WE‐FG‐206‐03: An Adaptive Model for Pharmacokinetic Nested Model Selection in Dynamic Contrast Enhanced MRI Data Analysis

摘要

Purpose: This study introduces an adaptive Model Selection (MS) technique to perform Pharmacokinetic nested MS from the time trace of longitudinal relaxation rate change, ΔR1 (R1 = 1/T1) in Dynamic Contrast Enhanced (DCE) MRI studies. Methods: Three physiologically nested models derived from the standard Tofts model along with an averaged (over 30 patients) arterial input function were used to simulate a set of ΔR1 profiles to describe possible physiological conditions of underlying tissue pathology: Model‐1: the vascular compartment is filled with contrast agent (CA) with no outward leakage. Model‐2: the vascular compartment is filled with CA with outward leakage but no evidence of back‐flux. Model‐3: the vascular compartment is filled with CA with both outward and backward‐flux. Three different sets of simulated ΔR1 profiles in presence of different signal‐to‐noise ratios (5, 10, 15, 30, 70, 100, and no noise) were used to train an Artificial Neural Network (ANN) for performing MS. A k‐fold cross‐validation method was used to validate and optimize the ANN architecture. The trained‐ANN was also applied on the DCE‐MRI data of 20 patients with Glioblastoma and results were compared to the models selected by the Log‐Likelihood‐Ratio (LLR) technique using Dice coefficient. Results: The confusion matrix and the strong similarity (Dice coefficients of 0.87, 0.89 for Models 2 and 3) between the models selected by the trained ANN and the LLR method confirms that the performance of the adaptive NMS technique is superior to the LLR method. The ANN showed a strong sensitivity for selecting models with higher orders; thus less type‐II errors (never misses any tissues with leaky vasculature (Models 2 and 3). Conclusion: The noise insensitivity, speed, and superiority of the ANN technique in choosing the best PK model would allow a less biased estimation of cerebrovascular permeability parameters in tumorous tissues. This work is supported in part by HFHS mentored Grant (A10237).