How best to approach generating in silico 'fake' representative images including speckle noise?

I am trying to generate a set of de novo Optical Coherence Tomography retinal B scans (only of the vitreous, NOT the retina) that form a reasonable approxomation to the real images aquired by our system. I would like to re-create pseudorandom sample images with near-identical statistical properties (spatial and intensity) of the expected real background, including the speckle noise. I have thought of a lot of approaches and it seems reasonable to analyse manually chosen regions of selected images taken with our machine, and use these statistics to populate entirely separate and de novo greyscale matricies. If this is acceptable, I am not sure how to A) extract relevant information from the real images or B) use these perameters to create a new, synthetic image.

This is the first step before I spike in objects / features of interest, but I'd like to get this right first. If the final result is good enough I'd like to use these fake images for optimisation / training purposes.

(I am new to MATLAB, and to programming as I've come from a biomedical background, so I apologise if my question is non specific and clumbsy.)

Thank you

[Daniel M]

Hmm, maybe I don't fully understand the situation, but what about this idea? Take a set of "training" images and take the fourier transform of them. Then try taking different linear combinations of your magnitudes (with a weighted mean) and/or randomly rearranging phases, and do an inverse transform. Maybe...?

It also might just help understanding the characteristics of your images by looking at the frequency domain as opposed to the spatial domain.