PosterTemplateLargeJLP

1
Fast Algorithm for Probabilistic Bone Edge
Detection FAPBED cepanovic D, Kirshtein J, Jain
A, Taylor R

• The Problem
• Find a fast, probabilistic framework for
  segmenting bone surfaces within a CT volume
• Analyze the accuracy of the segmentations
• Produce a bone surface probability map that can
  be used to register US scans to the CT volume

Automatic Segmentation
Ideal Bone Surface Overlap Validation Plot

• Results
• 6 methods are probabilistically compounded to
  yield final probability map (run on 1.8GHz W2K
  Matlab 6.5)
• Run time for 512x512 voxel image is 2.4 sec
• Estimated time for 60 slice CT volume is 3 min
• Validation map has a sharp peak at the origin,
  similar to the ideal map and better than any
  individual method

• Significance
• Registration is necessary for many CIS procedures
• Registration is often invasive, inconvenient and
  time consuming, but it does not have to be!
• Ultrasound is a convenient, real-time method for
  collecting many bone surface points
• Probabilistic US to CT registration is a
  promising alternative to invasive registration
  techniques
• In order for this to work a probabilistic bone
  surface map must be generated

Methods 6 methods used to generate final
image Many methods tested 3 gradient methods, 10
statistical methods, 4 discrete methods Figures
below probability curve given voxel value,
processed image, and a validation map of the
processed image

• Future Work
• Consider incorporating cross-slice information
  into probabilistic framework
• Increase accuracy and speed of algorithms
• Fine tune probability-given-feature functions by
  analyzing more manually segmented bones
• Extend FAPBED to different bones and anatomy

Absolute Threshold - Statistical

• Lessons Learned
• Automatic segmentation is not a trivial task
• Probabilistic framework is very flexible and
  enables enhancement of desirable qualities while
  suppressing noise
• Efficient planning and execution are critical for
  project success

XY Gradient - Gradient
People Grad Students Danilo cepanovic, Josh
Kirshtein Mentor Ameet Kumar Jain ERC Faculty
Dr. Russell Taylor Special Thanks Ofri Sadowsky
Edge From Threshold - Discrete
Engineering Research Center for Computer
Integrated Surgical Systems and Technology
The Johns Hopkins University