Machine Learning Methods for Fluorescence Lifetime Imaging (FLIM) Based Label-Free Detection of Microglia

Abstract

Automated computational analysis techniques utilizing machine learning have been demonstrated to be able to extract more data from different imaging modalities compared to traditional analysis techniques. One new approach is to use machine learning techniques to existing multiphoton imaging modalities to better interpret intrinsically fluorescent cellular signals to characterize different cell types. Fluorescence Lifetime Imaging Microscopy (FLIM) is a high-resolution quantitative imaging tool that can detect metabolic cellular signatures based on the lifetime variations of intrinsically fluorescent metabolic co-factors such as nicotinamide adenine dinucleotide [NAD(P)H]. NAD(P)H lifetime-based discrimination techniques have previously been used to develop metabolic cell signatures for diverse cell types including immune cells such as macrophages. However, FLIM could be even more effective in characterizing cell types if machine learning was used to classify cells by utilizing FLIM parameters for classification. Here, we demonstrate the potential for FLIM-based, label-free NAD(P)H imaging to distinguish different cell types using Artificial Neural Network (ANN)-based machine learning. For our biological use case, we used the challenge of differentiating microglia from other glia cell types in the brain. Microglia are the resident macrophages of the brain and spinal cord and play a critical role in maintaining the neural environment and responding to injury. Microglia are challenging to identify as most fluorescent labeling approaches cross-react with other immune cell types, are often insensitive to activation state, and require the use of multiple specialized antibody labels. Furthermore, the use of these extrinsic antibody labels prevents application in in vivo animal models and possible future clinical adaptations such as neurodegenerative pathologies. With the ANN-based NAD(P)H FLIM analysis approach, we found that microglia in cell culture mixed with other glial cells can be identified with more than 0.9 True Positive Rate (TPR). We also extended our approach to identify microglia in fixed brain tissue with a TPR of 0.79. In both cases the False Discovery Rate was around 30%. This method can be further extended to potentially study and better understand microglia's role in neurodegenerative disease with improved detection accuracy.

Keywords: CNS; FLIM; NAD(P)H; brain metabolism; machine learning; microglia; neural networks.