Sensitivity and Syntheses

• The following graph is not a very useful one
• Generated by mistake
• Computing the sensitivity to increases/decreases in the number of model syntheses
• Gives insight into the stability of evaluation as the number of synthetic images increases

Sensitivity and Syntheses - Graph #1

Sensitivity to change in number of model syntheses. Sensitivity is calculated over Specificity of a registered set model.

Sensitivity and Syntheses - Graph #2

Sensitivity to change in number of model syntheses. Sensitivity is calculated over Specificity of a model of mis-registered images.

Sensitivity and Syntheses - Observations

• There is no agreement among the 2 evaluations
• For more continuous curves, one needs
  • more models
  • more evaluations (with a varying number of modes)
  • multiple instantiations (perturbed image sets which make up models)

Number of Modes and Sensitivity

• Looking at 1,600 instances from the model
• The derivation agrees with our previous formulation of sensitivity
• The formulation: Difference between specificity values (registered, unregistered)...
• ...divided by the mean of the errors on Specificity

Number of Modes and Sensitivity - Results

Specificity sensitivity in models is shown for different number of modes. Limited number of modes is of course used in an evaluation which is finite.

Bending Energy - Graph

The effect of the number of knot-points on bending energy in images

Bending Energy and Number of Knot-Points

• The figure demonstates that bending energy behaves as we would expect
• Note: Y is scaled logarithmically
• "Points" refer to the number of knot-points, which are distributed at random
• For a number of points, say n², the number of point becomes (n+1)² in the subsequent stage

Label Propagation and Strong Edges

Selection of knot-points with placement on strong edges. Propagation of labels is shown after 5 warps were applied to each image.

Label Propagation and Strong Edges - Ctd.

• The method involves
  • Selecting points in n strong edges
  • Building model using NRR
  • Propagating labels from one (immutable) image to all others
• Performance is somewhat satisfying at the start

Summary/Conclusions

• Progress can be broken down into three parts:
  • A reasonable choice of number of modes, which are to be in evaluation, leads to better sensitivity
  • Bending energy increases significantly when the number of knot-point does
  • Segmention and label propagation, driven by strong edges

Present and Future

• Impending results from a 2-D registration of IBIM images and labels
• Comprehensive model evaluation experiments
• Decision on various parameters, e.g. number of modes, number of instantiations for better statistics
• Further experiments on segmentation in the interim