NIMG-22MORPHOLOGICAL RECONSTRUCTION BASED AUTOMATED SEGMENTATION OF GLIOBLASTOMA TUMOR VOLUME ON MULTI-MODAL MRI IMAGES

Abstract

BACKGROUND: Glioblastoma is the most aggressive brain tumor and, magnetic resonance imaging (MRI) is the most common imaging tool for brain tumor diagnosis and disease status monitoring post any intervention including craniotomy, irradiation and various chemotherapies. Tumor volume estimation using MRI is routinely used to determine responses, such as Macdonald's or RANO's criteria. Current techniques used for tumor volume measurements are manual measurements of the largest two-dimensional tumor area and they are inaccurate and operator-dependent. Automated techniques for estimating volumetric measurements are needed for minimizing human bias and to follow disease status objectively. We developed and implement automated segmentation of tumors by fusing information from multi-modal MRI. METHODS: High resolution three-dimensional (1 mm3 isotropic voxel) T2-weighted (T2), fluid attenuated inversion recovery (FLAIR), and T1-weighted (T1) pre- and post-contrast images were acquired on nine glioblastoma patients. The images were pre-processed for co-registration, skull-stripping, intensity non-uniformities correction, noise reduction, and intensity standardization. Regional maxima were obtained on T2/FLAIR images using grayscale morphological reconstruction followed by adaptive thresholding, and regional multi-modal density distribution, connectivity, and growing techniques. Enhanced region of the tumor was obtained from the T1 pre-/post-contrast images followed by the minimization of false classifications arising from enhancing vasculature. Classified regions from T2/FLAIR, and T1 pre-/post-contrast images were combined to obtain enhancing tumor component and the surrounding FLAIR changes representing vasogenic edema/tumor infiltration as the final segmentation for volumetric measurements. RESULTS: The automatic program is able to capture and calculate tumor volumes from T1 post contrast or T2/FLAIR images accurately, as judged by neuroradiologists, neurosurgeons and neuro-oncologists. Additionally, our technique performed consistently well across all the patients that we scanned without any human intervention. CONCLUSION: We developed and implemented automated technique that identifies tumor volumes precisely, improves the accuracy and efficiency of tumor status measurements significantly with minimal human bias.