2010 Winter Project Week RobustStatisticsDrivenActiveContourSegmentation

• 

  Projects List

• 

  Segmentation of kidneys, spleen, and liver. Segmented separately.

• 

  Simultaneous multi-object segmentation.

• 

  Simultaneous multi-object segmentation.

• 

  Highlighted in the circle, the objects are touching, but no overlapping. In fact, with shorter iteration, heart had been "invading" liver. But with long iteration, liver surface pushes heart surface out.

• 

  Recording the surface evolution history. Top left: result after 200 iterations. Then back to 160, 120, 80, 40, and 20 iteration. Though this level set evolution can not run backward. So when users found result at 190-th iteration is better than 200-th's, they don't have to run from 0 to 190 again.

Key Investigators

• Yi Gao, Allen Tannenbaum, Georgia Tech
• Andriy Fedorov, Katie Hayes, Ron Kikinis

Objective

Develop a general purpose active contour segmentor in Slicer3.

More detail and testing are at here.

Approach, Plan

Driven by the robust statistics, we provide an active contour segmentor module for general purpose in Sclier3. The algorithm will simultaneously extract multiple objects in the scene given some user provided labels.

Plan for project week: Integrated into Slicer3 using interactive module.

Progress

The CLI is already on Extension server and ready to be used through extension manager in Slicer.

The next step is to change the user input manner from fiducial points to label maps.

Furthermore, we plan to use multiple labels to simultaneously extract multiple objects.

Work done in this week

• Use label map, instead of fiducial points, as input. This brings two benefits:
  • User friendly: easier to set more seeds ===> better knowledge.
  • Assign different labels.

• Simultaneous multi-object segmentation. The surfaces of different object compete when they touch each other, and stop at balanced position.

• Keep the history. So can "iterate backward", though this level set evolution function can not go backward. So when users found the result at 190-th iteration is in fact better than the 200-th's, they don't have to run from 0 to 190 again, but just use a scroll bar to pull it back.

References

• NA-MIC Publications Database
• Yang, Y. and Tannenbaum, A. and Giddens, D. and Coulter, WH, "Knowledge-based 3D segmentation and reconstruction of coronary arteries using CT images", in IEEE EMBS 2004, pp1664--1666