A Cascaded Neural Network for Intravoxel Incoherent Motion Map Reconstruction

Motivation: Quantitative estimation of intravoxel incoherent motion (IVIM) maps is readily influenced by noise and artifacts in the DWI images and the processing time can be relatively high. Goal(s): Use deep learning to improve the accuracy and robustness of parameter estimation for IVIM and speed up the computation. Approach: We designed a simple cascaded network to reconstruct IVIM maps from noisy dMRI, consisting of one denoising network and one fitting network. Results: Accurately reconstructed three parameter maps, f, D, and Ds, achieving the first place in the IVIM-dMRI Reconstruction Challenge hosted by AAPM. Impact: With improved accuracy and robustness, IVIM maps can be used to assess tissue microstructural properties, which are critical to diagnosis of diseases, therapy planning and treatment responses assessment.