A learning-based wrapper method to correct systematic errors in automatic image segmentation: consistently improved performance in hippocampus, cortex and brain segmentation.

摘要: We propose a simple but generally applicable approach to improving the accuracy of automatic image segmentation algorithms relative manual segmentations. The is based on hypothesis that large fraction errors produced by are systematic, i.e., occur consistently from subject subject, and serves as wrapper method around given host method. attempts learn intensity, spatial contextual patterns associated with systematic training data for which segmentations available. then correct such in new images. One practical use proposed adapt existing tools, without explicit modification, imaging protocols different those tools were trained tuned. An open-source implementation provided, can be applied wide range problems. The evaluated four brain MRI methods: hippocampus using FreeSurfer (Fischl et al., 2002); multi-atlas label fusion (Artaechevarria 2009); extraction BET (Smith, tissue FAST (Zhang 2001). generates 72%, 14%, 29% 21% fewer erroneously segmented voxels than respective methods. In experiment method, average Dice overlap between reference 0.908 normal controls 0.893 patients mild cognitive impairment. Average overlaps 0.964, 0.905 0.951 obtained extraction, white matter gray segmentation, respectively.