EMS

Expectation-Maximization Segmentation

Fully automated model-based segmentation of MR images of the brain

Developed by Koen Van Leemput at Medical Imaging Computing, Leuven, Belgium

Press the 'Lesions' button in the EMS menu window. Select '1' for 'number of subjects', all three images (order is not important), '4' for the order of the bias field polynomial model, '3D' for the type of polynomial, '3.00' for 'Mahalanobis threshold?' (this determines the significance level at which voxels are considered as model outliers), 'no' for 'Use Markov random field?', empty string for 'Intensity constraint on lesions' (simply press ENTER), and 'no' for 'Look for other outliers as well?'. Now the multi-spectral data set is automatically being segmented into normal tissue types that follow the model assumptions on the one hand, and MS lesions as model outliers on the other hand.

As with the segmentation of normal brain images, the results are written in files with suffix '_seg' and '_bias' for the tissue class probability images and the estimated bias fields, but now there is also an additional file with suffix '_lesion' that contains the automatically segmented MS lesions, and a file with suffix '_garbage' that is empty in our case. Using the 'Check Reg' option of SPM, and selecting the original T2-weighted image and the lesion image will result in the figure shown on the left.

In the "quick & dirty"-section, all model outliers were assigned to the MS lesion class. Unfortunately, outlier voxels also occur outside MS lesions, typically on the border between two tissue types where so-called partial volume voxels violate the model assumption that every voxel belongs to only one single tissue type. True MS lesions can be discerned from such non-lesion outliers by incorporating extra a priori information about the appearance and the location of the lesions. For instance, lesions can be seen to be brighter than gray matter in the T2-weighted image. Further, it is known that the vast majority of MS lesions occurs inside white matter. Such extra information can be provided to the automatic segmentation program as described below.

Press the 'Lesions' button in the EMS menu window. Select '1' for 'number of subjects', all three images (order is not important, but do remember what number the T2-weighted image is), '4' for the order of the bias field polynomial model, '3D' for the type of polynomial, '4.00' for 'Mahalanobis threshold?', and select now 'yes' for 'Use Markov random field?'. Enter now 'i1>=gm1' for 'Intensity constraint on lesions' if the T2-weighted image was the first image you selected (otherwise use 'i2>=gm2' or 'i3>=gm3' if it was the second or the last image, respectively). This will prevent voxels with a T2-intensity lower than the estimated mean intensity of gray matter in the T2-weighted image from being classified as MS lesions. Select 'wm' (white matter) for 'To which tissue do lesions belong?'. An automatically trained Markov random field model will now discourage voxels from being classified as MS lesion in the absence of neighboring white matter. Finally, select 'no' for 'Look for other outliers as well?'.

The program now runs considerably slower, due to the incorporation of the Markov random field model and the automatic estimation of its parameters. However, the result, shown on the left, will be much "cleaner" than the one obtained with the "quick & dirty"-method.