A classification toolbox for the analysis of spectral microscopy images.
Many microscope systems equipped with spectral detectors come packaged with software for linear unmixing. These linear unmixing functions solve the least squares problem to decompose your spectral microscopy images into component channels based on a set of predefined reference spectra. Although linear unmixing is great for separating signals that overlap within a given pixel (ex. autofluorescence + fluorescent tag), the least squares equation is not well suited for classification problems. Techniques such as Combinatorial Labeling and Spectral Imaging Fluorescence in situ Hybridization (CLASI-FISH) use unique sets of fluorescent tags to identify spatially disparate targets of interest. Since the fluorescent signatures are not overlapping within any given pixel, we need only classify each pixel based on its observed spectrum.
The Spectra Learner toolbox provides a convenient way to analyze raw spectral microscopy images without the need for complex linear unmixing pipelines. Spectra Learner builds a classification model based on a set of reference spectra, and uses the learned model to accurately and efficiently predict the identity of unknown pixels within a raw spectral microscopy image. The Spectra Learner toolbox allows you to classify your images by simply dragging and dropping your files into the appropriate directories. Spectra Learner has been designed to handle sets of images from sequentially acquired spectral microscopy images. If you don’t mind getting your hands dirty, the modular functions of this toolbox can be modified for a tailored analysis routine.
There are many elements to spectral microscopy experiments, but techniques leveraging the spectral dimension can be extremely powerful. I have provided some tips below to get the most out of this toolbox and your experiment.
Spectra Learner was written in MATLAB 2016a, and has been tested on unix-based systems. This toolbox requires the following MATLAB toolboxes: Statistics and Machine Learning, Image Processing, and Parallel Computing.
There are two basic modes for a spectral microscopy experiment: simultaneous scan and sequential scan. In simultaneous scan, one or multiple lasers are used simultaneously to acquire one spectral image. In sequential scan, one or multiple lasers are used to capture a sequence of spectral images. Simultaneous scanning is useful when you have only a few fluorophores to classify, and those fluorophores do not have emissions peaks near the wavelengths of the dichroic mirrors. Sequential scanning is useful when you have many fluorphores to classify, or fluorophores have emission peaks on or near the wavelengths of the dichroic mirrors. Since sequential scanning results in separate spectral image files, a particular naming scheme must be followed for the toolbox to combine the correct data.
The Spectra Learner classification model makes predictions by first learning all possible signatures during the training phase. The user provides reference images of each spectral signature. Reference images should be acquired under similar physical conditions to those of the experimental images. The reference images are assumed to be grayscale TIFF stacks with a black background.
The resolution of your reference images will impact the duration of the training phase. A good balance between speed and accuracy can be obtained with images at 256x256 resolution. Higher resolution images provide more observations for training the model, but will increase the overall time of training.
Reference images must follow a particular naming scheme. Each reference image should be names using the fluorophore and the scan identifier separated by an underscore. For example, a reference image for Atto 488 using the 488nm laser would be named Atto488_488.tif. The next sequential scan using the 514nm laser would be named Atto488_514.tif. It is not necessary to have all possible sequential scans for each fluorophore. Any missing scans will be filled with zeros.
Reference images should be placed in the reference directory
Spectra Learner expects all raw experimental images to maintain the same properties (i.e. dimensions and bit depth). The images should be taken under similar conditions to the reference images to increase the accuracy of classification.
Spectra Learner can process batch files that follow the proper naming scheme. Each file name should consist of the group identifier and the scan identifier separated by an underscore. For example, here are the file names for two groups of files acquired in a sequential scan mode.
img1_488.tif
img1_514.tif
img1_561.tif
img1_594.tif
img1_633.tif
img2_488.tif
img2_514.tif
img2_561.tif
img2_594.tif
img2_633.tif
Raw experimental images should be placed in the raw directory.
For each group of raw files, Spectra Learner will produce two files: a JPG color preview and a TIFF stack. The first channel of the TIFF stack is the background channel trained on the background pixels of the reference images.
Valm, A.M., Welch, J.L.M. and Borisy, G.G., 2012. CLASI-FISH: principles of combinatorial labeling and spectral imaging. Systematic and applied microbiology, 35(8), pp.496-502.