UP-163 Deep learning-based automated sperm identification for non-obstructive azoospermia patients
Thursday June 27, 2019 from
TBD
Presenter

Ryan Lee, Canada

Research Engineer

University of British Columbia

Abstract

Deep learning-based automated sperm identification for non-obstructive azoospermia patients

Ryan Lee1,2,5, Luke Witherspoon4,5, Hongshen Ma1,2,3,5, Ryan Flannigan5,6.

1Department of Mechanical Engineering, University of British Columbia, Vancouver, BC, Canada; 2Centre for Blood Research, University of British Columbia, Vancouver, BC, Canada; 3School of Biomedical Engineering, University of British Columbia, Vancouver, BC, Canada; 4Division of Urology, The Ottawa Hospital, Ottawa, ON, Canada; 5Department of Urologic Sciences, University of British Columbia, Vancouver, BC, Canada; 6Department of Urology, Weill Cornell Medicine, New York City, NY, United States

Over 30 million men worldwide are infertile, and the most severe form of male infertility is non-obstructive azoospermia (NOA). NOA patients require andrologists to find viable sperm to proceed with vitro fertilization (IVF) intracytoplasmic sperm injection (ICSI), which often requires hours seeking rare sperm under a microscope. We evaluate the feasibility of using machine learning methods for the identification of rare sperm in microscopy images taken from a semen sample to improve IVF success rates.

We prepared samples using density gradient centrifugation to isolate healthy sperm with no debris or non-sperm cells. Sperm are stained using SYBR-14 and propidium iodide nucleic acid to be identified and then imaged using a fluorescent microscope. Images are then combined, binarized, and used as the ground truth to train a U-Net architecture using binary cross-entropy loss to segment sperm pixels. Individual sperm are identified using the watershed algorithm and evaluated through precision-recall metrics and receiver operating characteristic curves.

Unlike previous work, the model is trained on BF images with unwashed and unstained samples to mimic clinical practice. A custom metric was developed in Python to evaluate the model on sperm prediction precision and recall using nearest-neighbour, a k-d tree, and size/distance thresholding. Pilot tests were completed to optimize model performance and speed to determine the use of 10x magnification. Heavily unbalanced datasets were counteracted using weighted losses. At 10x magnification, our model achieves 91% precision and 96% recall in finding sperm in microscopy BF semen images.

Our results indicate it is feasible to use convolutional neural networks to semantically segment sperm to support andrologists for IVF-ICSI. Our custom lab protocol creates training data containing stained sperm and unstained miscellaneous cells, allowing for the first example of a real-world application of AI for assisted sperm identification.


© 2022 CUA 74th Annual Meeting