TU-C-12A-04: Quantification of Hepatic Perfusion Maps From DCE MRI Using Spatial Regularization

Authors


Abstract

Purpose:

To reduce uncertainty in hepatic perfusion quantification from DCE-MRI for assessing tumor and normal tissue response to RT in intrahepatic cancer.

Methods:

Pharmacokinetic modeling of liver DCE-MRI estimates arterial inflow rate (ka), portal venous inflow rate (kp), central venous outflow rate (kl) and bolus arrival time delays (ta, tp) from respective arterial and portal vein input functions to tissue. The parametric images estimated by the conventional approach could exhibit large variance and severe bias. To reduce the uncertainty in the quantitative images, we incorporated the prior knowledge, the slowly spatial variation of the time delays in the liver, into the model by adding spatial regulations on the time delays. The weighting factors of regularizations were determined by Stein’s unbiased risk estimate (SURE) of the mean squared prediction error (MSPE). The regularized model was optimized by the conjugate gradient minimization and implemented on the GPU to achieve fast computation. The method was evaluated by simulated and patient DCE-MRI data. The simulated perfusion image was generated with (k2, ka, kp) of (400, 20,100 ml/(100g min)), smooth (ta, tp) maps in a 64×64 matrix, and Gaussian random noise to have contrast-to-noise ratios (CNR) of 10 to 100.

Results:

Compared to the conventional estimation, the proposed method reduced estimate variances from 14% to 8% for k2, 38% to 9% for ka, and 14% to 10% for kp in the simulated data with 20 CNR. The SURE-based method consistently led to the optimal weight of the regularization determined by the true MSPE on the simulation data. Regularizing time delays improved the hepatic perfusion images of the patient data.

Conclusion:

The improved repeatability and possible accuracy in hepatic perfusion images quantified from DCE-MRI have the potential to increase statistical power in clinical application of the physiological images for predicting outcome of liver cancer treatment.

The work was supported by NIH/NCI grant RO1CA132834.

Ancillary