Improved unsupervised physics‐informed deep learning for intravoxel incoherent motion modeling and evaluation in pancreatic cancer patients

Purpose Earlier work showed that IVIM-NETorig, 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 diffusion-weighted imaging (DWI). This study presents a substantially improved version, IVIM-NEToptim, characterizes its superior performance in pancreatic cancer patients. Method In simulations (signal-to-noise ratio [SNR] = 20), the accuracy, independence, consistency of IVIM-NET were evaluated for combinations hyperparameters (fit S0, constraints, network architecture, number hidden layers, dropout, batch normalization, learning rate), by calculating normalized root-mean-square error (NRMSE), Spearman’s ρ, coefficient variation (CVNET), respectively. The best performing IVIM-NEToptim compared least squares (LS) Bayesian approach at different SNRs. IVIM-NEToptim’s independent dataset 23 patients with ductal adenocarcinoma. Fourteen received no treatment between two repeated scan sessions nine chemoradiotherapy sessions. Intersession within-subject standard deviations (wSD) treatment-induced changes assessed. Results (SNR outperformed IVIM-NETorig accuracy (NRMSE(D) 0.177 vs 0.196; NMRSE(f) 0.220 0.267; NMRSE(D*) 0.386 0.393), independence (ρ(D*, f) 0.22 0.74), (CVNET(D) 0.013 0.104; CVNET(f) 0.020 0.054; CVNET(D*) 0.036 0.110). LS SNRs < 50. vivo, significantly less noisy parameter maps lower wSD D f alternatives. treated cohort, detected most individual significant day-to-day variations. Conclusion is recommended accurate, informative, consistent IVIM DWI data.