN Convolutional neural network -- EL Ensemble learning -- ResNet Residual Network -- VGGNet Visual Geometry Group -- ROC Receiver operating characteristic -- AUC Area under the ROC curve -- ACC Accuracy -- SEN Sensitivity -- SPE Specificity -- TP True positive -- TN True negative -- FP False positive -- FN False negative; Assisted reproductive technology -- In vitro fertilization -- Embryo quality -- Deep learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102943
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: This study demonstrates the first attempt of constructing a deep learning-based classification system for 4- and 3-category classifications of day 3 human embryo from 1800 couples. An EL model was proposed integrating the most discriminative features from four CNN models, and showed excellent classification performance in both the 4- and 3-category classification. Our EL model was compared against four embryologists on an independent test cohort from 350 couples, and showed good potential as a simple platform to any embryologists or anyone. Abstract: Conventional embryo evaluations were based on morphological analysis by skilled embryologists. Although the method has been universally used in clinical practice, the embryo evaluation based on low resolution microscopic image represents a crude and subjective assessment of embryo quality, which is incomplete as well as time-consuming. In this study, a total of 3601 microscopic images of previously classified day 3 embryos from 1800 couples undergoing in vitro fertilization were clinically obtained between Sep. 2016 and Mar. 2018. The images were subjected to various convolutional neural networks and a proposed deep ensemble learning (EL) model for computer-assisted embryo grading analysis. An independent test cohort contained 699 microscopic images from 350 couples were gathered from Apr. 2018 to Oct. 2018 and used to test the models. The EL model achieved the highest average classification accuracy of 74.14% in our 4-catogory classification system. The accuracy was further improved to 89.16% when categories 1 and 2 were combined. The model displayed better discrimination power than the embryologist average in both 4- and 3-category classification systems in the independent test cohort, which suggested good potential for transfer in fertility clinics.