Improved unsupervised physics-informed deep learning for intravoxel-incoherent motion modeling and evaluation in pancreatic cancer patients
share
Purpose: Earlier work showed that IVIM-NET_orig, an unsupervised physics-informed deep neural network, was faster and more accurate than other state-of-the-art intravoxel-incoherent motion (IVIM) fitting approaches to DWI. This study presents: IVIM-NET_optim, overcoming IVIM-NET_orig's shortcomings. Method: In simulations (SNR=20), the accuracy, independence and consistency of IVIM-NET were evaluated for combinations of hyperparameters (fit S0, constraints, network architecture, # hidden layers, dropout, batch normalization, learning rate), by calculating the NRMSE, Spearman's ρ, and the coefficient of variation (CV_NET), respectively. The best performing network, IVIM-NET_optim was compared to least squares (LS) and a Bayesian approach at different SNRs. IVIM-NET_optim's performance was evaluated in 23 pancreatic ductal adenocarcinoma (PDAC) patients. 14 of the patients received no treatment between 2 repeated scan sessions and 9 received chemoradiotherapy between sessions. Intersession within-subject standard deviations (wSD) and treatment-induced changes were assessed. Results: In simulations, IVIM-NET_optim outperformed IVIM-NET_orig in accuracy (NRMSE(D)=0.14 vs 0.17; NMRSE(f)=0.26 vs 0.31; NMRSE(D*)=0.46 vs 0.49), independence (ρ(D*,f)=0.32 vs 0.95) and consistency (CV_NET (D)=0.028 vs 0.185; CV_NET (f)=0.025 vs 0.078; CV_NET (D*)=0.075 vs 0.144). IVIM-NET_optim showed superior performance to the LS and Bayesian approaches at SNRs<50. In vivo, IVIM-NET_optim showed less noisy and more detailed parameter maps with lower wSD for D and f than the alternatives. In the treated cohort, IVIM-NET_optim detected the most individual patients with significant parameter changes compared to day-to-day variations. Conclusion: IVIM-NET_optim is recommended for accurate IVIM fitting to DWI data.
READ FULL TEXT
