Blind Spectral Unmix with NMF ( IANMFUnmix )

Calculates unmixed fluorescence intensities using nonnegative matrix factorization (NMF).
If images contain saturated pixels mask them right after loading. Optionally use Inpaint Mask to fill in masks after unmixing.
Subtract background before unmixing. It is also advised to perform subpixel alignment before unmixing.
Masking background may improve convergence.
Crossbleed coefficients are iteratively calculated from a raw starting value.
Either uniform starting values are used, or if the Starting Crossbleed is set to zero then the Starting Coefficient matrix is used.
Coefficient matrix: each row corresponds to a channel, and tells that which fluorophores (columns) have how much contribution to this channel. The diagonal of the matrix contains ones for those channels which define a fluorophore.
There can be fewer fluorophores than channels, in this case the last one or more columns contain only zeros.
Use Preferences/Convolution kernels, Matrices to define the Coefficient matrix.
Use small (0.01-0.05) starting value, large values less likely converge to the correct result because of clipping initially erroneously unmixed pixels below zero. However too small starting values will be lengthy to converge.
The resultant coefficient matrix can be saved and used with the Spectral Unmix function.
Constrain to active ROI: Only pixels of the active ROI, from all frames are used for coefficient calculation.
Using the Starting Coefficient Matrix: Define a matrix of identical size to the spectral unmixing matrix (number of probes wide and number of channels height) in the Preferences dialog. Each row corresponds to a channel, and tells that which fluorophores (columns) have how much contribution to the channel. The iterations will start at these values. Provide here an underestimate of the value of crossbleed.
The matrix can be defined as Convolution kernels in the Preferences dialog, or optionally define it here as ={{1,"probe 2 to channel 1",..},{"probe 1 to channel 2",1,..},....}.
In some cases convergence reached in 1000 iterations, some cases 100000 iterations are required.
Fixed coefficients: By defining the "Fixed Coefficient Matrix" at Preferences/Convolution kernels, coefficients can be marked by "1" that are not changed during iterations from the initial value. Use this to lock known coefficients, and calculate others.
The algorithm is based on Eqs. A4 and A7 in Neher et al. 2009 Biophys J. 96:3791