Image Analysis - Segmentation, Registration and Shape Measurement  MIDAG - Medical Image Display and Analysis Group
Introduction
Of the presently available methods for segmentation of anatomic objects from medical images, principally interactive ones are the most commonly used in clinical practice. More automatic techniques are based on statistical pattern recognition (thresholding), mathematical morphology, deformable models with local geometric constraints (active surfaces), or deformable structural models, i.e., using models that involve nonlocal spatial relationships. The methods involving the deformation of structural models seem distinctly the most successful, and they lend themselves not only to segmentation but also to shape measurement and registration. It is on these that we focus. Access to our papers on object segregation and shape representation, subdivided into topic areas, can be made through the Bibliography page. A tutorial on object shape representation is available here.

A deformable structural model for one or more objects can be based on a variety of primitives. Among the possible primitives are voxels (with intensity) [e.g., Miller, Christensen, Joshi], landmarks [e.g., Bookstein, Kendall], sampled boundary points [e.g., Cootes & Taylor], parameterized continuous boundaries [e.g., Kelemen & Gerig, Staib & Duncan], and medial atoms [Pizer, Siddiqi, Székely, Kimia]. We focus on medial methods, especially those based on the m-reps representation (Fig. 1), on the parameterized continuous boundary representation based on spherical harmonics, and on tubes found by intensity or medial strength ridges. We are also working to develop methods for the validation of our segmentation methods in inexpensive, automatic, generalizable ways.
wireframe image of the m-rep model matching the m-rep model against a dataset
Fig. 1. An m-rep model for a kidney. Left, in 3D: the m-rep is made from 2 linked figures, each of which is represented by a mesh or chain of medial atoms. Here the two figures represent the whole kidney and the pelvis + ureter. The medial atoms are shown in white with the meshes in green and the medially implied boundaries in blue. Right, in 2D: a cut through the 3D image data showing 2D medial atoms interpolated from the 3D atoms for the figure representing the whole kidney. 

Fig 2. Subcortical brain structures via statistics on spherical harmonics coefficients. 

Fig. 3. Tree of vessels extracted from magnetic resonance angiogram, combined with volume rendered visualization of an arterial-venous malformation. 

 
Clinical Problems
The deformable models work is driven by a variety of clinical problems: 
  • Planning and verification of radiotherapy 
  • Neurosciences: science and diagnosis of structures of the brain and their relation to various mental illnesses 
  • Neurosurgery planning and delivery of neurosurgery involving the vascular system in the brain or involving the spine 
  • Analgesic nephropathy
  • Orthopedic surgery
These problems come both from the UNC School of Medicine and the Duke School of Medicine and from collaborating groups at Harvard University, Johns Hopkins University, and Memorial Sloan-Kettering Cancer Center. The problems involve us in segmentation and shape measurement of the cerebral ventricle from both MRI and 3D ultrasound, segmentation of the pelvic bones from CT and prostate from MRI, segmentation and shape measurement of the kidney from CT, segmentation and shape measurement of subcortical brain structures, especially the hippocampus, from MRI, and successive segmentation over 26-40 weeks of gestational age of the premature infant's cortex and basal ganglia from MRI. 
 
Subproject Links
The following pages go into further depth in the various aspects of this project: 
 
Participants
Faculty focusing on methods for segmentation, registration, and shape measurement include: 
  • Stephen Pizer, Kenan Professor (Computer Science, Radiology, Radiation Oncology, Biomedical Engineering) 
  • Guido Gerig, Taylor Grandy Professor (Computer Science, Psychiatry) 
  • Ed Chaney, Professor (Radiation Oncology) 
  • Julian Rosenman, Professor (Radiation Oncology, Computer Science) 
  • James Damon, Professor (Mathematics) 
  • Stephen Marron, Professor (Statistics) 
  • Valen Johnson, Professor (Statistics, Duke) 
  • James Coggins, Associate Professor (Computer Science) 
  • Stephen Aylward, Assistant Professor (Radiology, Computer Science)