Enhancements on Fluorescence Images using Machine Learning

Fluorescence microscopy has enabled a dramatic development in modern biology, and the output of conventional fluorescence microscopy is the intensity. However, it is hard to segment a sample with multiple cells using intensity. Specifically, fluorophores have overlapping emission spectra, and these cells cannot be segmented with only intensity information. Also, it is sensitive to the applied laser power, movement of the animal, and temperature. In this scenario, one can use fluorescence lifetime image (FLIM) of excited fluorophores, which enables the segmentation of the cells and also provide the vital information such as the ion concentration, the dissolved oxygen concentration, the pH, and the refractive index, which are the micro-environment in living tissues. FLIM methods are generally divided into two categories, namely, time-domain (TD) FLIM and frequency-domain (FD) FLIM. Time-domain consists of two methods: time-correlated signal photon counting (TCSPC) and time-gating (TG). However, the aforementioned methods suffer from either slow in computation time since more pulses are required to extract the lifetime information or additional hardware requirement, which is used for the phase measurements of an intensity-modulated excitation and emission signal. Therefore, we present a novel convolutional neural network (CNN) based approach to estimate the lifetime image from the intensity information. In this work, we create a dataset that consists of intensity and lifetime images. We train our model by considering the composite lifetime (HSV image: where the intensity and the lifetime are mapped to the brightness and hue, respectively) as the target image. The results show that the predicted lifetime image has less noise when compared to the ground truth image.