Is the image that you posted already a subtraction image, or do we need a second image? Could you draw (on another image) what you want to be segmented?
If you intend to segment the cerebral aneurysms only, you may follow these steps:
1. resize your image to make it smaller (if possible).
2. apply filter(s) for smoothness and sharpness purposes (enhance its contrast).
3. define your ROI,
4. Apply Region Growing method on your defined ROI.
If these steps did not work well on your images, you might want to use better quality images with better contrast. Let me know about the results to suggest further or different approaches.
I would start with a Gaussian filter to get rid of some noise.
Then fit a higher order polynomial to the pixel-values that have a small variance in their local neighborhood, and subtract that from the image. This step will make the image more homogeneous (see the inhomogeneity in the picture below).
For the polynomial fit you need to consult the documentation of the image-processing package you use.
Another way is to smooth the image with an 'appropriate' kernel-size, and subtract that from the original image. This technique is called unsharp-mask. I included the result for kernel-size 13 without taking boundaries of the field of view into account.
A third way is to go to the frequency domain (Fourier transform), filter out low frequencies, and go back to the spatial domain.
I've read about the pre-processing that the best way to enhance the inconsistency and inhomogeneity in DSA images is to use Anti-concentration Diffusion model ... but I need some more references and details about this technique ..
@ Lambert, I'll inspect "visually" all the data which I have and try to choose the one without motion artifacts.
I was thinking nearly the same way about how to homogenize the image, by first applying Gaussian smoothing to the first or last slice ... then finding the average of rows and subtract it from the corresponding ones in the vessel image ...
Better use the first (pre-contrast) slice. The last slice still shows contrast agents accumulating in veins.
I don't understand why you would want to subtract row-averages; just subtracting the whole pre-contrast slice is the best, even when the patient moves a little.
Maybe fixing the patient's head is too expensive for daily clinical routine. However, for research purposes... Ask the radiologist to be kind for once. The head-fixation does not need to be fancy, some wooden or plastic construction with Velcro belts is probably fine. Just don't use high density materials.
Have the images already been deconvolved with the point-spread function?