Name: Public defence: Akriti Sharma
Start: 2025-04-01
End: 2025-04-01
Location: Pilestredet, Room Athene 1, Pilestredet 46

10.00 – 10.45: Trial lecture. Title: “Current status and future perspective on multimodal AI in healthcare - with some example applications”.
12.00 – 16.00: Public defence

Ordinary opponents:

First opponent: Professor Kjersti Engan, Department of Electrical Engineering and Computer Science, University of Stavanger
Second opponent: Professor Alan Hanjalic, Department of Intelligent Systems, Delft University of Technology
The chair of the committee: Associate Professor Leiv Øyehaug, Department of Computer Science, OsloMet

Leader of the public defence: Associate Professor Norun Christine Sanderson, OsloMet

Main supervisor: Professor Hugo L. Hammer, OsloMet and SimulaMet

Co-supervisors:

Professor Michael A. Riegler, SimulaMet and OsloMet
Laboratory manager/PhD Mette H. Stensen, Volvat Spiren
Researcher Erwan Delbarre, OsloMet

Abstract

Background

Assisted reproductive technology (ART) is a leading medical procedure addressing infertility and involves the in vitro manipulation of eggs and sperm. Traditionally, embryologists manually assess embryo development and viability for transferring it to the uterus. In the Nordic countries, typically only one embryo is transferred, whereas in other regions around the world, multiple embryos may be transferred. It can be subjective and time-consuming.

Recent research studies have highlighted the potential of artificial intelligence (AI) to improve the accuracy and objectivity of embryo assessment, embryo grading and selection for implantation. Furthermore, several AI-based commercial software packages exist designed to streamline and digitize assessment workflows, yet there’s a disconnect between the clinical integration of AI and embryologists’ expectations.

Many embryologists view these AI-based solutions as ‘black boxes’ and would like these models to explain the rationale behind their decisions. While AI practitioners have focused on creating new assessment methods to enhance ART outcomes, the lack of understanding and trust among embryologists makes it challenging to adopt AI in ART. Therefore, developing AI methodologies that address embryologists’ needs and expectations is crucial for advancing ART quality and effectiveness.

Objectives

The main objective of this research is developing AI algorithms to aid embryologists in the ‘embryo assessment’ process. These methodologies would automate various assessment tasks, identify relevant patterns to devise new evaluation processes and ensure that embryologists trust and understand the outcomes.

Three research objectives are established to achieve the main objective:

How to promote trust and understanding among embryologists for AI use in ART?
What opportunities exist for integrating AI-based tools into clinical workflows to aid in
assessing embryos?
How can insights into embryo development can improve embryo assessment?

Methods

We formulated a set of AI-based methodologies addressing each objective:

We employed deep learning model architectures commonly used in research and commercial spaces to develop image and video classifiers for analyzing embryo development in ART and explained the algorithms’ learned features to embryologists.
Evaluating morphology at various embryonic development stages is a central task in ‘embryo assessment’. We developed AI-based methodologies to identify embryo cells, track cell divisions, and provide the duration of each developmental stage.
We developed a methodology forecasting the future pattern of embryo development, enabling embryologists to assess an embryo’s quality early on and make informed decisions about its use or disposal.

Highlighted Results

Embryologists successfully conducted model analysis by viewing visual representations of classifiers’ learning, and it improved their understanding and trust in predictions.
In addition to illustrating how classifiers identify key embryo morphological features, explanations helped embryologists recognize patterns that led models to misclassify embryo development stages.
The AI-based methodology for tracking embryo cell development could accurately track cells and their subsequent division up to ‘5 cells’ stage. The tool was recognized as beneficial for clinical practice by embryologists.
The cell tracking methodology was effective in locating patterns of abnormal cell divisions that are correlated to embryo quality.
The AI-based methodology for automated annotating of time duration of embryo development stages, performed the annotation task in real time, with a maximum of 1 minute for annotating a complete video.
The forecasting methodology identified specific biomarkers for assessing an embryo’s implantation potential up to 23 hours of development.

Conclusion

Our research results proved that AI has the potential to revolutionize the field of ART by providing solutions supporting embryologists in embryo assessment and selecting the embryo with the highest likelihood of successful implantation. The insights gained from this research work and the contributions made would benefit the ongoing research in the field of ART. However, the developed methodologies are validated on retrospective data from a single clinic and need to be evaluated prospectively across multiple clinical settings.