Abstract
Fluorescence Lifetime Imaging Microscopy is a highly sensitive analytical tech-nique used to monitor real-time metabolic events and cellular biochemistry by mea-
suring the decay time of fluorophores like nicotinamide adenine dinucleotide (NADH).
Applying this technique to live cells requires low laser power and short acquisition
times, resulting in low-photon datasets with significant Poisson noise. Traditional
analysis methods struggle to fit double exponential decays at low photon counts and
are computationally expensive. To address these limitations, this thesis introduces
a Locally Linear Embedding Bi-Exponential Autoencoder built with 1D ConvMixer
blocks to rapidly and accurately extract bi-exponential decay parameters.
The convolutional neural network compresses a 256-bin temporal decay histogram
into a concise three-dimensional latent space representing the fast lifetime, slow life-
time, and bound fraction. To train the network, a MATLAB-based synthetic data
pipeline was developed. This pipeline merged realistic cellular morphology from the
Human Protein Atlas with physically accurate simulated multi-exponential decays
and simulated noise. The model was subsequently evaluated against various simu-
lated perturbations, demonstrating high resilience to flat background noise and tri-
exponential decay signals, though it exhibited sensitivity to temporal misalignment
and extreme photon starvation.
Finally, the autoencoder was tested on experimental in vitro squamous cell car-
cinoma datasets and compared to the industry standard double exponential fitting.
The model achieved excellent agreement with traditional methods for the short life-time parameter, especially at higher photon counts. For the long lifetime and bound
fraction, while accurately capturing the slope of the values, there is a systematic
difference in the model prediction and least squared fit. The results confirm that
this deep learning architecture could provide a robust and computationally efficient
alternative to traditional curve fitting. Future work would go into establishing a
true ground truth comparison with traditional methods, and testing different model
configurations to improve systematic differences.