Automated and precise recognition of human zygote cytoplasm: A robust image-segmentation system based on a convolutional neural network

Abstract In this study, we evaluated the precision and reproducibility of convolutional neural network (CNN) measurements of zygote cytoplasm segmentation relative to experienced embryologists. We developed an automated system for segmenting human zygote cytoplasm using images and a CNN trained on a set of 550 zygote images labeled by embryologists. Further, we compared the consistency in cytoplasmic area measurements using images segmented by the CNN and the trained human eye (embryologists), as well as the precision and reproducibility of the measurements and the effect of several common image attributes. Moreover, the use of CNN results for 8377 zygote images for regularity classification revealed inter-observer agreement between two embryologists (98 ± 1.02 %) and between embryologists and the CNN regarding the cytoplasmic area (97.75 ± 1.45 %), with intra-observer agreement for the CNN system at 100 %. Furthermore, neither luminance nor image noise significantly affected CNN measurement precision; however, highly irregular zygote shapes reduced precision to 95 % for both. For classification, automatic segmentation of images of 8377 zygotes (area measurement: 9741.34 ± 7951.83 μm2; shapes: regular, 65.37 %; slightly irregular, 15.10 %; moderately irregular, 9.91 %; and highly irregular, 9.62 %) revealed good performance by the CNN in classifying both regularly and irregularly shaped zygotes (area under curve: 0.874 ± 0.043). These results demonstrated the efficacy of the CNN system for precisely measuring the zygote cytoplasmic area, classifying the shape regularity, and assisting with embryo assessment.