White matter lesion segmentation based on feature joint occurrence probability and χ2χ2 random field theory from magnetic resonance (MR) images

Lesions of the brain’s white matter are common findings in MR examinations of elderly subjects. A fully automatic method for segmenting white matter lesions is proposed here. The joint probability of multi-modality MR image intensities is used as a feature to segment lesions, because lesion intensities usually are outliers of the normal tissue intensities and the lesions’ joint intensity probability appears much smaller than those of normal brain tissues. The χ2χ2 random field theory is used to determine the significance of a detected lesion and provides a strict statistical analysis to exclude small-sized false-positive lesions. Experimental results show that the automatic segmentation of lesions is in high agreement with manual segmentation, and the χ2χ2 random-field-based statistical analysis greatly improves lesion segmentation results.