Fluid mechanics based quantitative transport mapping network for predicting gene expression of nasopharyngeal carcinoma (NPC) patients

Motivation: To use fluid-mechanics based deep learning method to predict perfusion parameters from dynamic imagesGoal(s): We propose to explore the possibility to use neural network trained on simulated data from fluid mechanics simulation to analyze dynamic medical images. Approach: We use quantitative transport mapping network (QTMnet), which is trained on simulated concentration propagation profile generated from constrained constructive optimization (CCO) and transport equation-based tracer propagation simulation, to predict perfusion parameters including flow rate, permeability, vasculature volume, from DCE MRI images. Results: QTMnet predict perfusion parameters accurately in simulation study and can distinguish different gene expression group patients comparing with using traditional kinetics model. Impact: Proposed QTMnet method can be used to predict different perfusion parameters related to dynamic images accurately and automatically without usage of arterial input function.