DBP3:Utah:AutoEndoSeg

Current Activities

• 11/30/11 Greg and Josh met with Yi via Skype
  • Algorithm is still very sensitive to image cropping -- registration seems to fail unless cropping is exact
  • We can input an atlas image and get a correct result
  • Yi will investigate multiscale registration to see if that improves result
  • Yi will look for any bugs in code that might be affecting registration
  • Greg will send Yi current data for new atlas with larger crop region

Notes on the Endocardial Segmentation Algorithm

Algorithm is a multi-atlas segmentation.

1. Register (cropped) heart region of patient DEMRI to all of the n atlas DEMRI images and record n minimum metric values (mutual information)
2. Transform all n atlas segmentations to patient DEMRI space
3. Segmentation is a weighted average of the transformed atlas segmentation, where weights are determined by the mutual information values from step 1
4. Pick a threshold as your segmentation

Notes on Yi's Slicer Module

• The exposed parameter is the ratio of pixel samples for mutual information to the total number of pixels in the image.
• We could also expose the threshold parameter in the final step
• We can change or add atlas datasets by copying into the appropriate directory and editing the text file that lists them
• Uses the ITK mutual information code. Does affine registration followed by b-spline
• We will partner with Yi for module development at the Jan 2012 project week

Initial Experiences

• Comparison of GA Tech and Utah-derived segmentations

(click to enlarge)

GA Tech Segmentation	Utah Segmentation	Thresholding Difference
Segmentation of an atlas image by GA Tech, using their endocardial auto-segmentation algorithm.	Segmentation of the same image with the same algorithm by Utah.	User-based variability in thresholding the probability mask output. (Green region - Utah; white region - GA Tech).

Larger Atlas Images 11/30/11

• A second atlas was generated from cropped images that included a larger region of the initial MRI scans (namely, the appendage and a greater extent of the PVs was included in this second atlas).
• This atlas showed no marked improvement over the initial atlas
• Segmentations with the second atlas tended to not capture the entire blood pool, whereas, the first atlas tended to overestimate the blood pool.
• Registration issues?

	Mask and LGE Image	Mask and Expert-defined Segmentation
Example 1
Example 2
Example 3
Example 4
Atlas Image
	Cropped MRI image overlaid with the algorithm-defined blood pool.	Expert manual segmentations (white) overlaid with the algorithm-defined blood pool (green).

Note: The last segmentation (red) was derived from an image in the atlas; the other segmentations (green) were derived from similarly-cropped images, not in the atlas.