Page 15 - HKSEMR2020 Programme book

Oral Presentation (Basic Science) Abstracts

Establishment of an artificial intelligence model for morphologic

prediction of sperm fertilization potential

TY Leung, CL Lee, PC Chiu
Department of Obstetrics and Gynaecology, The University of Hong Kong, Pokfulam Road, Hong Kong SAR;
Shenzhen Key Laboratory of Fertility Regulation, Department of Obstetrics and Gynaecology, The University of
Hong Kong-Shenzhen Hospital.

Introduction / Background / Objectives: attending our assisted reproduction program and processed
using our established protocols. ZP-bound and -unbound sperm
Human fertilization begins when a capacitated spermatozoon
binds to the zona pellucida (ZP) of a mature oocyte. In fertile men, were collected using a sperm-oocyte binding assay. ZP-binding
only a small proportion of motile sperm in the ejaculate (< 14%) protein expressions, DNA integrity and morphology of the
has the capacity to bind to the ZP. This observation is suggested isolated sperm subpopulations were analyzed by flow cytometry,
to be due to the capability of ZP selectively interacting with immunofluorescence, TUNEL/comet assay and Diff-Quik staining.
the sperm subpopulations characterized by normal chromatin Diff-Quik stained sperm images were pre-processed and fed into
structure, superior morphology and fertilization ability. Defective the CNNs using the Matlab software (MathWorks, MA, USA). The
spermatozoa-ZP interaction causes subfertility and is a major computational complexity (i.e. cost of prediction) and accuracy of
cause of low fertilization rates in clinical in-vitro fertilization (IVF). the CNN model were determined.
While intracytoplasmic sperm injection benefits patients with
defective spermatozoa-ZP binding, a standard method to identify Results / Outcomes:
such patients prior to conventional IVF is lacking.
Our results showed that ZP-bound sperm had significantly (P <
Deep learning, a core element of artificial intelligence (AI), has 0.05) higher ZP-binding protein expression, improved DNA integrity
recently emerged as a subset of machine learning that can and better morphology compared to the corresponding controls.
efficiently leverage various forms of structured data, such as A total of 1,334 and 885 images of ZP-bound and -unbound
images (arrays of pixel values). Convolutional neural networks sperm were collected, respectively. A novel algorithm for sperm
(CNN)s, a main component of many deep-learning algorithms, is a image transformation and segmentation were developed to pre-
specific approach originally proposed for image analysis, including process the images. CNN architecture was then applied on these
image classification as well as object and scene recognition. CNNs pre-processed images for feature extraction and model training.
systematically process all the data, composing a digital image A CNN model was successfully established to classify ZP-bound
through layers to identify the underlying relationship between and -unbound sperm with a high accuracy of 85% and low
them. The application of CNNs to cell morphology analysis has computational complexity.
become a topic of growing interest owing to the fact that the
traditional morphological assessment is highly subjective and time-
consuming. Conclusion:
We hypothesize that the CNN model will be a useful tool for the Evaluation of the fertilization potential of sperm samples will
morphologic prediction of human sperm ZP-binding capacity/ provide a comprehensive insight into treatment plans. The
fertilization potential. successful establishment of an AI model in this study suggests

its potential use as a simple, objective and quick evaluation for
morphologic prediction of the ZP-binding capacity/fertilization
Methods:
potential of semen samples in clinical assisted reproductive
Human spermatozoa and oocytes were obtained from donors settings.

13