September 2022 Ravata Newsletter
Introduction
Thank you for reading Ravata’s monthly newsletter where we aim to keep our friends, customers, and collaborators up-to-date. In this issue, we will cover Ravata’s experience with computer vision, SBIR results, and the availability of Pilots nationwide.
Join our newsletter mailing list for updates here!
Ravata National Institute of Health SBIR Results
In collaboration with the University of California Davis Mouse Biology Program and the National Institute of Health, Ravata Solutions developed and tested a platform for the mechanical assessment of embryo quality using time lapse microscopy. The goal of this work was to identify mechanical stress coefficients between viable and non-viable embryos at the zygote stage of development.
Embryos were aspirated using Ravata’s pneumatics and microfluidics designs, cultured to development individually, and annotated with computer vision derived landmark tracking. Over 500 cells spanning 8 different data collection days were used to identify the accuracy of such a methodology. The results are promising however are subject to similar issues plaguing computer vision in IVF applications.
In figure 1, all data was aggregated and subsequently split into both train and test datasets. The training data in this case comprised some cells from every experiment and the test data included some cells from every experiment. The results indicate very strong predictive power in embryo selection achieving an accuracy close to 99%!
Figure 1
In figure 2 training and test data sets were taken from different runs entirely which occurred on different days. The results here show only a 60% predictive power, slightly better than the flip of a coin and about where published performance for computer vision with embryo selection stands today.
Figure 2
Challenges in Computer Vision for IVF
The main issue with computer vision in IVF is the lack of annotation, and inability to measure for proper annotation, all the various differences between experiments that are innate in a laboratory setting. These differences include the effect of different osmotic gradients due to variations in media and cell handling which will affect the size of cells, differences in the development stage and time of day an embryo is assessed compared to what the model is trained on; seemingly minor differences in image quality due to lighting, focus, and image capture devices that have a major effect on the results; and rotation of a 3 dimensional spheroid embryo as captured by a 2 dimensional image.
While results can be presented as very strongly predictive, the methodology used to arrive at those results needs to be questioned before a computer vision model based product can truly be applicable across multiple laboratories with varying standard operating practices.
Several companies have attempted to source and use vast quantities of embryo development microscopy videos to bring a computer vision model to market. Currently a “state of the art” computer vision model, including from Invitrolife which produces the industry standard for time lapse microscopy incubators, had a prediction accuracy of 63% to 69%.¹ Their products have been on the market for over 20 years however they still have not brought to market a complementary tool to their incubators that would enable predictive embryo selection. Others in the space also include Presagen, AiVF, Embryonics, and others in varying degrees of development.
Technical Advantages of Impedance Spectroscopy over Microscopy
Microscope images require white light to be applied to cells and the resulting change in how the light interacts with the cell is captured via a black and white images. Due to this approach, the data being generated by an embryo is a one dimensional response to a fixed stimulus. The effect of light is known to also cause DNA damage resulting in lower pregnancy rates.²
Impedance spectroscopy alternatively can be used simultaneously at multiple frequencies to capture a broader spectrum of cell responses giving orders of magnitude more data per cell than traditional microscopy. This access to higher density of data per cell will be more effective in training machine learning algorithms due to the higher potential for identifying obscure patterns. Furthermore, signals up to 2Ghz have been tested on cells to cause no increase in DNA damage.³
Pilots Shipping Now!
Ravata Solutions has overcome a range of productization challenges to scale its deployment of devices to customers and academic partners. We are now ready to begin assessing embryo quality on-site across the country! Contact us here to receive more information on the Pilot and check out our website for more information.
Embryologist Input Wanted!
We are surveying the assisted reproduction landscape to identify the needs of embryologists, lab techs, and directors nationwide regardless of the mammalian species they work with. Please help us gain traction in this endeavor by forwarding our survey to anyone you believe fits the demographic. Survey results will be shared with respondents and via newsletter!
Regards,
Ravata Team
References
Berntsen, Jørgen, et al. Robust and Generalizable Embryo Selection Based on Artificial Intelligence
and Time-Lapse Image Sequences. PloS One, Public Library of Science, 2 Feb. 2022,
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8809568/.
Campugan, Carl, et al. The Effect of Discrete Wavelengths of Visible Light on the Developing
Murine Embryo. Journal of Assisted Reproduction and Genetics , 15 June 2022,
https://link.springer.com/content/pdf/10.1007/s10815-022-02555-4.pdf.
Miyakoshi, Junji, et al. Effects of High-Frequency Electromagnetic Fields on DNA Strand Breaks.
Research Gate, Electrical Engineering in Japan, Wiley, Dec. 2002,
https://www.researchgate.net/publication/229925768_Effects_of_high
frequency_electromagnetic_fields_on_DNA_strand_breaks_using_comet_assay_method.