Abstract
Fluorescence Lifetime Imaging Microscopy (FLIM) is essential in biomedical research, providing insights into cellular mechanisms and disease diagnosis by measuring the time molecules remain in an excited state before emitting a photon. Analyzing FLIM data is challenging due to the complex, computationally intensive curve fitting required, especially with large datasets or in noisy conditions. These data may be analyzed by FLINET, a deep learning architecture designed specifically for lifetime parameter prediction. The goal of this study is to train the existing FLINET model on synthetic data that best represents FLIM images on UMSCC74A cells exposed to different mitochondrial inhibitors and uncouplers.
The first part of this study ensured that the models perform well on synthetic datasets. Models were trained and evaluated on their ability to predict lifetime parameters (τ₁, τ₂, A₁, A₂, and A₀) of images that weren’t part of its training dataset. Also, we studied how model performance was affected by different training data modifications, the number of epochs, and loss weight tuning. The second was to use the best performing model on the synthetic dataset to predict the lifetime parameters of UMSCC74A cells and compare to fittings.
Results from this study revealed that lifetime predictions from our models match fairly with traditional lifetime fittings predictions for most lifetime parameters except the offset. Samples with uncoupled mitochondria (FCCP) consistently show the highest structural similarity index measure (SSIM) values across all parameters, suggesting the model performs best under these conditions. Control samples demonstrate stable, intermediate SSIM values, while inhibited mitochondria (Rotenone) samples consistently show the lowest SSIM values, particularly for τ₁ and A₀, indicating reduced model reliability under these conditions.