Novel segmentation of dynamic 18F-FDG PET bypasses the need of arterial plasma input function, delivering a robust quantification of the tumor microenvironment
Prateek Kaiyar, Werner Siemens Imaging Center, Eberhard Karls University Tübingen
Introduction: Non-invasive imaging has greatly enriched our understanding of the factors underlying cancer progression. 18F-FDG PET plays a pivotal role in tumor characterization and treatment planning in many cancer subtypes. Although compartmental modeling extracts the crucial information about blood flow and receptor status, estimation of kinetic parameters not only relies upon the acquisition of time activity curves (TACs) with low noise, but also on a precise measurement of the arterial input function (AIF). Accurate measurement of the AIF is an invasive and tedious procedure in humans and highly challenging in mice, nearly prohibiting applications in longitudinal studies. We propose a new algorithm for accurate mapping of the tumor micro-environment without the necessity of an AIF. We show that our method is robust to varying levels of noise, and, thus, can be used for the voxel-wise characterization of dynamic PET data. We also present exhaustive simulations to compare the proposed algorithm with standard compartmental modeling, and assess the kinetic parameter variability caused by various distortions in the AIF.
Methods: Several (n=12, 4 mice x 3 scans) 18F-FDG dynamic PET scans were acquired using the Inveon small-animal PET scanner for 60 min with 27 frames. The plasma input curves of all the measurements were approximated using a minimal blood-sampling scheme [1]. A two tissue compartmental model was applied to the mean TAC of the entire tumor for each measurement. The observed kinetic parameter range was used to simulate the TACs of three tumor tissue classes. All the tumors were excised into 2-3 mm thin slices parallel to the axial field of view and processed for histological staining. The TACs of the simulated and real example were segmented using a novel clustering algorithm.
Results: To perform an objective evaluation, simulated TACs were corrupted with different levels of noise and segmented using the suggested algorithm. The proposed method showed substantial prediction accuracy in comparison to SUV based clustering. Spectral clustering derived segmented regions of the real example showed a significant correlation with histology slides stained with CD31. Cluster-wise averaged time activity curves were also in congruence with histology, with well perfused areas characterized by a higher activity concentration in contrast to that of remaining tumor regions. Kinetic modeling rate constants were underestimated with decrease and overestimated with increase in the peak amplitude of the plasma input function. The rate constants also showed substantial variability with the increase in amount of noise.
Conclusion: Unlike compartmental modeling, spectral clustering is independent of the AIF, thus eliminates the need of arterial blood sampling and associated sampling errors. The variability in the estimation of kinetic rate constants indicates its sensitivity for voxel level analysis and stresses the need of robust models for precise quantification of the tumor microenvironment.
[1] Estimation of the 18F-FDG input function in mice by use of dynamic small-animal PET and minimal blood sample data, JNM 2007