Ultrasound volumes registration

Dear Imfusion support,

I am testing different optimisers for registering ultrasound volume pairs and they typically only have a partial overlap. I am performing a rigid registration and using the NCC loss. This is an example of output obtained:
Nelder-Mead simplex algorithm (local, no-derivative)
Adding Zero-Mask to 1mov_2_vol9_30deg
Adding Zero-Mask to fixed_image 7
509 evaluations in 2.87387 seconds, parameter tolerance
-22.4134 -2.68474 -0.296537 2.20769 -3.97804 -0.786304 → 0.989506 (0.603707)
I was wondering which formulation of the NCC loss is implemented in the software please. Is the masking used to remove the zero pixels part of padding? What are these 2 values obtained 0.989506 (0.603707)? Thanks a lot for the support! Best, Maria

Hi Maria,

Indeed, the zero-mask is automatically added as the SDK detects the images’ modality and assumes background (zero-valued) pixels should be ignored. You can either turn this off or assign some other mask if you’d like before running the registration.

All pixels in the non-masked overlap region are used to compute the similarity measure. In case of NCC, we just follow this paper: http://scribblethink.org/Work/nvisionInterface/nip.pdf.

The log output is formatted like this:
all optimized parameters → optimal similarity (initial similarity),
so in your case, the optimizer improved from 0.60 to 0.99.