Towards robust bubble detection in diverse microphysiological systems by machine learning – Nadezhda Koriakina

Microphysiological systems (MPSs) provide the capability to continuously monitor cell cultures for an improved understanding of various biological phenomena. A common problem, however, is the formation of air bubbles that can affect the experiments. Receiving timely notifications about bubble formation in MPSs can be beneficial for an experimenter, as early detection makes it possible to eliminate the bubble or direct it into a waste outlet.

Brightfield (BF) microscopy imaging is a common method for monitoring MPSs, and it also allows for the observation of bubbles. However, if bubbles are to be monitored visually using BF images, the observer must remain vigilant and continuously monitor the situation. Modern machine learning (ML) techniques offer both automation and accuracy. Bubble detection using ML could be useful for both offline post-processing and also become part of an online system for bubble notification in MPSs. Despite the potential, there are no studies focused on applying ML models for bubble detection in MPSs. This absence can be attributed to the lack of sufficiently large datasets from MPSs, as these systems are often used in research environments where it is challenging to collect large amounts of data from the same device or under consistent settings. We explore whether combining BF microscopy data from different MPSs could enable the development of a versatile ML model for bubble detection in MPSs. Considering that training of ML models usually requires a substantial amount of data, we utilise transfer learning approach as a strategy for learning from small datasets.

About Nadezhda Koriakina