Neuroradiologists, medical physicists and biomedical scientists who are interested in pharmacokinetic modeling and brain tumor studies.In cancerous tumors, systemic drug delivery of macromolecular therapeutic agents is challenging due to the lack of lymphatic drainage, chaotic microvasculature and heterogeneity of the tumor tissue which all together leads to an increased central Interstitial Fluid Pressure (IFP), and an inadequate and uneven uptake of the drug^{1, 2}. An increased central IFP negatively affects the tracer distribution volume within the whole tumor². Thus, non-invasive prediction of tumor IFP would be the special of interest in diagnosis and treatment phases and can be applied toward patient-specific treatments and to more accurately investigate effects of flow patterns on heterogeneous tumor in drug delivery and quantification of response to treatment². Our group has already shown that three physiologically Nested Model (NM) selection^3,4 can be driven from the standard Tofts model. This pilot study introduces Model Evolution (ME) concept as a novel approach that is framed and formulated based on the Model Selection (MS) technique^3,4 in Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) studies. To this end, the ME technique is applied on DCE-MRI data of 19 U251n rat tumor models of cerebral tumors with known in-vivo measures of their tumor IFPs to train an adaptive model (AM) for predicting tumor IFP.In this study, using NM selection technique, ME concept is framed to discover and quantify the evolutions of three different model regions corresponding to three different physiologically NMs throughout the course of DCE-MR experiment. We believe that in leaky tumors, the temporal formation and transformation of Model 1, 2 and 3 to each other highly depend on the inward diffusion and outward convection of Contrast Agent (CA) concentration which are the main components affecting the IFP of tumor. We hypothesized that three evolutionary profiles in the course of DCE-MRI experiment generated from the ME concept, that are linked to the transformation of the pharmacokinetic conditions of underlying tumor tissue pathology contains abundant information relevant to the tumor IFP and can provide required information for training adaptive models for characterizing the tumor IFP.To test the hypothesis, 19 Nude rats with U251n rat tumor models of cerebral tumor were studied using Look-Locker T1 mapping and DCE-MRI experiments (Dual Gradient Echo, DGE, 150 image sets at 4.0 sec intervals over 10 min: matrix = 128x64, five 2.0 mm slices, NE= 2, NA=1, TE/TE/TR = 2.0/4.0/40ms with bolus intravenous injection of the Magnevist at 0.25 mmol/kg) acquired at 7T field strengths. Using the first and second echoes of the DCE study and maps of T1 before and after the DCE-MRI study, a trace of ΔR1(t) was constructed across the 9 min duration of the DCE-MR experiment³. In each animal, in-vivo measurement of tumor IFP (Figure 2-A) was done right after the DCE-MRI experiment using a wick–in–needle technique⁵. The three ME profiles along with the measure of tumor IFP were used to construct and train an AM for predicting the tumor IFP. An Artificial Neural Network (ANN)⁶ with Multi-Layer-perceptron architecture was recruited as the adaptive predictor. Pharmacokinetic parameters of the tumors and their corresponding model choice maps were estimated from DCE-MRI data using Maximum-Likelihood Estimation technique and the Log-Likelihood- Ratio test respectively. Using the ME concept, three model profiles were generated and the second- and third-order moments (variance and skewness, see Figure 2-B) of the three profiles were used to construct a set of descriptive features for training the ANN. K-Fold Cross-Validation with Leave-One-Out strategy⁶ along with the Area-Under-the-Reciver-Operating-Characteristic (AUROC) curve were used for training, optimization, and validation of the ANN. The in-vivo measure of tumor IFPs (1.70, 2.30, 2.50, 3.50, 4.16, 5.55, 5.70, 5.86, 10.00, 10.20, 10.36, 10.89, 11.88, 12.80, 12.90, 15.15, 16.98, 19.20 mmHg) served as the ground truth for training and validation of the ANN.Figure 1-A and 1-B show the volume evolution of 3 different nested models (ME(1), ME(2), and ME(3)) in a U251n rat-tumor model of cerebral tumor during the course of MR experiment. These figures clearly confirm the three models being nested and also illustrates the latency level between different models as well as the behavioral changes of each model during the course of MR experiment (~ 8.5 min). Results imply that the trained AM (optimal ANN, Architecture:7:3:1) can estimate the IFP of the embedded tumors from ME profiles that are generated from DCE-MR data with predictive power of 87% (AUROC=0.87). This pilot study confirms that the ME concept can make a paradigm shift in noninvasive estimation of tumor IFP from DCE-MRI studies.