Interactive segmentation of cerebral gray matter, white matter, and CSF: photographic and MR images

Digital photography of postmortem brain slices was compared with magnetic resonance imaging (MRI) for morphological analysis of human brain atrophy. In this study, we used two human brains obtained at autopsy: a cognitively defined nondemented control (70-yr-old male) and a demented Alzheimer's disease (AD) subject (82-yr-old female). For each of two brains, interactive manual image segmentation was performed by two observers on two image sets: (a) four coronal T1-weighted MR images (5 mm slices); and (b) four digitized photographic images from comparable rostrocaudal levels. Microcomputer image analysis software was used to measure the areas of three segmented cerebral compartments--gray matter (GM), white matter (WM) and CSF--for both image types. Resegmentation error was defined as the absolute difference between the areas derived from two segmentation trials divided by the value from trial 1 and multiplied by 100. This yielded the percent difference between the area measurements from the two trials. We found intra-observer agreement was better (error rates 1-18%) than inter-observer agreement (3-70%) with best agreement for WM and least for CSF, the smallest object class. MRI overestimated GM area relative to digitized photographs in the control but not the AD brain. The results define limitations of manual image segmentations and comparison of MRI with pathologic section photographic images.