I have an image processing problem that I need help with. I am looking at cell-size and morphology in a cell-structure and although my problem does not directly concern any particular MatLab functions, it relates to methods that are easily implemented in MatLab (which is the software that I use) and probably someone here can give me some useful and general, or specific advice. To help illustrate my problem I have put some images on this webpage <<a href="http://www.homepages.ucl.ac.uk/~smgxro1/page1.html">http://www.homepages.ucl.ac.uk/~smgxro1/page1.html</a>> and I will be referring to those images below.
The cell-structure is imaged with a fluorescent Microscope at 40x. To determine the individual cell size I look at a region-of-interest (ROI) in the Frequency domain where the peak spatial frequency can easily be converted into the average cell-size (top-right corner of image 4 and 5). The tissue sample is around 5 cm square and it is quite a delicate process to prepare and flat-mount it (image 1). To get the Fourier transform to register a distinct peak frequency I use the following pre-processing steps (results can be seen on the left of image 4 and 5): 1) Blind deconvolution - to improve image resolution and compensate for out-of-focus areas of the flat-mounted tissue. 2) Uniform local histogram equalisation (CLAHE) - to optimally enhance image contrast. 3) Gaussian low-pass filtering - to eliminate low-frequency changes (due to uneven illumination, tissue thickness etc.).
Once I have the peak frequency of a given ROI and I know the image px/µm ratio I can derive the mean cell-size. Finally, I plot the mean cell-size and the standard deviation against the distance from the centre of the tissue (the cell-size increases with distance from centre). In the centre of the tissue the cells are tightly packed in a hexagonal lattice, which makes it ideal for an isotropic Fourier analysis (image 2 and 4). However, towards the extreme periphery several morphological changes take place: The cell shapes becomes severely deformed, the cell contents change (and hence, the pixel brightness, edge gradients etc.) and none-RPE cells are mixed into the lattice. Image 3 illustrates but one of countless and random cell configurations and as a consequence the frequency spectra becomes even harder to evaluate (top-right of image 5).
These are my two questions: 1) How can I improve on the pre-processing of the microscopic images to enhance the cell-walls, or edges and eliminate other non-continues cell artefacts? 2) The mean cell size and it's standard deviation give a very accurate measurement of the central and highly regular lattice, but when the cell structure becomes more irregular, it is less meaningful. Are there better spatial descriptors - and methods to measure them - for the random cell shapes that I encounter at the periphery of the tissue?