3D segmentation in MRI of brain tumors: preliminary results

Tissue segmentation based on 2D and 3D feature map derived from high resolution MR images was performed in phantoms, normal humans and particularly those with brain tumors (four benign and two malignant). Three inputs: proton density, T2- and, as a third, T1-weighted MRI, were utilized. Statistical and anisotropic diffusion filters were applied to the data and k-Nearest Neighborhood segmentation algorithm was utilized. The inclusion of T1 based images into segmentation produced dramatic improvement in tissue identification. The authors' technique utilizing all three inputs provided better segmentation (p<0.001) than that based on any combination of two inputs. In benign brain tumors, the authors identified tumor volume prior to the injection of gadolinium-DTPA. In malignant tumors, up to four abnormal tissues were identified: (1) solid tumor core, (2) cyst, (3) edema in white matter and (4) edema in grey matter. Subsequent neurosurgery confirmed the authors' model. These results encourage further investigation.