Three-Point Deep Learning Framework for Protocol-Independent and AIF-Free DCE-MRI Parameter Estimation in Gliomas

Motivation: Quantitative assessment of brain tumor by dynamic contrast enhanced MRI(DCE-MRI) using non-linear least squares(NLLS) method is slow, scan length dependent and estimate noisy maps. Deep-learning(DL) approach shows promise, inheriting temporal sampling or AIF dependencies which are acquisition protocol dependent. Goal(s): Develop and compare convolutional neural network(CNN) and U-Net architecture for estimating perfusion parameters(GTKM) using three strategic time points and compare with NLLS. Approach: Networks map three concentration-time points(bolus arrival, peak, and tail concentrations to tracer kinetic(TK) parameters . Networks performance was evaluated against NLLS on in-vivo data. Results: Our approach demonstrated high accuracy and substantial reduction in computation time for TK parameter estimation. Impact: The proposed DL approach enables robust DCE-MRI quantification in gliomas using minimal temporal sampling, eliminating AIF dependencies while maintaining accuracy in substantially less time. Facilitating multi-center clinical adoption and efficient pre-operative tumor characterization and treatment monitoring.