[Insight-users] registration non-convergence?

[Carolyn Johnston]

Hi Luis et al,

I actually do have another question (I guess this is just my day).

I am performing a standard registration with the mean squares metric and 
the reg step gradient descent optimizer. I have set my fixed image region 
to be 1000 by 1000 for speed (on a 3600 by 3600 fixed image). My moving 
image is just the same as the fixed image with a translation I've added.

I begin with initial parameters 0,0.

0 = 2185.867696: [0.000052, -0.000019]

1 = 2123.551007: [0.000104, -0.000039]

2 = 2072.931500: [0.000156, -0.000058]

3 = 2033.078306: [0.000209, -0.000077]

4 = 2003.966331: [0.000261, -0.000097]

... <snip> ...


Now right about here, I pass through a metric minimum and keep on going for 
some reason...


76 = 467.887291: [0.003999, -0.001529]

77 = 467.045557: [0.004051, -0.001549]

78 = 466.963944: [0.004103, -0.001569]

79 = 467.642714: [0.004154, -0.001589]

80 = 469.093127: [0.004206, -0.001609]

... <snip> ...


and I end up here, with a metric value significantly higher than the lowest 
one I achieved during the registration. Any idea what might be causing 
this? I thought the registration would try to converge around the minimum 
by reducing the stepsize.

95 = 502.706704: [0.004986, -0.001905]

96 = 508.162350: [0.005038, -0.001925]

97 = 512.621524: [0.005090, -0.001944]

98 = 517.175693: [0.005142, -0.001964]

99 = 520.728025: [0.005194, -0.001984]


*** Report for registration of dem 1 ***
Translation parameter along X: 0.005194
Translation parameter along Y: -0.001984
Iterations to reach optimal value of metric: 100
Optimal value of metric: 520.728025