Ensemble of expert deep neural networks for spatio‐temporal denoising of contrast‐enhanced MRI sequences

HighlightsWe present a novel spatio‐temporal denoising framework, based on Deep Neural Networks (DNNs), for the analysis of DCE‐MRI sequences of the brain.Specifically, DCE‐MRI denoising facilitates the extraction of the pharmacokinetic (PK) parameters from the DCE‐MRI concentration curves (CTCs) and thus allows a quantitative assessment of blood‐brain barrier functionality.This is accomplished by an ensemble of expert DNNs (deep‐autoencoders), where each is trained on a specific subset of the input space to accommodate different noise characteristics and curve prototypes.We introduce a novel sampling scheme, for generating realistic training sets that faithfully model the DCE‐MRI data and accounts for statistical similarity of neighboring CTCs.Experimental results on synthesized realistic data demonstrate that PK parameters reconstruction is more accurate for denoised CTCs using our method in comparison to CTCs denoised by other methods.The proposed approach is also applied to real DCE‐MRI scans from patients with stroke and brain tumors and is shown to favorably compare to state‐of‐the‐art denoising methods. Graphical abstract Figure. No caption available. ABSTRACT Dynamic contrast‐enhanced MRI (DCE‐MRI) is an imaging protocol where MRI scans are acquired repetitively throughout the injection of a contrast agent. The analysis of dynamic scans is widely used for the detection and quantification of blood‐brain barrier (BBB) permeability. Extraction of the pharmacokinetic (PK) parameters from the DCE‐MRI concentration curves allows quantitative assessment of the integrity of the BBB functionality. However, curve fitting required for the analysis of DCE‐MRI data is error‐prone as the dynamic scans are subject to non‐white, spatially‐dependent and anisotropic noise. We present a novel spatio‐temporal framework based on Deep Neural Networks (DNNs) to address the DCE‐MRI denoising challenges. This is accomplished by an ensemble of expert DNNs constructed as deep autoencoders, where each is trained on a specific subset of the input space to accommodate different noise characteristics and curve prototypes. Spatial dependencies of the PK dynamics are captured by incorporating the curves of neighboring voxels in the entire process. The most likely reconstructed curves are then chosen using a classifier DNN followed by a quadratic programming optimization. As clean signals (ground‐truth) for training are not available, a fully automatic model for generating realistic training sets with complex nonlinear dynamics is introduced. The proposed approach has been successfully applied to full and even temporally down‐sampled DCE‐MRI sequences, from two different databases, of stroke and brain tumor patients, and is shown to favorably compare to state‐of‐the‐art denoising methods.