ESHRE 2021

How machine-led approaches can improve outcomes in ART

Nikica Zaninovic: The key to future IVF is the combined decisions of humans and AI.

Published 30 June 2021

The potential of artificial intelligence in many of the IVF lab’s processes was a recurring theme in some of the 1300 abstracts submitted for inclusion in this year’s Annual Meeting. In an invited session embryologist Nikica Zaninovic from Cornell summarised how AI is already changing the way babies are made.

The future of IVF lies not solely in replacing human decision-making with artificial intelligence (AI) but in combining the benefits of both effectively. This was the vision of Nikica Zaninovic in his presentation on how machine-led approaches could – and already are – changing the way babies are made.

Much hype has surrounded AI, but, he said, there is already emerging evidence to back its potential in improving IVF success rates. A recent study co-authored by Zaninovic, an embryologist from Cornell University in the US, found that an algorithm developed by the research group was able to classify blastocyst quality with 97% accuracy.(1) The results were based on AI analysis of 12,000 time-lapse embryo images, a technique also used by Imperial College London in the UK in partnership with Zaninovic. This work showed a high accuracy rate (77%) in predicting whether a viable embryo would lead to a live birth or miscarry by using a convolutional neural network, an image recognition and classification algorithm.

Objectivity, standardisation, precision and big data analysis are among the benefits of AI in reproductive medicine, said Zaninovic. In IVF, treatments are based on experience and knowledge, yet also on trial and error. What AI can do is emulate the human brain, which performs tasks using neural networks, and thereby improve performance over time as the machine-led system is exposed to more data.

Zaninovic proposed that AI can ‘sense, reason, act and adapt’ using algorithms in an efficient way. He demonstrated how it ‘sees’ the embryo through features such as shapes, shadows and textures. Not having to fertilise every oocyte, ‘only competent ones’, is a very exciting prospect in his opinion.

The abilities of AI, however, are not just limited to embryology. They can also be applied in clinical settings such as in evaluation and suggestion of optimal stimulation protocols. Overall, this could be applicable in other applications in ART, such as donor recipient matching; indeed, some clinics are already using AI-based facial recognition technology to screen applicants so the child will resemble the would-be parents.

Inevitably, products driven by AI are already on the market - such as machines to perform automatic sperm assessment, tracking and selection by morphology. The first AI software analysis tool to assess and grade oocytes non-invasively and their probability of reaching the blastocyst stage has already been developed - and nicknamed ‘Violet’. Cornell’s own computer has a more macho title – ‘the Beast’ - although Zaninovic pointed out that more modest-sized kit such as cell phones can also be deployed.

Nevertheless and despite the commercialisation, AI has limitations. Quality and diversity of data to train algorithms, standardisation of methods and universal use were among those listed by Zaninovic, as well as the accuracy of the training set's classification for supervised learning techniques. AI is only as good as the quality of the data on which its predictions and decisions are based, he said. Problems occur if the data are not ‘clean and correct’ and clinicians also need to be clear about the exact task they want machines to achieve. ‘You’re not getting the full picture if you only want (to predict) live birth rate.’

Multiple models of AI are in existence which means there are no standardised methods for applying the technology, and widespread implementation is not yet feasible because AI is not available in every laboratory. Furthermore, critics have already questioned the overall accuracy rate (70%) of AI in embryo ploidy image analysis, although Zaninovic summed this up as ‘not bad’, especially as PGT-A itself is not totally accurate.

So he emphasised that AI-based systems in IVF are not a ‘magic egg’. The path to success potentially lies in combining AI with other non-invasive technologies such as time-lapse morphokinetics. And echoing earlier comments on the importance of quality data, Zaninovic called for information such as images of embryos to be shared publicly.

1. Khosravi P, Kazemi E, Zhan Q, et al. Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization. NPJ Digital Medicine 2019; doi.org/10.1038/s41746-019-0096-y

Get notified of new articles with our ESHRE newsletter.

Sign up and never miss an update