Idea is to capture the spatial characteristics of the tumor by detecting Texture-sensitive edges within the tumor at different scales and different apperture weights.

EMD texture edges, Scale=3, Apperture=constant

 Here are the corresponding Edge strength distribution histograms:  They have a widely varying StdDev/Mean and so on.

We can also Build the "Concentric Profiles" of Edge Strength distribution ( in corresponding order)

"Tumor blob analysis"

Tumor is thresholded.

Dark blob info is available

"Tumor ring profile" trials:

Mean=Blue, StdDev=red,  1 vox step -> inwards

Full Resolution Femur sample:
Run the registration Using URAL, default parameters, Subsample=8,with Finetune=1.
Timings:
Primitive generation:   3 min
AutoFocus:                  50 sec
FineTune:                     25 sec

Result is below.  I don't see a need to change anything about the registration FineTune.
MI problem has to be investigated.

Femur Full-resolution registration sample:

URAL registration works well with the present default parameters. Below is the result for the AutoFocus stage (no Finetune)

Using URAL, AutoFocus-only, no Finetune

However, for the MutualInfo this stage fails with the result grossly misplaced.

FGT segmentation: Implemented a hybrid FAT thresholding method based on Z-alg for when histogram has 2 distinct peaks. New criteria was developed for the absence of such peaks. In that case the PVV segmentation is invoked.
Processed all the 16 "challenge" cases with good results.

Developed completely new application-oriented AirThresholding method. It is based on Z-alg idea.
Validating it on the failed breast FGT samples where selecting across-the-board threshold = 60 did not work.  Specifically, in REIC cases the proper thresholds were substantially higher :  {67,108}
in KELL cases the proper threshold were substantially lower {17,15}

REIC_R

KELL_R

KELL_L

REIC_L

Added the {BiGauss_PVV,BiLaplace_PVV} methods to the menu of Histogram segmentation methods in
a) RoiStats3D dialog
b) "ROI>Split ROI with threshold" menu

Improved the performance of the "Interval Arithmetic" library. This is used in all Histogram segmentations and dynamic modelling.

Developed new thresholding methods  "Bi{Gaussian,Laplacian} with PVV correction".
These methods work on the idea that the histogram segment between 2 peaks is unreliable due to the Partial Volume Voxels (PVV) and should be completely disregarded.  Intuition here is how human thresholds complex fused histograms, just by looking on the side tails of the distribution.

Tested on the difficult case ABBE-R. With the good result.

Histogram modelling

Final segmentation

problematic case ABBE-R
seems the BiLaplacian split found a wrong threshold:

ABBE-R.  BiLaplacian split.  Residual = 2549

However, when we re-segment the BiGaussian model, we obtain much lower residual.

BiGaussian fit.  Residual = 645

ABBE-R:   Split using BiGaussian

However, for Phantoms the BiLaplacian was much better than Gaussian at least for most cases and all phantoms were processed using exclusively BiLaplacian.


Proposal:  We evaluate histogram fit using both {BiLaplace,BiGauss}.  We choose the method that provides smaller fitting residual.

Determine the Air threshold in Breat FGT images:
We inspect the histogram that has a gross-dominant peak around the Signal=0. We fine first evident valley and assign the valley bottom to be the AirThreshold.

New workflow on case ABBE_L:

result of the segmentation:

Testing new FGT workflow on human data:
BiCal + BiLaplace provide excellent result. Exactly the parameters  from Phantom processing
were used

BiCal followed by BiLaplace

Breast FGT phantom workflow:  implemented completely new workflow  BiCal+BiLaplace+MaxCC+EdgeWaveML.  Processed all 24 cases with very good result matching the physical measurement.

Some voxels are marked as FGT, but should be they included at all?  They could be fairly easily excluded by demanding that the FGT+FAT is a Connected object and does not consist of the several
disconnected components. This should be true for Phantom and Human data?

"EdgeWave MultiLabel - Boundary Only" variant was implemented.  Only the boundary voxels of the total region are modified by Multilabel Peel\Grow.
Provides slightly different result

FGT=187.8 cm3, FAT = 944, Total = 1131.8, Fgt%=16.5%

Implemented new algorithm "EdgeWave MultiLabel". This is a general algorithm: given an initial segmentation of the region with multiple labels it applies EdgeWave morphological criteria to correct these regions while completely covering the total region.

This was applied to p1s0_L example that had issues with partial volume voxels being misclassified as FGT.

This algorithm contains both internal and external morphological competition.  So additional algorithm "EdgeWave ML Boundary" will be attempted.

FGT: 180.23, FAT=950, total=1130.23,  FGT percentage=15.9%

Implemented the original variant of the FGT breast workflow.  It consists of BiCal, Bimodal Laplace thresholding.  While providing great (+-4% ground truth) result for some images, other images are problematic. This requires to develop an additional Morphological module, as none of the existing variants of EdgeWave don't seem to fit.

This is the "fused peak" case that Henry sent earlier.  As far as I understand this was the case "p4s0.fvx"-RIGHT (DICOM5)

Again BiCal was applied with the exact parameters as with other cases. Then BiLaplacian with exactly same parameters too.  We see a great histogram split as a result.

Initial fused peaks case resolved.

Corresponding segmentation (top and bottom slices)