Brain%20Tumor%20Segmentation:%20Label%20each%20voxel%20in%20MR%20image%20as%20{%20tumor,%20non-tumor%20}

1
Segmenting Brain Tumors using PseudoConditional
Random Fields Chi-Hoon Lee, Shaojun Wang, Albert
Murtha, Matthew Brown, and Russell Greiner
S-38

• Brain Tumor SegmentationLabel each voxel in MR
  image as tumor, non-tumor
• Use only individual voxels
• Discriminative classifier (Logistic Regression
  SVMs)
• Also use spatial correlations of labels among
  neighboring voxels
• Random Fields potential for voxel potential
  for neighboring voxels
• Extension Pseudo-Conditional Random Fields
• Learn
• Learn discriminative iid classifier for each
  voxel
• Hand-tune potential for neighbors
• Inference
• Uses both potentials
• Incorporates label correlations in 2-D MR image
• Contributions
• Learning is significantly faster than typical
  CRFs
• Quality of resulting segmentation ? typical CRFs

Brain Tumor Analysis Project
http//www.cs.ualberta.ca/btap