When reading large tif images for the jaw localisation, only read the relevant pixels and pages

[richard-lane]

We dont need the whole 2000x500x500 pixel tiff stack to locate the jaw, we can (in principle) do it with a much smaller, downsampled version of the image. At the moment, (in my testing notebook at least), i read in the full-sized tifs and then downsample them (with a bit of anti aliasing)

it may be sufficient, and would probably be faster, to read only the required slices in

• Check if this is actually significantly faster for a single image (if not, dont bother)
• change the function to do that
• check if the model still works well with the new downsampled images