Browse for the Trained Network (*.gnn) to validate. Select the BestModel.gnn generated with Train Neural Network.

Choose the Dataset Folder created with Create Training Data containing the corresponding training data in the two folders Train and Test.

Then select the Subfolder that should be used for validation:

• Train contains the data that was used for training. The neural network should perform well on this data, as it was explicitly trained for this.
• Test contains the data produced explicitly for the validation. This was not used in training and contains different data with similar statistical properties. If the network also performs well on this data, it can be applied on real scans with the required properties.

The Threshold is an internal (or expert) parameter that is mostly applied by proficient users at Math2Market. The parameter is only available if the chosen network is trained to differentiate between two material phases.

It is used for the decision whether a voxel is a target material voxel or not. The ROC curve shown here is based on the chosen threshold.

Find a visualization of the confidence field (*.npz) and threshold here.

Specify the minimum required confidence to select a class for a voxel.

The material to analyze should be split into different materials by the neural network. For each of the resulting materials a confidence field is computed. These fields contain values between 0 and 1, where 1 means, the neural network is perfectly sure, that this voxel should be assigned to the corresponding class and 0 means it is definitely not the corresponding class.

If none of these fields has a higher probability value than specified the material will be set to uncertain in these locations and material ID 255 is assigned.

The values of the confidence fields of all classes sum up to 1. Thus, not all confidence values can be smaller than 1/n, where n is the number of classes. That is why setting the Confidence value below 1/n will not have any impact.

An example is shown here.