Corrective Learning: Learning classifiers for correcting segmentation errors for a host segmentation method. usage:

  • bl dim [options] AdaBoostOutputPrefix

required options:


Image dimension (2 or 3) -ms sub1_mseg.nii ... subN_mseg.nii Manual segmentations -as sub1_aseg.nii ... subN_aseg.nii Automatic segmentationds produced for the training images by the host segmentation method

-tl Tlabel

Our method learns to correct one label at a time. This parameter specifies the target label that is going to be learned by this learning task.

-rd dilationR

This parameter specifies the region of interest (ROI) for this learning task. A label's working ROI is defined as the region assigned with the Tlabel by the host segmentation plus some dilation (unless a mask is provided, see below). dilationR specifies the dilation radius. The dilated ROI should cover most or all voxels that are manually assigned to Tlabel Default: 2x2x2

-rf featureRadius

Patch radius for feature extraction. scalar or vector (AxBxC) Default: 2x2x2

-rate sampleRate

0<= sampleRate <=1. When the ROI is large or there are too many training images, loadng every voxel in ROI as a training sample may be impossible due to the memory limit. To address this problem, sampleRate specifies the percentage voxels from ROI that will be problem, sampleRate specifies the percentage voxels from ROI that will be used used. If sampleRate=0.01, 1 percent voxels will be used. Default: 1

-i iteration

Training iteration for AdaBoost learning. Default: 100

  • options:

-c featureChannel

Number of feature channels

-f sub1_feature1.nii sub1_feature2.nii ... subN_feature1.nii subN_feature2.nii ...

Feature images for the training subjects. The number of feautres for each subject should equal featureChannel.

-m mask1 ... maskN:

Specify labels' working ROI for the training images. ROI will be derived by performing dilation on this mask then.