The four different AI Models Boosted Tree, Random Forest, Unet2D, and Unet3D are described in this chapter. In the following table, the learning methods are compared directly.

Random Forest and Boosted Tree

Since GeoDict 2023, the AI-segmentation via the widely used Random Forest method is featured in addition to the Boosted Tree method. Both methods work very similar, but the Boosted Tree model is usually faster, since it is an advanced version of the Random Forest.

Boosted Tree and Random Forest both work well for many examples and are very fast. For the training data, information about the neighboring voxel data are used and different filters are applied to the input image:

• a Gauss filter blurring the image, and
• a Sobel filter, emphasizing the edges.

Define the filter parameters for these images and to which distance the neighboring voxels should be considered in the Select Features dialog explained below or simply use the default values.

In the figure below, three filtered images are used, one with a small sigma for the Gauss filter, one with a larger sigma, and one image filtered with the Sobel filter.

Then, for each voxel a vector is built containing the corresponding data from the original image, from each of the filtered images, from the neighboring voxels, and from the labeled image. In the diagram below this is shown for two voxels.

The Random Forest / Boosted Tree algorithm learns from these vectors and segments the image.

In the case of big-scaled relevant features, the Random Forest and Boosted Tree methods can be limited. That is because the decision for each pixel depends on the surroundings, especially on the Gauss kernel of a fixed size. If, e.g., the image has pores of different materials but with a similar gray value and the only difference is the border to the solid material, for big pores the boosted tree method could decide wrong. In such cases, the Unet methods are recommended. Also, if more than two materials are in the structure, they can be limited and the Unet method is preferable.

Click on the gearwheel icon gives access to the parameters defining the model setup. The Set Model Parameters dialog opens. The following parameters can be set for the Boosted Tree model:

• Number of Trees: Defines the number of parallel trees constructed during each iteration. More trees result in a higher learning ability but can lead to overfitting. An overfit model does not have a good generalization ability.
• Maximum Depth: Maximum depth of a tree. Increasing this value will make the model more complex and more likely to overfit.
• Learning Rate: The learning rate is a step size shrinkage used in each update to prevent overfitting. The learning rate must be >0 and ≤1. The lower the learning rate, the longer the runtime but the higher the generalization ability of the final model.
• Minimum Split Loss: Minimum loss reduction required to make a further partition on a leaf node of the tree. The larger the minimum split loss, the more conservative the algorithm will be.

The following parameters can be set for the Random Forest model:

• Number of Trees: Defines the number of parallel trees in the forest constructed during each iteration. More trees result in a higher learning ability but can lead to overfitting. An overfit model does not have a good generalization ability.
• Maximum Depth: Maximum depth of a tree. Increasing this value will make the model more complex and more likely to overfit.
• Minimum Samples Split: Minimum number of samples required to split an internal node. The larger the minimum samples split value, the more conservative the algorithm will be.

For Boosted Tree and Random Forest the Select Features dialog controls the parameters for the applied filters for the Boosted Tree and the Random Forest methods.

These filtered images are considered in the training of the model as shown in the figure above to gain more information for each voxel. They are used also in the segmentation, but only as a reference for the model, since they can help to identify which label is correct for which voxel in the original image. These filters always were applied for Boosted Tree with the current default settings, but since GeoDict 2023, their parameters can be changed by you.

Check the features that should be used and enter a list of parameters for the feature runs, separated by “;”:

• The Neighbors method has the parameters Distances and Directions. For the default settings, the neighborhood is examined three times with a distance of 1 voxel – once for each direction. The voxels in the neighborhood within the given distance are considered as information for how the current voxel should be labeled.
• The Gauss filter has the parameter Sigmas. Sigma is the standard deviation for the Gauss function used for this filter. For the default settings this filter is run twice, once for a sigma of 1 and once for a sigma of 3. The higher the standard deviation, the more the image is blurred. This filter is also described here.
• The Sobel filter has the parameter Directions. In the default settings, it is applied in all three directions. This filter emphasizes the edges. This filter is similar to the Compute Gradient filter.

Unet

The deep learning methods Unet2D and Unet3D require more training data. Thus, more labels must be provided manually. However, they are more capable to analyze the scan correctly and thus, achieve better results.

The name Unet refers to the U-shape of the neural network diagram, consisting of a constricting branch on the left and an expanding branch on the right. The number of layers in each branch defines the depth of the Unet.

To use the Unet models a good graphics card is needed, i.e., a NVIDIA graphics card (GPU) with compute capability of at least 3.5. Make sure that the drivers are installed and up-to-date. See more information on graphics cards at https://developer.nvidia.com/cuda-gpus. This webpage contains a helpful section with Frequently Asked Questions.

For Linux, the Gnu C Library (glibc) must be at least version 2.17. We recommend Ubuntu 20.04 LTS, but for a current glibc other Linux distributions should also work.

Unet3D considers the complete image, while Unet2D learns from a single slice in the specified direction. The default is the Z-direction.

Click on the gearwheel icon to access the parameters defining the model setup. The Set Model Parameters dialog opens. The following parameters can be set for the Unet model:

• Window Size X, Y, and Z: For the training, the Unet algorithm subdivides the gray value image into windows of the given size in voxels. A bigger window size requires more training data and time, but the learning potential increases. The windows are placed according to the given Stride or around every labeled voxel.
• Number of Epochs: The number of times, the training data is used to train the neural network. For more epochs, the learning time grows linearly, but the learning potential increases. Too many epochs, however, can lead to overfitting.
• Depth of the Unet: The number of levels in a Unet. A higher depth requires more training data and increases runtime and GPU memory, but the learning potential increases.
• Number of Features in First Layer: The number of features that the model can learn in the first layer. More features require more training data and time, but the learning potential increases.
• Kernel Size: Control the size of the convolutional kernel. A kernel is a feature detector, i.e., an n x n x n voxels box (or n x n square for Unet2D) with a defined pattern. The size n is the number of voxels in one direction and must be an odd number. Common values are 3 or 5.
• Use Strides: Use a window stride defined by the values Stride X, Y, and Z. These strides determine how many voxels the window is moved for each new training sample, starting in the upper left corner. Then, only the windows containing labeled voxels are taken into account. If not checked, windows around every labeled voxel are used for the training.
• Normalize: Normalize the input images using the minimum and maximum gray values in the image. If this option is unchecked 0 and the maximum possible values are used.

For a more detailed explanation about the Unet parameters refer to the GeoDict-AI user guide.

Unet2D means that the windows determined by Window Size are 2D slices, i.e., one of the three Window Size parameters is 1. The default for Window Size Z = 1 usually leads to good results. If all three size parameters are greater than 1, the Unet3D model is applied automatically.

The choice of Batch Size is only available for the Unet models and is the number of samples the training uses simultaneously. For example, with a batch size of 4, during training the error is computed, and the network updated for one window at the same time, before moving on to the next four windows. In literature there is usually a warning about too large values for batch size due to the potential of overfitting. However, in 3D image analysis the batch size is limited by the available GPU memory. If the value is chosen too high, a warning dialog appears, when starting the training. It is recommended to set the value as high as possible with the available GPU memory, so that the training works without giving an error. This is often a value between 1 and 8 depending on the GPU hardware. Modern GPUs can allow for even higher values for the batch size.

If multiple GPUs are available, select on which it should run, by checking the corresponding checkboxes. If licensed, multiple GPUs can be used for the training and segmentation. Then, the batches are distributed to equal numbers on the different GPUs.

If no GPU is detected, the training and the AI-segmentation are run on the CPU, which usually needs much more runtime.

The Select Additional Images dialog controls which of the loaded images should be considered for the training additionally to the current image. Multiple images can be loaded to the Image Processing dialog by having Keep existing Volume Fields selected in the Import Geometry dialog. Thus, it is possible to take multiple images from the same structure’s cutout into account for training. For this, the images must be the same size and fit to the used labels. If the network was trained on several images, these images also are needed for the segmentation.

Different scanning settings can produce different kinds of images. For example, if an image has three material phases it can sometimes be hard to get a good contrast for all three of them at the same time. Then, the image can be taken twice, once with a good background separation and once with a good contrast between the solid materials. For the training the neural network considers both data sets and thus, will generate better results, compared to only considering one of these images.

For all models, choose the Number of Materials to be labeled in the scan.

Specify the different materials with the material database. The checkbox next to the material boxes determines the current material for painting the labels.

The Navigation Mode allows to pan the image in the 2D Slice Visualization section with the left mouse button and zoom in and out with the right mouse button.

Select the Painting Mode to paint in the 2D Slice Visualization area. Ensure to have the Visibility turned on as described here and AI-Labels selected for Overlay as described here. The Painting Mode allows to paint labels on the scan with the left mouse button. Switch between the materials by selecting the corresponding checkboxes. When pressing and holding shift, the navigation in the 2D Slice Visualization section works the same as in the Navigation Mode. Press and hold control to erase labels by clicking or holding the left mouse button.

Make sure to have approximately the same amount of labeled area for the different materials, especially for the Unet models. Validate this in the Label Data tab of the Histogram section.

If a label was not painted correctly switch to the Erasing Mode and erase labels by clicking or holding the left mouse button and moving the mouse over the labels.

Change the Brush Size fitting to the areas to label.

When you enable the Magic Brush the images are analyzed on clusters recognizing edges. The Magic Brush clusters are computed using the SLIC (Simple Linear Iterative Clustering) algorithm, which generates the cluster based on gray value similarity and proximity in the image plane. Then, labels can be created by simply left-clicking on these clusters in Painting Mode. To show the clusters, enable Show Magic Brush Outlines. You can change the size of the created clusters by changing the Brush Size. For many image datasets this feature makes labeling much easier.

Using the different view directions, you can observe the already existing labels from slices labeled in the other two directions, respectively.

In the example below, the center slices in X- and Z-direction were labeled. Thus, if viewed in Y-direction, two thin lines with white (material 1) and red (material 2) sections are already labeled. If labels are erased on these lines, in the respective slices in the other directions thin lines are erased. This leads to less training data.

If there are more than two materials, it is important to label boundaries for all different boundary combinations.

Labels close to each other give better results. For each labeled area paint as many labels as possible to get a fully filled output window.

While it can be important to label boundaries, only label as near to the boundary as it can be ensured to label it correctly. Otherwise, wrong training data will be generated. Paint in all three directions, especially if the gray values differ much.

Especially for the Unet3D method ensure to label some consecutive slices in the same areas.

Save the painted labels as *.gld file by clicking Save Labels. Load the training data again whenever needed by clicking Load Labels. Delete all labels by clicking Clear Labels.

If a trained model is loaded, the Overlay can also be set to Preview to preview the segmentation results for the current slices. When changing the Overlay to Mixed, both the current Labels and the Preview of the current model can be viewed. If the preview shows that the model needs more training, using the mixed view helps to concentrate the further labeling on the areas, where the model did not classify the voxels correctly.

If enough labels are painted click Train to train a machine learning model.

When the training is completed, select Preview to control the resulting segmentation. The Overlay is switched to Preview.

In the example above, the Boosted Tree method was used. Thus, only few labels are needed for this gray value image of a Berea sandstone.

A previously trained and currently loaded model can be improved by clicking Continue Training. A model is loaded either by training it before in the current Image Processing session or by loading a previously trained model by clicking Load Model.

For example, with more labeled gray values, Continue Training can improve the model. Note that not only the new labels are considered, but also the labels used for the loaded model to use all available information for the training.

Additionally, a model can be improved if another similar image is loaded and labeled to provide more training data.

Click Save Model to save a trained machine learning model. This model can be loaded again by clicking Load Model whenever similar scans have to be segmented. For the different AI models the formats are *.XGBM for Boosted Tree, *.FOREST for Random Forest, *.UNET2D for Unet 2D, and *.UNET3D for Unet 3D.

As final step click Create Segmentation to apply the trained AI model to the complete image dataset.

After clicking Create Segmentation for one of the segmentation methods, the segmentation is applied for the gray value image. For Global Thresholding, AI Segmentation, Multi-Phase Segmentation, and Hysteresis Thresholding a result file (*.gdr) is generated and saved in the project folder. The generated result file with the name entered for Result File Name is opened automatically in the Result Viewer.

The Report tab lists some basic Structure Information for the resulting structure and all steps done in the Image Processing dialog, e.g., Segmentation methods and the used Image Filters.

In addition to the result file, a result folder with the same name is saved inside the current project folder, containing the segmented structure file (*.gdt) and, for the AI segmentation, also the used labels (*.gld), and the trained model (*.FOREST, *.XGBM, *.UNET2D or *.UNET3D).

For a more detailed description of the Result Viewer options refer to the Result Viewer user guide.