Model-based segmentation of brain tissue and tumor

Guido Gerig, Nathan Moon, Sean Ho, Elizabeth Bullitt

Research output: Contribution to journalConference articlepeer-review


Combining image segmentation based on statistical classification with a geometric prior shows a significant increase in robustness and reproducibility. Using a probabilistic geometric model of sought structures helps with initialization of probability density functions and defines spatial constraints. A strong spatial prior, however, prevents segmentation of structures that are not part of the model. In practical applications presenting pathology, e.g. brain tumors, we encounter the presentation of new objects not modelled by a statistical atlas of healthy anatomy. We developed an extension to an existing expectation maximization segmentation (EM) algorithm that modifies a probabilistic brain atlas with individual subject's information which results in an automatic, atlas-based classification of healthy tissue and pathology. Object segmentation requires spatial grouping of classified voxels. For tumors, we use the geometric model of blobby shaped objects and use region-competition level-set evolution. The statistical classification of the classification step serves as a probabilistic measure for object and background. The geodesic snake, initialized by the noisy tumor classification and locally driven by object/background probabilities, generates smooth, blobby structures presenting tumors. The new brain and tumor segmentation method has been tested on five cases presenting different tumor types and shapes. A preliminary validation compared these segmentations with manual expert's results.


  • Brain tissue segmentation
  • Expectation maximization
  • Level-set evolution
  • Tumor segmentation

ASJC Scopus subject areas

  • Signal Processing
  • Biomedical Engineering
  • Computer Vision and Pattern Recognition
  • Health Informatics


Dive into the research topics of 'Model-based segmentation of brain tissue and tumor'. Together they form a unique fingerprint.

Cite this