Dynamic contrast enhancement (DCE) MRI has been used to study tumor microvasculature in terms of parametric measures such as blood flow and rate of leakage [1]. In practice, to maintain high temporal resolution and derive relevant parameters accurately (e.g. wash-in rate), SNR of DCE-MRI acquisition is compromised and can be worsened by fast imaging artifacts (e.g. streaking due to under-sampling) [2]. Consequently, pharmaco-kinetic (PK) model [1] fit to DCE data is poor and results in “pixelated” maps even within a homogenous tissue. Moreover, if such DCE temporal data is used for subtraction from baseline, it can obscure a lesion and reduce clinical confidence in the results. Previously, an ICA based source separation method was investigated for removal of under-sampling artifacts [2]. One problem with generic ICA based methods is that they can result in negative source component(s), which can be difficult to interpret physiologically and more importantly result in loss of useful information from remaining components. Since DCE images are typically fit to PK model and shape correlated to tumor malignancy [8], it is essential that all sources are non-negative and repetitive to ensure physiological plausibility while maintaining the shape of DCE curves. One source separation method which can mathematically satisfy these conditions is Convex Analysis of Mixtures of Nonnegative Sources (CAMNS) [3,4]. In this work we demonstrate the utility of CAMNS and subsequent weight estimation methodology to filter “noise” from DCE concentration data and its impact on the PK model parameters in liver DCE-MRI.MRI Data: MRI data for this study was obtained on 1.5T GE SIGNA EXCITE and GE Signa HDxt system (GEHC) for patients with liver fibrosis (N =2) and tumor (N =1). The liver-DCE protocol was: 3D SPGR sequence, TE/TR = 1.12/4.8 to 1.3/4.5 ms , FA = 15-30°, FOV = 400 to 460 mm2, 32 to 82 bolus volumes. DCE signal data was converted into concentration units using the baseline images and fixed tissue T₁ (1.5T = 550 ms, 3T = 800ms) and used for CAMNS based filtering. Physiologically implausible voxels were removed. CAMNS based filtering: CAMNS algorithm was used to identify underlying sources of DCE-MRI concentration data. Three sources were assumed: vascular, leakage and noise [5]. Once sources are obtained, a constrained optimization problem is solved to calculate weights corresponding to the sources as follows: If data is $$$ X \varepsilon \Re^ {m x n} $$$ , (m = bolus phases, n = no. of voxels) and p-sources $$$ s \varepsilon \Re^{m x p} $$$, weights w calculated by solving $$$ \min_{w} \left \| X-ws \right \| $$$ such that w $$$ \geqslant 0 $$$. Components corresponding to j^th source are obtained by multiplying j^th row of w with j^th column of s. Thus, we reconstruct 4D-data as: Data_{4D, filtered} = $$$ \sum_{\forall i }^{ } component_i $$$ where is the component_i corresponding to the i^th source. It should be noted that weight estimation is done over the entire FOV, including those voxels which were removed because of being implausible or extremely noisy from DCE standpoint. PK Fitting: The original 4D concentration data and post CAMNS filtered concentration data were both fit using a dual input (aorta and portal vein), single compartment Materne model [6]. Coefficient of determination (R²) metric was computed at each voxel to assess goodness-of-fit before and after CAMNS based filtering. The methodology was implemented using functionality in ITK [7].Figure 1 shows a representative liver tumor case. Notice that extremely noisy voxels (Figure 1B) have some DCE relevant shape, but nevertheless cannot be filtered using direct filtering methods. Sources from CAMNS (Figure 1C) are used to estimate source weights (Figure 1D). Figure 2 demonstrates efficacy of CAMNS approach in filtering DCE curves with different wash-in and wash-out characteristics [8]. Notice that important characteristics such as bolus arrival time, initial upslope and wash-out characteristics are well preserved by CAMNS based filtering. We did notice that in voxels which represent AIF shape (sharp peak), there was a tendency to smoothen the peak (Figure 2c is good example). Further investigation indicated that number of AIF–like voxels was very less in data provided to CAMNS (over-aggressive pre-filtering was responsible). Figure 3 demonstrates that R² metric for PK fit on CAMNs filtered data (Figure 3D) is much higher (0.99) as compared to R² obtained with original concentration data (~0.7) (Figure 3C). Figure 3B (arrow) also demonstrates how some of missing K_trans signatures are highlighted post filtering.CAMNS based filtering improves accuracy of DCE-MRI quantification and will enhance confidence of clinicians in DCE data and PK maps.