[Resources] [Terminology] [Units] [MCQs]


   

Threshold Based Segmentation


90 min. ASSESSED
Objectives:
At the end of this practical you should understand:
- the characteristics of anatomical MR brain images
- the use of thresholding for region segmentation
- the value of adaptive thresholding for industrial scene analysis

Terminology:


See:
Image Histograms
Thresholding
Grey Levels
This exercise is to be carried out using MATLAB and a report is to be written as part of your course assessment. A moderate degree of collaboration in carrying out the work is acceptable but reports should be entirely your own. You can expect to have to draw upon both course work and previous practical experience. In each case you should explain the objective of the exercise, outline the combination of processes you used, explain the results you obtained, the scope (ie general usefulness) of the method and try to address any specific questions raised. You do not need to explain the basics of MATLAB but inclusion of the MATALB routines you develop will be expected. Tackle the following two problems and produce a write-up of approximately 1-2 pages for each.

Download the .pgm images above for analysis in MATLAB.

The upper image is a Nuclear Magnetic Resonance image of the skull and brain. In such images, the distinct tissue types have distinct grey level values.

Use the ROI histogramming routines you already have to identify regions of white matter in the histogram of the upper image. For reference a white matter segmentation is shown in the lower image of the main "Region Segmentation" unit page.

Use a combination of thesholds above and below the white matter peak and some simple image arithmetic to produce your own pixel level binary segmentation of white matter.

Now attempt the same process with the lower image, which was obtained from a CCD camera in an industrial environment. This time attempt to extract the darker contents of each circular cell.

You will notice that illumination artefacts cause position dependant results. Now filter the image with a large scale smoothing kernal, subtract the result from the original image and try again.

Can you improve upon your initial segmentation?



(c) Imaging Science and Biomedical Engineering 2000 [paul.bromiley@man.ac.uk]

Valid HTML 4.01!