Repeatability of tumor perfusion kinetics from dynamic contrast-enhanced MRI in glioblastoma

Abstract

Background Dynamic contrast-enhanced MRI (DCE-MRI) parameters have been shown to be biomarkers for treatment response in glioblastoma (GBM). However, variations in analysis and measurement methodology complicate determination of biological changes measured via DCE. The aim of this study is to quantify DCE-MRI variations attributable to analysis methodology and image quality in GBM patients. Methods The Extended Tofts model (eTM) and Leaky Tracer Kinetic Model (LTKM), with manually and automatically segmented vascular input functions (VIFs), were used to calculate perfusion kinetic parameters from 29 GBM patients with double-baseline DCE-MRI data. DCE-MRI images were acquired 2–5 days apart with no change in treatment. Repeatability of kinetic parameters was quantified with Bland–Altman and percent repeatability coefficient (%RC) analysis. Results The perfusion parameter with the least RC was the plasma volume fraction (vp), with a %RC of 53%. The extra-cellular extra-vascular volume fraction (ve) %RC was 82% and 81%, for extended Tofts-Kety Model (eTM) and LTKM respectively. The %RC of the volume transfer rate constant (Ktrans) was 72% for the eTM, and 82% for the LTKM, respectively. Using an automatic VIF resulted in smaller %RCs for all model parameters, as compared to manual VIF. Conclusions As much as 72% change in Ktrans (eTM, autoVIF) can be attributable to non-biological changes in the 2–5 days between double-baseline imaging. Poor Ktrans repeatability may result from inferior temporal resolution and short image acquisition time. This variation suggests DCE-MRI repeatability studies should be performed institutionally, using an automatic VIF method and following quantitative imaging biomarkers alliance guidelines.

Structural and practical identifiability of contrast transport models for DCE-MRI

“…Our findings for both the LTK model (discussed in Section 3.2) and the nested PM, TK, and eTK models (discussed in Section S.1.2) reveal how practical parameter and, thus, model identifiability is a↵ected by the quality of the data: the more noisy the data and/or the individual-based VIF function, the lower the confidence levels of identifiability of the di↵erent parameters. This is evident from the change in the shape of the profile likelihood transforming from a parabola-like shape in the (AA) case to the flat shape in the (RR) case, shown in Figures 3, 4 The analysis proposed here is of significant importance considering the wide use of DCE-MRI data in research [52,7] and, thus, the need for ensuring reliability and reproducibility of transport model results [53]. DCE-MRI has been shown to be associated with tumor angiogenesis and may be used to assess glioma grading [54,55,56,57,58,59], predict genetic mutation status of brain tumors [60,61,62], distinguish pseudoprogression from true progression in glioblastomas [63,64], and predict response to antiangiogenic treatment [65].…”

“…The analysis proposed here is of significant importance considering the wide use of DCE-MRI data in research [52, 7] and, thus, the need for ensuring reliability and reproducibility of transport model results [53]. DCE-MRI has been shown to be associated with tumor angiogenesis and may be used to assess glioma grading [54, 55, 56, 57, 58, 59], predict genetic mutation status of brain tumors [60, 61, 62], distinguish pseudoprogression from true progression in glioblastomas [63, 64], and predict response to antiangiogenic treatment [65].…”

“…However, variability in DCE acquisition (e.g. in scan duration) and imaging analysis with various transport models may produce diverging results, which makes repeatability a challenging issue and hinders its wider adaptation for both clinical practice and research [53,68]. Prospective longitudinal clinical trials, preferably in a multicenter setting with standardized imaging techniques and quantification of DCE parameters, are needed to validate the utility of transport parameters as imaging biomarkers for treatment response and survival prediction.…”

Model discovery approach enables non-invasive measurement of intra-tumoral fluid transport in dynamic MRI

“…Here, we develop a strategy that overcomes these limitations by integrating DCE-MRI with a method developed for the discovery of equations governing timeseries data through the concept of function-space regression, known in data sciences as sparse identification of nonlinear dynamics (SINDy) (28). Given that the time-and space-resolved data obtained by DCE-MRI is often noisy (29), we modified weak SINDy, a method that leverages the weak form of governing equations to provide an efficient and accurate method for simultaneous model discovery and parameter estimation from noisy data (30). This method bypasses the use of discrete approximations of derivatives on noisy or sparse data, which are used by SINDy (28), or iterative gradient descent methods used in modern DCE-MRI model inversion which require estimating derivatives of the objective function with respect to the data (21)(22)(23).…”

Model discovery approach enables noninvasive measurement of intra-tumoral fluid transport in dynamic MRI

“…Here, we develop a strategy that overcomes these limitations by integrating DCE-MRI with a method developed for the discovery of equations governing time-series data through the concept of function-space regression, known in data sciences as sparse identification of nonlinear dynamics (SINDy). 31 Given that the time- and space-resolved data obtained by DCE-MRI is often noisy, 32 we utilized a variation of SINDy which leverages the weak form of governing equations to provide an efficient and accurate method for parameter estimation from noisy data. 33,34 Weak-form methods bypass the use of discrete approximations of derivatives on noisy or sparse data, which are used by the original SINDy implementation 31 or gradient descent methods used in Jacobian estimation for standard ODE-fitting techniques.…”