First denoise the image. This could be done in a variety of ways, from the simple box filter and median filter to more sophisticated methods like anisotropic diffusion, BM3D, non-local means, K-SVD, K-LLD or the like.
Then process image to a point where you can threshold it. This may require locally adaptive methods like stdfilt, adapthisteq, or others depending on what the image looks like.
Then threshold, label, and call regionprops(). Then look at the perimeter^2/(4*pi*area) ratio to pick out those that are irregularly shaped, as opposed to smoother.
Posting an image would get you better recommendations.
