Attached is a pipeline that might be usable. The image is pre-processed by illumination correcting each channel before proceeding. Since brown is a mixture of red, green and blue, the problem is finding which of these, if any, be used to constitute brown. I have found that brown pixels often have a red intensity greater than the blue or green. So, by dividing the red channel by the blue channel, the intensity of the brown pixels can be enhanced (red/blue> 1) and the non-brown regions (including the unstained tissue) can be suppressed.
ClassifyObjects will permit the binning of the objects into brown/not-brown based on this divided intensity. Right now, the cutoff is set to 1; you will need to adjust it to find what the proper value should be; you can look at the outlined images and see the results. Keep in mind that whatever you adjust the cutoff value to be in ClassifyObjects, the same value must be set as the minimum value in FilterByObjectMeasurement.
That said, the quality of the image probably needs to be improved before this sort of segmentation or color classification is successful. From the sample image, it is not clear what the nuclei are; the brown seems to be staining larger regions of tissue than the the black (green?) specks (which I assume are the nuclei). The image also seems to be fairly low-resolution; perhaps a higher-magnification image which better resolves the nuclei might be in order?
2009_07_16_PIPE.mat (2.05 KB)