$\mathbf{W_{s}}$ Weighting Experiments

I then changed the 3 stages at the start to work in a coarse-to-fine approach. This was to make sure that they get close to the 'best' rigid and affine transformations. I tried to investigate the weighting between shape and texture. I started with 0.5 and stuffed a zero at the right side of the decimal point. NRR now comprised 10 iteration with 3 knot-points along each dimension.

• 20112004-6: weight = 0.5
• 20112004-7: weight = 0.05
• 20112004-8: weight = 0.005
• 20112004-9: weight = 0.0005
• 20112004-10: weight = 0.00005
• 20112004-11: weight = 0.000005
• 20112004-12: weight = 0.0000005

Videos depicting a sequence of the results above suggested that we should try and see what happens at weight > 0.5. There appears to be better alignment when the weight is set to high values¹. Here are some of the next experiments:

• 20112004-13: weight = 0.9 (trying higher values which result in better alignment, according to visual appraisal)
• 20112004-14: weight = 2
• 20112004-15: weight = 10

Perhaps this is obvious, but values above 1 do not seem to affect registration. The results do not surpass these which were obtained at weight = 0.9. They appear almost identical, in fact, but they are not. There is a difference between weight = 2 and weight = 10 and it seems like a small 'vibration' in the brain. weight = 0.9 and weight = 2 give identical results, maybe because they are the same order of magnitude. The same behaviour exactly is observed for weight = 0.9 and weight = 0.5. Comparison of the results is in 21112004-1.