Tracking and Modeling Tumor Growth Rates
in Patients with Malignant Gliomas

Annual Meeting of the American Academy of Neurology, 2000 [submitted]

Surface modeling approaches help to capture the complex morphology of brain tumors. They can be used in conjunction with tissue classification approaches to assess growth rates and therapeutic response, and are sufficently detailed to track subtle changes in the dynamics of tumor growth.

Objective. As part of a comprehensive study of patients with malignant gliomas, two image analysis algorithms were applied to serial MRI scans. Rates of change were tracked for gadolinium-enhancing tissue, peritumoral edema and cystic compartments. The purpose of this study was to investigate the accuracy of two 3-dimensional image analysis algorithms in tracking tissue changes by comparing the results to a manually segmented gold standard.

Background. Methods currently employed to estimate tumor response to therapy are not reliable. A method capable of accurately and systematically tracking changes in tumors would be valuable in patient management and in therapy evaluations, providing a quantitative index of comparison for multicenter trials. 3D structural maps also provide a stereotaxic framework for the correlation of data from other modalities. Tumor changes were tracked with a reliable, objective tissue classification algorithm and with a surface modeling algorithm which has successfully been used to track minute changes in the corpus callosum in children.

Design/Methods. 3D T2-weighted and gadolinium-enhanced T1-weighted (256x256 resolution, 3 mm spacing) SPGR MRI volumes were acquired over a 3.5 year period from 10 patients with histopathologically confirmed glioblastoma multiforme (GBM; age 4-54 yrs at first scan). Scans were automatically aligned into Talairach stereotaxic space and one or both of the following protocols were applied: (1) Tissue Classification; tissue maps were created by determining a Gaussian mixture distribution reflecting the image intensities of specific tissue classes in the scan series. A nearest neighbor algorithm was used to differentiate tissue types and its accuracy confirmed by tagging points. Tissue maps were manually adjusted to better delineate class boundaries. (2) Surface Modeling; tumor enhancement volumes were traced and a 3D surface reconstruction algorithm based on tiled, parametric mesh modeling was applied. Results from the nearest neighbor and surface modeling algorithms were compared with manual segmentations.

Results. T1-based segmentation maps, generated by the nearest neighbor algorithm, provided excellent tissue differentiation and growth rate evaluation for tumor enhancement (r-squared=0.99). Volumetric changes in tumor components ranged from -73% to +2600%, corresponding to a halving time of 130 days and a doubling time of 13.8 days. Volumetric measurements of edema were systematically 16.5%+/-8.8% lower in T1-based maps (average volume 70.5+/-55.9cm³) than T2-based maps (average volume 82.5+/-63.3 cm³; p<0.001, paired t-test) but values were highly correlated as were volumes from surface modeling and the manually derived gold standard (r-squared=0.94).

Conclusions. Tissue maps, based on the nearest neighbor algorithm, provide significant power in tracking the regional dynamics of multiple tissue classes in GBM patients. The resulting tissue-classification and growth rate detection systems also provide a 3D structural framework to correlate spectroscopic, diffusion and histopathologic indices in interventional studies and quantitatively evaluate therapeutic response in individual patients or clinically-defined groups.

Grant Support: (to P.T. and A.W.T.): NIMH/NIDA (P20 MH/DA52176), P41 NCRR (RR13642); (A.W.T.): NLM (LM/MH05639), NSF (BIR 93-22434), NCRR (RR05956) and NINCDS/NIMH (NS38753).

Related Publications