Network structure impacts on quantitative transport mapping network for compartment model inversion in perfusion image quantification

Motivation: The accuracy of deep learning methods for multi-compartment perfusion modeling may be affected by the choice of network structure. Goal(s): To compare various deep learning network structures on the accuracy of perfusion parameter estimation using a previously proposed deep learning method QTMnet. Approach: Four different architectures were trained on synthetically generated perfusion data and tested on nasopharyngeal carcinoma DCE MRI images. Results: Among the architectures studied, CMUNeXt provided the highest accuracy in simulation while Attention U-net provided the best contrast between tumor and normal tissue in the DCE MRI data. Impact: CMUNeXt and attentation U-net improve accuracy QTMnet based perfusion parameter estimates.