Principal Investigator: Hadi Shafiee
Identification of the structure and the features of human sperm cells still continues to depend heavily upon manual image based assessment in laboratories as the other available technologies are either too expensive or too inaccurate for clinical cost-effectiveness. Supervised deep learning based methods have shown promise in automated feature and structure based differentiation of the human sperm but they also have been restricted to only high quality images acquired from expensive and bulky imaging hardware. Here, we show that adversarial learning can be used to develop high-performing networks which are trained on unlabeled medical images of varying image quality specifically using low-quality images obtained using cheap ($5) and portable optical systems to train networks for the quantification of human sperm morphology. Adaptive adversarial networks may expand the usage of authenticated neural-network models for the assessment of data collected from multiple imaging systems of varying quality without negotiating the knowledge stored in the network beforehand. The results can potentially allow for improved access to and regulation of care even in settings with limited resources.
Clinical semen analysis continues to depend heavily upon manual image-based sperm morphology assessment, as all of the proposed alternative technologies have been either too expensive or too inaccurate for clinical cost-effectiveness. While supervised deep learning-based approaches have shown promise in automated morphological analyses of sperm, they have been limited to only high-quality images resulting from expensive and bulky imaging hardware. Here, we show that adversarial learning can be used to develop high-performing networks trained on unannotated medical images of varying image quality specifically using low-quality images acquired using inexpensive ($5) portable optical systems to train networks for the quantification of human sperm morphology. Adaptive adversarial networks may expand the use of validated neural-network models for the evaluation of data collected from multiple imaging systems of varying quality without compromising the knowledge stored in the network. The findings could allow for improved access to and standardization of care even in resource-limited settings.
If the PDF viewer does not load initially, please try refreshing the page.
