Lab 02 : CT Images


  • Visualization We will take a look at the MITK software and inspect the 3D visualization of CT images.
  • Classification We will go through a short introduction to classification and look at the ilastik software.

HW02 Homework

(A) CT Data Processing

Download the MITK Workbench application. Load the testing image lowdose_CT.nii, if the data are too large for your PC to process, you can try the smaller lowdose_CT_cropped.nii.gz file. Optionally you can look at T1-weighted images of lower leg YM5_MRI_Data_Combined.nii to see the differences between CT and MRI images.

After the file is opened, the image's name will appear as entry in the Data Manager view, you will also see four rendering windows, they show the standard anatomical planes and a 3D-visualization of the image. The Level Window slider appears on the right.

  1. [0.5 pt] Adapt the Level Window slider to show only bone structures in the image.
    • purpose of the Level Window is to define how the image values, which are in Hounsfield (internally as int32 datatype) are mapped to the standard display range. Is the level window set to a range $[W_l, W_u]$, then all values lower or equal $W_l$ will have intensity 0, values equal or larger than $W_u$ will have intensity 255 and the intensity of points with values within $[W_l, W_u]$ are linearly interpolated.
    • it is typically defined by its mean value $(W_u + W_l) / 2$ and its span $W_u - W_l$, these values can be entered into the boxes below the slider
    • Which settings for LW did you use? Insert a screenshot for each image in your report.
  2. [0.5 pt] Create a segmentation of bone structures. Open the Segmentation plugin and use the UL Threshold segmentation tool from the 3D Tools pane. Select a lower and upper threshold and create a segmentation. Create a surface representation (right-click on the segmentation node in the DataManager opens a context menu, choose the entry Create polygon model) and make a screenshot for the report. Comment on your choice of thresholds used for segmentation
    • first, you need to select Create new segmentation in the upper part of the segmentation plugin, the tools will be active afterwards
  3. [1 pt] Manual segmentation task
    • manual annotation is still the most common task done by radiologists to create segmentation.
    • Add a new segmentation to the femur image and use a tool of your choice to create a segmentation of muscle structure in three consecutive axial planes
    • Report which tool you used, export the segmentation as '.nrrd' format (Save segmentation image from Data Manager) and upload it together with your report (the segmentation image should not be larger than 10MB).
  4. [1 pt] Volume Visualization Open the Volume Rendering plugin, adapt the transfer function to get a 3D visualization of bone structures, put a screenshot (from the 3D rendering window) into your report.

(B) CT Image Segmentation

Segmentation You have experienced, how tedious a manual segmentation can be, which is one of the main motivations for development of (semi-)automated segmentation methods. In this part, we will look at segmentation by means of pixel classification. It is about finding set of values (features) for each pixel, that allows to construct a criterion which assigns a class label to each pixel. The threshold-based segmentation from first part is a simple example of this principle. We have the pixel value $I(x,y)$ as feature and define the criterion for the bone class as $\theta_l \leq I(x, y) \leq \theta_u$.

In this task, we will look for a larger set of features for each pixel, that will lead to a criterion for classes bone, muscle, fat and background.

Pixel classification

[2 pts] Homework task

Download the ilastik application. Download the input data archive

In ilastik:

  1. Create a new Pixel Classification project.
  2. Load the training images from the training folder within the data archive.
  3. Select a set of features to be considered, ilastik offers pixel value, edge information, laplacian and structure tensor at different smoothing levels.
  4. Define four classes: Background,Bone,Fat,Muscle
  5. Draw some annotation examples for each class
  6. Turn on 'Live Update' to see, how the classifier recognizes the four classes, turn on the Uncertainty layer to see areas, where the classifier may need additional labels, switch also between training images, and add labels if needed.
  7. Save the segmentation maps for all testing images and include them in the report.
courses/zsl/labs2024_02_ct_images.txt · Last modified: 2024/02/29 13:13 by anyzjiri