An artificial neural network for GFR estimation in the DCE-MRI studies of the kidneys

The dynamic contrast-enhanced magnetic resonance imaging is a diagnostic method directed at estimation of renal performance. Analysis of the image intensity time-courses in the renal cortex and parenchyma enables quantification of the kidney filtration characteristics. A standard approach used for that purpose involves fitting a pharmacokinetic model to image data and optimizing a set of model parameters. It is essentially a multi-objective and non-linear optimization problem. Standard methods applied in such scenarios include nonlinear least-squares (NLS) algorithms, such as Levenberg-Marquardt or Trust Region Reflective methods. The major disadvantage of these classical approaches is the requirement for determining the starting point of the optimization, whose final result is a local minimum of the objective function. On the contrary, artificial neural networks (ANN) are trained based on a large range of parameter combinations, potentially covering whole solution space. Thus, they appear particularly useful in fitting complex, non-linear, multi-parametric relationships to the observed noisy data and offer greater ability to detect all possible interactions between predictor variables without the need for explicit statistical formulation. In this paper we compare the ANN and NLS approaches in application to measuring perfusion based on DCE-MR images. The experiments performed on a dataset containing 10 dynamic image series collected for 5 healthy volunteers proved superior performance of the neural networks over classical methods in terms of quantifying true perfusion parameters, robustness to noise and varying imaging conditions.