Labelling, modelling, and predicting cell biocompatibility using deep neural networks

Abstract

With an approximate annual cost of £1.7 billion for the NHS, osteoporosis and osteoarthritis alone represent a major socio-economic burden to the UK. While bisphonates and nonopioids such as non-steroidal anti-inflammatory drugs are prescribed for osteoporosis, these types of drugs may lead to serious medical complications when high doses are taken for a long time or when someone is at an advanced age or in poor health, and may even accelerate cartilage destruction in osteoarthritis. Tissue engineering is a successful alternative or additional approach, but the use of grafts and implants is not without risk, specifically the risk of rejection and failure. Therefore, there is a growing need for innovative techniques to promote implant integration and reduce the failure rate of osteopathic intervention. Additionally, at the other end of the age scale, poor placental function can compromise fetal development and placental function directly determines fetal growth. Poor fetal growth is linked with higher chronic disease rates, so leads to an increased probability of health conditions in later life. The interaction between fibroblasts, pericytes, and endothelial cells in placenta can improve understanding of how cell placement and behaviour affect placental health. To model and analyse these complex interactions, 3D visualisation is required, and 3D-labelling is therefore a necessity to the medical imaging field, which can take months of researcher-time. Similarly, the experimentation required in tissue engineer, from in-vitro to ex-vitro, can be equally lengthy and complex, as cell response to biochemical and biophysical cues remains poorly understood. The three areas of research in this project include labelling cells for 3D visualisation, modelling cell response to biophysical cues, and predicting biocompatibility of tissue engineering scaffolds. The data used for these approaches includes 3D nanoscale-resolution images of placenta for labelling and images of 3D bioengineered scaffolds for biocompatibility analysis, which were provided by collaborators. Data also included images of stem cells cultured on topographically-varied surfaces, to analyse stem cell response to biophysical cues, which was collected for this project. This data was used to train multiple deep neural networks, with the goal of applying deep learning to label cells in placenta, and both model and predict cell responses to biophysical and biochemical cues. It was found that deep neural networks can be used to replace labour-intensive manual labelling, with automated labels comparable, pixel-to-pixel, to manual labels by over 98% on average. The response of cells to physical cues can also be modelled by a deep neural network. With a probability of 𝑃 < 0.001, it can therefore be used as a model, with potential implications for tissue structure development and tissue engineering. Deep learning can also be used to predict biocompatibility, which may act as a future replacement for animal models in place of traditional computer modelling. When predictions were compared with experimental results, images displayed excellent agreement. Tissue engineering is an increasingly important area of regenerative medicine, partly due to aging populations around the world, combining cell biology, bioengineering and clinical research. However, focus on health from much earlier in life, such as through analysis on placental tissue, may also aid in overcoming the high chronic-disease levels in old age. Applying deep learning to regenerative medicine research may not only help increase efficiency of 3D-image processing, but also potentially help increase understanding of stem cell behaviour and reduce levels of necessary animal testing in research. In the future, application of deep learning to regenerative medicine could help increase quality of life in our later years.

More information

Identifiers

ORCID for Benita Scout MacKay: orcid.org/0000-0003-2050-8912

ORCID for Benjamin Mills: orcid.org/0000-0002-1784-1012

Catalogue record

Export record