Parameter Estimation in FRAP using Deep Learning
Typ
Examensarbete för masterexamen
Program
Complex adaptive systems (MPCAS), MSc
Publicerad
2019
Författare
Wåhlstrand Skärström, Victor
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Fluorescence recovery after photobleaching (FRAP) is a method used in optical microscopy
for determining properties of diffusion in organic and inorganic solutions, including cells,
membranes and gels. FRAP may be used to determine parameters such as the diffusion
coefficient and binding rates of particles in a sample, and is used in wide-ranging disciplines
such as the medicine, soft materials and food science. In FRAP, fluorescently labelled
particles in a sample are bleached in a region of interest using a high-intensity laser.
The recovery of the mean fluorescence intensity in the region of interest is known as the
recovery curve. Conventional methods for inference rely on least squares and models for
the recovery curve, but recent work has come to use the entire spatio-temporal image data
for estimation.
In this work, we have implemented a set of deep neural network architectures for estimating
parameters such as the diffusion coefficient in FRAP. This is to our knowledge
a novel approach with some potential advantages over conventional methods. We have
implemented a set of both spatio-temporal and purely temporal neural network models,
where operating on the full image data gives the best performance in terms of error on
simulated data. The downsampler neural network model is easy to implement and parallels
the extraction of the recovery curve, and can be trained from numerically simulated
data. We show that the downsampling neural network can be trained on limited computational
resources, using the combined power of continuously generating training data and
batch-mode optimization. The neural networks demonstrate a robustness against noise
and computational speed unlike the conventional least squares methods.
The performance of the neural networks versus the conventional methods is tested on
simulated FRAP data and finally validated on experimental data, yielding good agreement
with the expected values of the parameters and those obtained from the conventional
methods.
Beskrivning
Ämne/nyckelord
confocal microscopy , diffusion , fluorescence recovery after photobleaching , deep learning , machine learning , neural networks