Quantitative MR Estimation of Interstitial Fluid Pressure in a Pre-Clinical Rat Model of Glioblastoma Tumors Using an Adaptive Model and Principle Component Analysis
Molaeimanesh Z, A NVD, Faghihi R, Elmghirbi R, Ewing J, Wen N, I JC, and Bagher-Ebadian H. Quantitative MR estimation of interstitial fluid pressure in a pre-clinical rat model of glioblastoma tumors using an adaptive model and principle component analysis. Med Phys 2017; 44(6):3018-3019.
Purpose: In many solid tumors, an elevated Interstitial Fluid Pressure (IFP) results in inefficient uptake of therapeutic agents. This study investigates the feasibility of constructing an Artificial Neural Network (ANN) for prediction of IFP in a rat-based model of glioblastoma (U251n) tumors. Methods: Dynamic-Contrast-Enhanced (DCE)-MRI data was studied in 35 athymic rats with orthotopic-U251-glioma. Tumor IFP was measured by the Wick-in-needle technique. Using a Model-Selection (MS) technique applied on Toftsmodel, three physiologically-nested models with the following PK parameters were constructed for the tumors: Model-1 with one parameter (vp: plasma volume), Model-2 with two parameters (vp and Ktrans: forward transfer constant), and Model-3 with three parameters (vp and Ktrans and kep: reverse transfer constant). A principle component analysis (PCA) technique was recruited to identify and construct three major components with the most discriminative power from PK parameters. The PCA features were used to train an ANN. A K-fold-Cross-Validation technique and an Area-Under-Correct- Classification-Fraction (AUCCF) were employed for training, structure optimization, and evaluation of the ANN. Results: Three discriminant features were identified and constructed by the PCA as follows: X1 = 0.95 × vp (Model-1) + 0.869vp (Model-3) + 0.76 × Ktrans (Model-2) + 0.68×Ktrans (Model-3) + 0.94×ve (Model-3), X2 = 0.69×Ktrans (Model-2) + 0.99×kep (Model-3), and X3 = 0.98×vp (Model-1). The performance of the trained-ANN at the optimum epoch was 79%. The cumulative variance of the three PCA-components was 89%. The optimal architecture (3, 6 and 1 neuron in its input, hidden and output layers) of the ANN was found at the epoch number of 1200 with AUCCF of 0.79. The generalization error of the trained- ANN for estimating the IFP was 20%. Conclusion: This pilot study confirmed the feasibility of constructing an adaptive model (ANN-PCA) for prediction of IFP of U251n tumors using DCE-MRI and MS technique. The proposed methodology can be used for predicting the IFP of glioblastoma in human.