9.6K Views
•
11:38 min
•
August 23rd, 2017
DOI :
August 23rd, 2017
•0:05
Title
0:53
Preparing the Workplace, Data Representation, and Generating Appropriate Super-regions
3:19
Introduction to Annotation
4:12
Segmentation Using Model Training
5:41
Segmentation Using Super-regions
6:42
Classifiction and Analysis of Data Objects Based on Inherent Characteristics
9:00
Results: Model Training, Super-region Segmentation, and Label Splitting Using the SuRVoS Workbench
10:55
Conclusion
Transcription
The overall goal of this procedure is to semi automatically segment complicated volumetric datasets of a biological nature. This method can help answer key questions in the field of structural biology by allowing for the qualitative and quantitative analysis of volumetric data. The main advantage of this technique is that in Though this method can provide insight into biological datasets, it can also be applied to other fields such as material science.
This software uses a different workflow as segmentation programs which can cause initial difficulties when first using SuRVoS. Therefore, visual demonstration is critical to understand this new workflow. Segmentation of large well-defined regions using model training is demonstrated on a phase contrast x-ray tomography dataset.
To begin this procedure, launch the SuRVoS workbench. Click the Open Dataset button and select the data file to be segmented in the resulting popup. Choose an appropriate orientation for the dataset.
Next, choose a folder in which to store the workspace and the associated files. Click the Load button. In the Select ROI tab, input the X, Y, and Z start and end coordinates for our region interest and click Add.
Once a section is added, make sure it is selected by checking the box to the right. After this, in the Features Channels tab, use the dropdown menu to choose a feature and add it to the queue. Modify any options specific to the feature and choose the input dataset on which to run the feature.
Click the checkbox to the right of the feature name to compute. Multiple features can be run on the same dataset and computed features can be used as input datasets for further processing. Then choose the Super Regions tab.
In the Super Voxels section, use the dropdown menu to select the filtered dataset from which the super voxels will be created. Specify the shape, spacing, and compactness. Click the Apply button to generate the super voxels.
To assess the quality of the super voxels, display them alone without the data overlaid. If the features of interest from the data are still visible in the super voxels alone, they represent the data well. In the Annotation tab, click the Add Level button to add an annotation level.
Click the Add Label button in the newly created level to add a label for the annotation. Next, in the tool shortcut section, select the pen icon. A set of options will appear at the top of the visualization pane.
Select the super voxels option and a middle width pen to begin creating training data for the model training. Click the box to the far right of the label information to select the label to be annotated. Then click and drag in the visualization pane to annotate multiple super voxels.
In the Model Training tab, set the predict level to the level that contains the manual training annotations. Then in the descriptor section, set the region to super voxels. Click on the Select Sources dropdown and check the boxes for the features or filters of choice to select the descriptors that will be used to differentiate regions of data.
Click on the Predict button. When the computation is complete, the visualization pane will update with predictions of which annotation label each non-labeled voxel belongs to. After assessing the effect of the training methodologies and choosing one, click the Refine dropdown in the refinement section to apply additional refinement.
At the bottom of the model training tab in the Update Annotations section, ensure that the visualization dropdown menu is set to predictions. Next, use the confidence slider to assign more or less of the unannotated super voxels to the selected annotations labels. After an appropriate level of confidence has been selected, click the Save buttons next to the labels at the bottom of the confidence tool to save the predictions into specific labels.
The visualization pane will update to reflect the changes made. If necessary, repeat the model training with additional refinement and high confidence predictions until there are few or no unlabeled super voxels. Segmentation of smaller more complex regions using super regions is demonstrated on a cryo-electron tomography dataset.
After adding appropriate levels and labels in the annotation tab, select the label. Using a middle width pen, begin annotating with super voxels selected. One strategy for super region segmentation is to quickly segment on one slice, move a few slices above or below, and fill in the gaps on the new slice.
In this way, the intermediate slices will be annotated as well with less effort by the user. Next, to clean up the annotations further, select the segmentation label and one of the morphological refinement methods. Enter a radius value and choose how to apply the refinement method.
Then click Refine. Classification and analysis of data objects is demonstrated on a cryo-soft x-ray tomography dataset. After fully segmenting the data, click on the second tab in the visualization pane called Label Splitter.
A new area, the rule creation pane, will be added to the right-hand side of the window. At the top of the rule creation pane, select the level and labels for label splitting. Next, select the dataset to query and click Label.
Each object in the selected labels will now be outlined in blue as separate objects in the visualization pane. The rule creation pane will now contain a plot showing the average intensity of the objects. Click on the dropdown box in the upper right-hand side to change the measure being shown.
Next, click Add New Label at the bottom of the rule creation pane to begin splitting the objects into relevant classes. Click Add New Rule and use the dropdown and free form entry boxes to define the rule to be applied. Click Apply to see the effects of the new rule in the visualization pane and on the plot in the rule creation pane.
After all objects of interest are classified, go to the Annotations tab. Create a new empty level. Then choose this new level in the rule creation tab and click Save Labels.
Click on the Label Statistics tab on the edge of the visualization pane to open a new visualization pane that can be used to understand the relationships between object classes. At the top, select an appropriate level and labels and the dataset to query. Then select a few measures of interest by checking the boxes next to them.
Click Label to produce pairwise comparison plots for each of the selected measures. If a measure needs to be added or removed, click the appropriate checkbox and then click Update Plot. In this study, two segmentation strategies and one classification tool in the SuRVoS workbench are demonstrated.
For model training, a relatively high contrast dataset with region-defining boundaries is loaded. The data is filtered and clamped to make the background, foreground, and inner structures more distinguishable. Super voxels are then built on top of the filtered dataset.
After the quality of the super voxels is assessed, manual annotations are made to train the classifier to predict the areas corresponding to the background, the fruit bristle, the seed material, and the surrounding flesh. Morphological refinements are used to clean up the segmentation by filling holes. For super region segmentation, a noisy and complex dataset is loaded.
Next, an appropriate filter set is applied to the selected region of interest. After model training and segmentation, quick manual annotations are used every five to 10 slices to select the super voxels that describe the feature of interest. Label splitting is then demonstrated using a dataset with many varied organelles.
First, an appropriate total variation filter is used to enhance the organelle boundaries. Organelles are semi-manually segmented using mega voxels and super voxels and refining is used to fill holes and smooth edges. The label splitter is used to visualize each organelle as an object in the dataset and then to visualize various characteristics about each object in the data plot.
Here, rules were created to separate the organelles into five distinct classes based on their inherit properties, for example their size or average intensity. The label splitter can be used to output quantitative information about the data and begin to understand the cellular context. After watching this video, you should have a good understanding of how to use SuRVoS workbench for semi-automatic segmentation.
Using this procedure, comparisons of multiple biological states can be made in order to answer questions, for example about wild-type disease and then treated conditions. Once mastered, this technique can speed up the segmentation process by approximately five times if it is performed properly. Following this procedure, visualization programs can be used to render the results for publication and movie making.
Segmentation of three-dimensional data from many imaging techniques is a major bottleneck in analysis of complex biological systems. Here, we describe the use of SuRVoS Workbench to semi-automatically segment volumetric data at various length-scales using example datasets from cryo-electron tomography, cryo soft X-ray tomography, and phase contrast X-ray tomography techniques.