0

1. Mattes Mutual Info Doubts

In SimpleITK Mattes Mutual information is a similarity metric measure, is this a maximizing function or minimizing function?

I have tried a 3D registration(image size : 480*480*60) with Metric Mattes Mutual Info metric and Gradient Descent Optimizer

Output

numofbins = 30

Optimizer stop condition: RegularStepGradientDescentOptimizerv4: Step too small after 24 iterations. Current step (7.62939e-06) is less than minimum step (1e-05).

Iteration: 25

Metric value: -0.871268982129

numofbins = 4096

Optimizer stop condition: RegularStepGradientDescentOptimizerv4: Step too small after 34 iterations. Current step (7.62939e-06) is less than minimum step (1e-05).

Iteration: 23

Metric value: -1.7890

If it is a minimization function then the lower one is better, which I suspect.

2. Transformation matrix final Output

TranslationTransform (0x44fbd20) RTTI typeinfo: itk::TranslationTransform Reference Count: 2 Modified Time: 5528423

What is Modified Time?

3. Final Metric is a registration accuracy measurement?

Is metric a sign of registration accuracy? Is a higher metric value mean better registration? Or it is just a value at optimum point after optimization?

4. Random sampling for registration

10-20% of random sample points suffice for a registration. But the doubt arises whether the samples are taken from the main ROI or outside the ROI? Masking is an option, is there any other option in SimpleITK?

Thanks

Subhajit
  • 43
  • 9

1 Answers1

0

Similarity metrics in ITK usually give the cost, so the optimizers try to minimize them. Mutual information is an exception to this rule (higher MI is better), so in order to fit into the existing framework it has negative values - bigger negative number is better than small negative number, while still following the logic that it should be minimized.

Modified time is used to check whether a certain filter should be updated or not.

Generally lower metric means better registration. But it is not comparable between different metrics, or even between different types of images using the same metric.

Random sampling will take 10-20% of samples in your RoI. I am not sure whether it picks randomly within RoI, or picks randomly within image and then checks whether it is in RoI.

Dženan
  • 3,329
  • 3
  • 31
  • 44
  • Metric can be used to estimate Number of bins. right? because we are just changing number of bins. Nothing else. Is there any output parameter to check how much time registration took ? – Subhajit May 16 '17 at 08:16
  • Random sampling doesn't have knowledge about the ROI right? So I think it will consider the whole image as sample space. No? If I use metric X for methodA and methodB, for same image pair, the metric will not be helpful comparing two. Though I don't know when they can be misleading because for a image pair, the cost function will be same , as we are using same metric, only optimization algorithm will be different. It will just give the transformation to use and metric will be calculated using same formula in two cases. So, I think metric value 1.23 in case 1 should be same as case 2 , no? – Subhajit May 16 '17 at 08:35
  • For the case of same metric and same image pair, metric value can be used to compare different registration transforms or methods used to achieve them. And lower value is better! Lower means more negative for mutual information. – Dženan May 17 '17 at 14:14
  • Yes! Only MI related metrics are opposite. Same thing can be concluded for RMS (root mean square ) also? I mean lower is better. Which is used in case of demon registration! – Subhajit May 17 '17 at 15:00
  • Yes it can - same metric and same image pair leads to comparable registrations. – Dženan May 17 '17 at 16:38