Deep Learning Assisted Fourier Transform Imaging Spectroscopy

Abstract

Fourier transform spectroscopy (FTS) has been combined with fluorescence microscopy to allow for high-throughput screening using multi-color probes. In principle, FTS is built on an interferometer, and the intensities of interferograms recorded for varying optical path differences (OPD) provide the emission spectrum of the fluorophores. Here we use deep learning to reduce the sampling number, and thus to increase the data acquisition speed of FTS-based fluorescence microscopy. Even though compressed sensing has been demonstrated to reduce the sampling number, the deep-learning-based approach is able to classify the types of fluorescent dyes without reconstructing the emission spectrum; thereby, it can further reduce the required sampling number. Further, using deep learning, we aim to demonstrate a robust classification without relying on the laser interferometer, which is typically installed in parallel with the main beam path to monitor the actual OPD. We use a 1-D convolutional neural network (1DCNN) together with weight decay regularization and ReLU activation in the hidden layers. For the classification, we use a categorical cross-entropy loss function and an optimization algorithm with adaptive learning rate. For the proof of concept, we simulate 10 fluorescence emission spectra with close emission peaks, then show the NN can distinguish all the different types with 95% accuracy from about 1/10 of the interferograms typically required in FTS. We also experimentally demonstrate our method using bovine pulmonary artery endothelial (BPAE) cells labeled with three fluorophores. Our approach may lead to compact, fast, robust FTS-based fluorescence microscopy.