Difference between revisions of "2013 Project Week:4DShapeAnalysis"

From NAMIC Wiki
Jump to: navigation, search
Line 57: Line 57:
 
<h3>Volume analysis after regression</h3>
 
<h3>Volume analysis after regression</h3>
  
Here we measure the volume of the estimated baseline shape as well as the volume of the estimated shape at the final time point.
+
Here we measure the volume of the estimated baseline shape as well as the volume of the estimated shape at the final time point.  We show the volume estimated from regression as well as the observed volume in parentheses.  Recall that there is also a third observation somewhere in between the initial and final time point that also influences estimation.  Finally, note that the estimation is performed jointly on all shapes simultaneously.
  
 
'''HD Subjects'''
 
'''HD Subjects'''

Revision as of 18:05, 11 January 2013

Home < 2013 Project Week:4DShapeAnalysis

Key Investigators

  • James Fishbaugh, Marcel Prastawa, Guido Gerig, SCI
  • Hans Johnson, Iowa

Project Description

Objective

In past project weeks, we have explored the application of shape regression to study sub-cortical changes associated with Huntington's disease. This has led to the development and release of open source shape analysis software. During this project week, we would like to investigate a new algorithm recently developed for our software. The method is the extension of linear regression to shapes, where we jointly estimate a baseline shape (intercept) and initial momenta vectors (slope) which define the dynamics of shape change over time. This is in contrast to the current implementation, which assumes the baseline shape is fixed at the earliest time point.

Approach, Plan

  • Individual regression with control points at shared locations
    • Compare to previous results
    • Compare differences between HD subjects and controls

Progress

Segmentation

Here we consider a multi-shape complex of caudate, putamen, and thalamus, which were provided as binary images. Triangular meshes were extracted using marching cubes, giving us 3 shapes at 3 time points for each subject.

Regression

The continuous evolution of all 3 structures is estimated jointly, using a shared control point parameterization.

Conclusion

  • As before, segmentation quality is the most important factor in the analysis. The changes we are interested in detecting could potentially be of the same order as segmentation error.
  • Estimating the baseline shape may improve the robustness of the estimation,
  • However, we need to be careful about the estimation of the baseline shape. The changes over time are very subtle, therefore small differences in the baseline shape could lead to completely different dynamics of shape change.

Summary

Volume analysis after regression

Here we measure the volume of the estimated baseline shape as well as the volume of the estimated shape at the final time point. We show the volume estimated from regression as well as the observed volume in parentheses. Recall that there is also a third observation somewhere in between the initial and final time point that also influences estimation. Finally, note that the estimation is performed jointly on all shapes simultaneously.

HD Subjects

Shape Initial Volume (mm^3) Final Volume (mm^3) Percent Change
Subject 10001 Caudate 2298 (2338) 2289 (2266) -0.39%
Putamen 3042 (2815) 2894 (2743) -4.86%
Thalamus 6384 (6270) 5961 (5685) -6.63%
Subject 10018 Caudate 3059 (3261) 3065 (3091) 0.20%
Putamen 3521 (3468) 3416 (3375) -2.98%
Thalamus 6992 (7214) 7449 (7568) 6.54%
Subject 10027 Caudate 2397 (2330) 2352 (2296) -1.88%
Putamen 3194 (3091) 2993 (2846) -6.29%
Thalamus 7278 (6959) 7287 (7291) 0.12%


Controls

Shape Initial Volume (mm^3) Final Volume (mm^3) Percent Change
Subject 10016 Caudate 4015 4051 0.90%
Putamen 4750 4741 -0.19%
Thalamus 8088 8049 -0.48%
Subject 10017 Caudate 2094 2093 -0.05%
Putamen 3871 3879 0.21%
Thalamus 5915 5846 -1.17%

Regression results