Determination of unscaled blood input for human dynamic FDG brain PET

Abstract

Objectives: Many existing techniques for the non-invasive quantification of the blood input function in dynamic FDG-PET imaging require strong historical information or user input. The technique proposed in this work utilizes the assumption that a dynamic PET scan can be modeled by the Patlak plot to determine an unscaled blood input function. Materials and Methods: The time activity curve (TAC) for each voxel in a dynamic image can be considered as an n-dimensional vector. In this context, a TAC follows the Patlak plot if and only if the TAC is a linear combination of the blood input function and the integral of the blood input function. Given a set of TACs which follow the Patlak plot, we can thus use PCA to determine a basis which spans the same vector space as the blood input function and the integral of the blood input function. We then seek to find two TACs in this vector space which best satisfy that the estimated anti-derivative of one of the TACs is close to the other TAC; such TACs are candidates for the blood input function and the integral of the blood input function. We were able to construct a low (2) dimensional optimization problem to find such TACs. Results: We applied our results to obtain predicted blood input functions and Ki maps for twelve normal subjects. Scaling the predicted blood input function to best match the ground truth, we achieved an average SSE of $0.042 \pm 0.032$ and an average DTW distance of $0.141 \pm 0.053$. Matching the means of the predicted and ground truth Ki maps, we achieved an average MAPE of $2.539 \pm 0.928$ and an average SSIM of $0.991 \pm 0.005$. Conclusion: While not often viewed as such, the assumption that some dynamic data follows a kinetic model gives strong prior information. In the case of the Patlak plot, we can use this assumption to estimate an unscaled blood input function and unscaled Ki map.