Rare sperm screening and retrieval with a domain-adaptive deep learning-enabled microwell system.

Project Summary Reproductive challenges are a growing concern worldwide, marked by a 50% decline in fertility rates over the past 50 years. Approximately 10% of couples face infertility, affecting both men and women. A condition called azoospermia, where semen lacks sperm, impacts up to 20% of infertile men. Azoospermia is categorized into obstructive and non-obstructive types, and assisted reproductive technologies (ART) are often necessary. Microdissection testicular sperm extraction (microTESE) is the preferred surgical method for non-obstructive azoospermia but relies on time-consuming manual examination of samples under an optical microscope. Furthermore, it has been shown that in some cases of non-obstructive azoospermia, rare sperm may exist in the ejaculate, but they can only be identified through laborious manual searches of large-volume samples. Such occurrences can potentially save these individuals from undergoing microTESE surgery. The need for an accurate, reliable, and consistent method for sperm identification and retrieval is crucial for improving the chances of men with azoospermia having biological children. While microfluidics and machine learning have shown potential, they face limitations. Microfluidic solutions suffer from usability and design complexity issues, while machine learning struggles with domain bias. None of the existing systems are suitable for replacing manual sperm examination, highlighting the ongoing need for a clinically viable method to retrieve sperm efficiently and reliably from testicular biopsies, aspirates, and severe factor ejaculated specimens (oligozoospermia/cryptozoospermia). To address these challenges, our proposed approach includes two components: 1. A microwell device designed to digitize cells in ejaculated samples from men with low sperm counts (cryptozoospermia) and microTESE samples. This device creates confined nanowells for efficient sperm localization. 2. A robust deep learning framework tailored to handle diverse and heterogeneous data, enabling swift and precise detection of rare sperm.