WMLS: White Matter Lesion Segmentation
Brain Lesion Segmentation
We are working on effective ways of using multi-modality image parameters for segmentation of brain lesions. We obtain multi-modal image parameters by aligning images of a same subject from different modalities and contrasts through a rigid image co-registration algorithm that maximizes a mutual information-based similarity measure. The collection of image intensities corresponding to the same location from all available modalities then forms a multi-modality imaging signature. We use a manually delineated set of lesion definitions to characterize the imaging signatures of lesions in contrast to the imaging signatures of healthy brain tissue. To this end, we formulate a classification problem that invokes a very flexible support vector machine learner to separate the classes of multi-modal image signatures of the lesion tissue from those of the healthy tissue. We further invoke spatial statistics of lesion distribution as well as additional statistics that correct false positives due to compounding factors such as co-registration errors that are most prominent near the cortical surface.
Fig.1. Visual demonstration of the discriminating abilities that different combinations of MR acquisition protocols have. One direction of research in our group aims to combine multi-parametric MR images via machine learning methods, for the segmentation of brain lesions, but also for the identification of brain tissue that presents subtle abnormalities that might be predictors of subsequent pathology or clinical progression.
Fig.2. Left is the result of voxel-wise evaluation map showing different lesion rating for each voxel. Right is WML segmentation result after thresholding the map on the left.
Fig.3. ROC curve showing a general performance of our WML segmentation method over a population of 45 subjects.
- Dinggang Shen
- Zhiqiang Lao
- Songyang Yu, Dzung Pham, Dinggang Shen, Edward H. Herskovits, Susan M. Resnick, Christos Davatzikos,""Automatic Segmentation of White Matter Lesions in T1-Weighted Brain MR Images",IEEE International Symposium on Biomedical Imaging, Washington, D.C. 7-10 July 2002.