Knowledge-based tumor segmentation in MR images

Conventional measurements of tumor volume are prone to ambiguity and error, so a method which semi-automatically segments and labels glioma tumors in magnetic resonance (MR) images of the human brain is presented. With Chamfer matching, brain MR images, roughly segmented by threshold and morphological operations, are matched with a fuzzy brain atlas to get a fuzzy segmentation of brain structures. After that, a supervised histogram estimation method will perform another fuzzy segmentation of brain structures on the basis of the gray feature information of MR images. The above two fuzzy segmentations are fused to make a decision Finally a boundary optimization is applied with deformable contour. Results include segmented areas of cerebrospinal fluid, gray matter, white matter, and brain tumor (glioma). The tumor segmentation was compared with "ground-truth" tumor segmentations that were generated by radiologist hand labeling.