Medical Imaging Interaction Toolkit  2022.04.00
Medical Imaging Interaction Toolkit
The Segmentation View
Icon of the segmentation view

Some of the features described below are closed source additions to the open source toolkit MITK and are not available in every application.


Segmentation is the act of partitioning an image into subsets by either manual or automated delineation to create i.e. a distinction between foreground and background. A multilabel segmentation can contain more than one label and more than one layer. This allows you to create different labels for different regions of interest encapsulated in one single image. The difference between labels and layers is that labels on one layer cannot overlap but labels on different layers can.

The MITK segmentation plugin allows you to create multilabel segmentations of anatomical and pathological structures in medical images. The plugin consists of two views:

In this documentation, the features and usage of the segmentation view will be described. For an introduction to the segmentation utilities please be referred to the respective documentation pages.

Segmentation plugin overview


The segmentation plugin offers a number of preferences which can be set via the MITK Workbench application preferences:

Segmentation preferences
  • Slim view: Allows to show or hide the tool button description of the segmentation view
  • 2D display: Specify whether the segmentation is drawn as outline or as a transparent overlay
  • Show only selected nodes: Enable if only the selected segmentation and the reference image should be visible
  • Smoothed surface creation: Set certain smoothing parameters for surface creation
  • Default label set preset: Start a new segmentation with this preset instead of a default label

Technical issues

The segmentation plugin makes a number of assumptions:

  • Images must be 2D, 3D, or 3D+t.
  • Images must be single-values, i.e. CT, MRI or "normal" ultrasound. Images from color doppler or photographic (RGB) images are only partially supported (please be aware that some tools might not be compatible with this image type).
  • Segmentations are handled as multilabel images of the same extent as the original image.

Data selection & creating new segmentations

To select a reference image for the segmentation, click on the Selected image selection widget and choose a suitable image from the selection available in the data manager. Once an image is selected, a new segmentation can be created on this reference image by clicking the button to the right of the Selected segmentation selection widget. A new multilabel segmentation with an initial label is automatically generated. The new segmentation will be added to the data manager as sub-node of the reference image. This item is then automatically selected in the data selection, which allows to start editing the new segmentation right away.

Data selection

Alternatively to creating a new segmentation, an existing one can be edited as well. If a reference image is selected for which a segmentation already exists in the data manager, the auto selection mode will automatically select a fitting segmentation. Clicking on the segmentation selection widget a drop down list will open, containing all suitable segmentations for the selected reference dataset available in the data manager.

Segmentation layers

For each multilabel segmentation different layers can be added or deleted. The layers can be used independently and layers can be switched using the left and right arrows. A layer is a set of labels that occupy a non-overlapping anatomical space. The best way to describe them is by a real use case: Imagine you are working on a radiotherapy planning application. In the first layer of your segmentation session, you would like to trace the contours of the liver and neighboring organs. You can accommodate all these segmentations in separate labels because they all occupy different anatomical regions and do not overlap. Now say you would like to segment the arteries and veins inside the liver. If you don't trace them in a different layer, you will overwrite the previous ones. You may also need a third layer for segmenting the different irrigation territories in the liver and a fourth layer to contain the lesion you would like to treat.

Layer selection

Segmentation labels

For each layer, one or more labels can be added. Pressing the double arrow on the right, all created labels are shown in the 'Label Table'. The following label properties are available:

  • Name: the name of the label. Can be a predefined one or any other.
  • Locked: whether the label is locked or editable. A locked label cannot be overwritten by another.
  • Color: the color of the label.
  • Visible: whether the label is currently visible or hidden.
The 'Label Table' shows all labels in the current segmentation session

The 'New Label' button can be used to add a new label. This will automatically add a new label with a distinct name and color to the list of available labels.
In the current implementation of the plugin, the maximum number of labels is restricted to 255. If you need more, you will have to create a new segmentation session.

Saving and loading label set presets

Label set presets are useful to share a certain style or scheme between different segmentation sessions or to provide templates for new segmentation sessions.

The properties of all labels in all layers like their names, colors, and visibilities are saved as a label set preset by clicking on the 'Save label set preset' button. Label set presets are applied to any segmentation session by clicking on the 'Load label set preset' button. If a label for a certain value already exists, its properties are overridden by the preset. If a label for a certain value does not yet exist, an empty label with the label properties of the preset is created. The actual segmentations of labels are unaffected as label set presets only store label properties.

Applying label set presets by default

If you work on a repetetive segmentation task, manually loading the same label set preset for each and every new segmentation can be tedious. To streamline your workflow, you can set a default label set preset in the Segmentation preferences (Ctrl+P). When set, this label set preset will be applied to all new segmentations instead of creating the default red "New label 1" label.

Searching for a label

It may happen that many labels (e.g. > 200) are present in a segmentation session and therefore manual searching can be time-consuming. The 'Label Search' edit box allows for quickly finding a label by providing assistance for label name completion. If the label is found, it will become the active one after pressing 'enter'.

To start editing a label needs to be activated by clicking on the corresponding row in the 'Label Table'. Only one label can be active at the time. Then the segmentation tools in the toolbox can be used for mask generation.

Operations on labels

Depending on the selection in the 'Label Table', several actions are offered:

Operations with single label selection

Right-clicking on any label opens a pop-up menu that offers the following actions to be performed on the selected label:

  • Rename... : change the name and / or color of the selected label.
  • Remove... : delete the selected label.
  • Erase... : only clear the contents of the selected label.
  • Merge... : merge two labels by selecting a second label.
  • Random color : assign a random color to the label.
  • View only : make all labels except the current selected label invisible.
  • View/Hide all : make all labels visible / invisible
  • Lock/Unlock all : lock or unlock all labels.
  • Create surface : generate a surface out of the selected label.
  • Create mask : generate a mask out of the selected label. A mask is a binary image with "1" inside and "0" outside.
  • Create cropped mask : generate a binary mask out of the selected label. Crop changes the extent of the resulting image to the extent of the label.
Context menu for single label selection

Operations with multiple label selection

Shift-clickink on multiple labels allows to select more than one label. If more than one label is selected, different options will appear in the menu:

  • Merge selection on current label : transfer the contents of the selected labels in the 'Label Table' into the current one.
  • Remove selected labels : delete the selected labels.
  • Erase selected labels : only clear the contents of the selected labels.
Context menu for multiple label selection

Segmentation tool overview

MITK offers a comprehensive set of slice-based 2D and (semi-)automated 3D segmentation tools. The manual 2D tools require some user interaction and can only be applied to a single image slice whereas the 3D tools operate on the whole image. The 3D tools usually only require a small amount of user interaction, i.e. placing seed points or setting / adjusting parameters. You can switch between the different toolsets by selecting the 2D or 3D tab in the segmentation view.

An overview of the existing 2D and 3D tools in MITK.

2D segmentation tools

With 2D manual contouring you define which voxels are part of the segmentation and which are not. This allows you to create segmentations of any structures of interest in an image. You can also use manual contouring to correct segmentations that result from sub-optimal automatic methods. The drawback of manual contouring is that you might need to define contours on many 2D slices. However, this is mitigated by the interpolation feature, which will make suggestions for a segmentation.

To start using one of the editing tools, click its button from the displayed toolbox. The selected editing tool will be active and its corresponding button will stay pressed until you click the button again. Selecting a different tool also deactivates the previous one.
If you have to delineate a lot of images, shortcuts to switch between tools becomes convenient. For that, just hit the first letter of each tool to activate it (A for Add, S for Subtract, etc.).

All of the editing tools work by the same principle: using the mouse (left button) to click anywhere in a 2D window (any of the orientations axial, sagittal, or coronal), moving the mouse while holding the mouse button and releasing the button to finish the editing action. Multi-step undo and redo is fully supported by all editing tools by using the application-wide undo / redo buttons in the toolbar.

Remark: Clicking and moving the mouse in any of the 2D render windows will move the crosshair that defines what part of the image is displayed. This behavior is disabled as long as any of the manual segmentation tools are active - otherwise you might have a hard time concentrating on the contour you are drawing.

Add and subtract tools

Add and subtract tools

Use the left mouse button to draw a closed contour. When releasing the mouse button, the contour will be added (Add tool) to or removed (Subtract tool) from the current segmentation. Adding and subtracting voxels can be iteratively repeated for the same segmentation. Holding CTRL / CMD while drawing will invert the current tool's behavior (i.e. instead of adding voxels, they will be subtracted).

Paint and wipe tools

Paint and wipe tools

Use the Size slider to change the radius of the round paintbrush tool. Move the mouse in any 2D window and press the left button to draw or erase pixels. Holding CTRL / CMD while drawing will invert the current tool's behavior (i.e. instead of painting voxels, they will be wiped).

Region growing tool

Region growing tool

Click at one point in a 2D slice widget to add an image region to the segmentation with the region growing tool. Region Growing selects all pixels around the mouse cursor that have a similar gray value as the pixel below the mouse cursor. This allows to quickly create segmentations of structures that have a good contrast to surrounding tissue. The tool operates based on the current level window, so changing the level window to optimize the contrast for the ROI is encouraged. Moving the mouse up / down is different from left / right: Moving up the cursor while holding the left mouse button widens the range for the included grey values; moving it down narrows it. Moving the mouse left and right will shift the range. The tool will select more or less pixels, corresponding to the changing gray value range.

Fill tool

Fill tool

Left-click inside a segmentation with holes to completely fill all holes. Left-click inside a hole to fill only this specific hole.

Erase tool

Erase tool

This tool removes a connected part of pixels that form a segmentation. You may use it to remove single segmentations (left-click on specific segmentation) or to clear a whole slice at once (left-click outside a segmentation).

Live wire tool

Live wire tool

The Live Wire Tool acts as a magnetic lasso with a contour snapping to edges of objects.

Steps for using the Live Wire Tool
  • (1) To start the tool you have to double-click near the edge of the object you want to segment. The initial anchor point will snap to the edge within a 3x3 region.
  • (2) Move the mouse. You don't have trace the edge of the object. The contour will automatically snap to it.
  • (3) To fix a segment you can set anchor points by single left mouse button click.
  • (4) Go on with moving the mouse and setting anchor points.
  • (5) To close the contour double-click on the initial anchor point.
  • (6) After closing, the contour can be edited by moving, inserting and deleting anchor points.

The contour will be transferred to its binary image representation by deactivating the tool.

2D and 3D Interpolation

Creating segmentations using 2D manual contouring for large image volumes may be very time-consuming, because structures of interest may cover a large range of slices. Note: Interpolation is currently disabled for segmentations containing more than one label. The segmentation view offers two helpful features to mitigate this drawback:

  • 2D Interpolation
  • 3D Interpolation

The 2D Interpolation creates suggestions for a segmentation whenever you have a slice that

  • has got neighboring slices with segmentations (these do not need to be direct neighbors but could also be a couple of slices away) AND
  • is completely clear of a manual segmentation, i.e. there will be no suggestion if there is even only a single pixel of segmentation in the current slice.
2D interpolation usage

Interpolated suggestions are displayed as outlines, until you confirm them as part of the segmentation. To confirm single slices, click the Confirm for single slice button below the toolbox. You may also review the interpolations visually and then accept all of them at once by selecting Confirm for all slices.

The 3D interpolation creates suggestions for 3D segmentations. That means if you start contouring, from the second contour onwards, the surface of the segmented area will be interpolated based on the given contour information. The interpolation works with all available manual tools. Please note that this is currently a pure mathematical interpolation, i.e. image intensity information is not taken into account. With each further contour the interpolation result will be improved, but the more contours you provide the longer the recalculation will take. To achieve an optimal interpolation result and in this way a most accurate segmentation you should try to describe the surface with sparse contours by segmenting in arbitrary oriented planes. The 3D interpolation is not meant to be used for parallel slice-wise segmentation, but rather segmentations in i.e. the axial, coronal and sagittal plane.

3D interpolation usage

You can accept the interpolation result by clicking the Confirm-button below the tool buttons. In this case the 3D interpolation will be deactivated automatically so that the result can be post-processed without any interpolation running in the background.

Additional to the surface, black contours are shown in the 3D render window, which mark all the drawn contours used for the interpolation. You can navigate between the drawn contours by clicking on the corresponding position nodes in the data manager which are stored as sub-nodes of the selected segmentation. If you do not want to see these nodes just uncheck the Show Position Nodes checkbox and these nodes will be hidden.

If you want to delete a drawn contour we recommend to use the Erase-Tool since undo / redo is not yet working for 3D interpolation. The current state of the 3D interpolation can be saved across application restart. For that, just click on save project during the interpolation is active. After restarting the application and load your project you can click on "Reinit Interpolation" within the 3D interpolation GUI area.

3D segmentation tools

The 3D tools operate on the whole image and require usually a small amount of interaction like placing seed-points or specifying certain parameters. All 3D tools provide an immediate segmentation feedback, which is displayed as a transparent green overlay. For accepting a preview you have to press the Confirm button of the selected tool. The following 3D tools are available:

3D Threshold tool

The thresholding tool simply applies a 3D threshold to the patient image. All pixels with values equal or above the selected threshold are labeled as part of the segmentation. You can change the threshold by either moving the slider of setting a certain value in the spinbox.

3D Threshold tool

3D upper / lower threshold tool

The Upper/Lower Thresholding tool works similar to the simple 3D threshold tool but allows you to define an upper and lower threshold. All pixels with values within this threshold interval will be labeled as part of the segmentation.

3D upper / lower threshold tool

3D Otsu tool

The 3D Otsu tool provides a more sophisticated thresholding algorithm. It allows you to define a number of regions. Based on the image histogram the pixels will then be divided into different regions. The more regions you define the longer the calculation will take. You can select afterwards which of these regions you want to confirm as segmentation.

3D Otsu tool

3D Region growing tool

The 3D Region Growing tool works similar to the 2D pendant. At the beginning you have to place a seedpoint and define a threshold interval. If you press Run Segmentation a preview is calculated. By moving the Adapt region growing slider you can interactively adapt the segmentation result.

3D Region growing tool

Picking Tool

The Picking tool allows you to select islands within your segmentation. This is especially useful if e.g. a thresholding provided you with several areas within your image but you are just interested in one special region.

Picking tool

nnU-Net Tool (Ubuntu only)

nnUNet tool

This tool provides a GUI to the deep learning-based segmentation algorithm called the nnUNet. With this tool, you can get a segmentation mask predicted for the loaded image in MITK. Be ready with the pre-trained weights (a.k.a RESULTS_FOLDER) for your organ or task concerned, before using the tool. For a detailed explanation of the parameters and pre-trained weights folder structure etc., please refer to
Remark: The tool assumes that you have a Python3 environment with nnUNet (pip) installed. Your machine should be also equipped with a CUDA enabled GPU.


  1. Select the "Python Path" drop-down to see if MITK has automatically detected other Python environments. Click on a fitting environment for the nnUNet inference or click "Select" in the dropdown to choose an unlisted python environment. Note that, while selecting an arbitrary environment folder, only select the base folder, e.g. "myenv". No need to select all the way until "../myenv/bin/python", for example.
  2. Click on the "nnUNet Results Folder" directory icon to navigate to the results folder on your hard disk. This is equivalent to setting the RESULTS_FOLDER environment variable. If your results folder is as per the nnUNet required folder structure, the configuration, trainers, tasks and folds are automatically parsed and correspondingly loaded in the drop-down boxes as shown below. Note that MITK automatically checks for the RESULTS_FOLDER environment variable value and, if found, auto parses that directory when the tool is started.
    nnUNet Segmentation Settings
  3. Choose your required Task-Configuration-Trainer-Planner-Fold parameters, sequentially. By default, all entries are selected inside the "Fold" dropdown (shown: "All"). Note that, even if you uncheck all entries from the "Fold" dropdown (shown: "None"), then too, all folds would be considered for inferencing.
  4. For ensemble predictions, you will get the option to select parameters irrespective on postprocessing files available in the ensembles folder of RESULTS_FOLDER. Note that, if a postprocessing json file exists for the selected combination then it will used for ensembling, by default. To choose not to, uncheck the "Use PostProcessing JSON" in the "Advanced" section.
    nnUNet Segmentation Settings
  5. If your task is trained with multi-modal inputs, then "Multi-Modal" checkbox is checked and the no.of modalities are preloaded in the "No. of Extra Modalities" spinbox. Instantly, as much node selectors with corresponding modality names should appear below to select the Data Manager along including a selector with preselected with the reference node. Now, select the image nodes in the node selectors accordingly for accurate inferencing.
    nnUNet Multi Modal Settings
  6. Click on "Preview".
  7. In the "Advanced" section, you can also activate other options like "Mixed Precision" and "Enable Mirroring" (for test time data augmentation) pertaining to nnUNet.
    nnUNet Advanced Settings
  8. Use "Advanced" > "GPU Id" combobox to change the preferred GPU for inferencing. This is internally equivalent to setting the CUDA_VISIBLE_DEVICES environment variable.
  9. Every inferred segmentation is cached to prevent a redundant computation. In case, a user doesn't wish to cache a Preview, uncheck the "Enable Caching" in the "Advanced" section. This will ensure that the current parameters will neither be checked against the existing cache nor a segmentation be loaded from it when Preview is clicked.
  10. You may always clear all the cached segmentations by clicking "Clear Cache" button.


  1. In case you want to reload/reparse the folders in the "nnUNet Results Folder", eg. after adding new tasks into it, you may do so without reselecting the folder again by clicking the "Refresh Results Folder" button.
  2. The "Advanced" > "GPU Id" combobox lists the Nvidia GPUs available by parsing the nvidia-smi utility output. In case your machine has Nvidia CUDA enabled GPUs but the nvidia-smi fails for some reason, the "GPU Id" combobox will show no entries. In such a situation, it's still possible to execute inferencing by manually entering the preferred GPU Id, eg. 0 in the combobox.
  3. In the RESULTS_FOLDER directory, inside the trainer-planner folder of every task, MITK keeps a "mitk_export.json" file for fast loading for multi-modal information. It is recommended not to delete this file(s) for a fast responsive UI. Tip: If multi-modal information shown on MITK is not correct for a given task, you may modify this JSON file and try again.

Additional things you can do with segmentations

Segmentations are never an end in themselves. Consequently, the segmentation view adds a couple of "post-processing" actions, accessible through the context-menu of the data manager.

Context menu items for segmentations
  • Create polygon model applies the marching cubes algorithm to the segmentation. This polygon model can be used for visualization in 3D or other applications such as stereolithography (3D printing).
  • Create smoothed polygon model uses smoothing in addition to the marching cubes algorithm, which creates models that do not follow the exact outlines of the segmentation, but look smoother.
  • Autocrop can save memory. Manual segmentations have the same extent as the patient image, even if the segmentation comprises only a small sub-volume. This invisible and meaningless margin is removed by autocropping.

Segmentation of 3D+t images

For segmentation of 3D+t images, some tools give you the option to choose between creating dynamic and static masks.

  • Dynamic masks can be created on each time frame individually.
  • Static masks will be defined on one time frame and will be the same for all other time frames.

In general, segmentation is applied on the time frame that is selected when execution is performed. If you alter the time frame, the segmentation preview is adapted.

Technical information for developers

For technical specifications see Technical design of QmitkSegmentation and for information on the extensions of the tools system How to extend the Segmentation view with external tools.