Bone Segmentation
Kevin Baldor
Steps
- Choosing the Dataset
1. Cryoslices
2. freshCT
3. frozenCT
- Image Processing
1. thresholding
2. morphology
- Model production
1. iso-contours
2. tiling
1. Choosing the Dataset
To produce the bone geometry dataset, I considered three datasets:
- Cryoslices : these images were obtained by photographing slices of the body. They posess several attributes which make them desirable.
- high resolution - allows for fine details to be resolved
- low noise - the high resolution is usable
- color - bone,internal organs,etc are segmentable
- FreshCT : these images were obtained through computed tomography. This technology has its advantages and disadvantages
- non-destructive - can be perfomed on living people
- noisy - X-ray opaque objects produce shadows
- fresh - organs are more visible
- Frozen CT : these images were obtained after the subject was frozen
- high contrast - bones are fairly easy to find
- some noise - when the arms protrude beyond the imageing area, they produce shadows which are not accounted for in other scans producing streaks
- missing bones - the elbows are outside of the imaging area
Of these, I chose the frozen CT dataset because I was unable to produce an automated segmentation scheme for finding bone in the cryoslice data, and the frozen CT dataset had the lowest noise.
2. Image Processing
To segment the bone, some image processing was required:
- Thresholding : a variable threshold was used to highlight bone
- variable threshold - allows for different gray values associated with bone. This was caused by the different imaging distances
- Morpholgy : these techniques alter the shape of the data to remove artifacts
- erode - removes small objects but decreases the size of everything
- dilate - fills small holes but increases the size of everything
- close - tends to connect fine lines which were lost - useful in the skull
3. Model Production
While this was not intended to be part of my contribution, it became necessary. We considered two approaches to model production.
- Isocontouring : the model is defined by a mesh which is defined by a constant value.
- benifits - it is useful for models for which the values have significance
- detractors - doesnt take advantage of the properties of the desired body part in the visual human dataset
- compormise - preprocess the image to accentuate the desired feature
- Tiling : the model is produced by connecting the outline of a body on two adjacent slices. This was looked into by Florian Mayer
- benifits - well suited to the dataset
- detractors - unavailable, the code to perform this task was non functional
This task was acheived through image processing and isocontouring. However, the isocontouring code did not handle binarized data well and created models which could not be downsampled using the Zaiqing attempted to use.
4. Lessons Learned
The most important lessons that I have learned from this course and project involve the techniques which did not work well for me.
- Isocontouring :
- not well suited to the Visible Human Dataset
- Tiling :
- well suited to the dataset
- Dataset Size Reduction :
- Should be performed on the geometry dataset, not the volume data