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

Misha P. T. Kaandorp, Sebastiano Barbieri, Remy Klaassen, Hanneke W. M. van Laarhoven, Hans Crezee, Peter T. While, Aart J. Nederveen, Oliver J. Gurney-Champion

Research output: Contribution to journalArticleAcademicpeer-review

36 Citations (Scopus)

Abstract

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 diffusion-weighted imaging (DWI). This study presents a substantially improved version, IVIM-NET optim, and characterizes its superior performance in pancreatic cancer patients. Method: In simulations (signal-to-noise ratio [SNR] = 20), the accuracy, independence, and consistency of IVIM-NET were evaluated for combinations of hyperparameters (fit S0, constraints, network architecture, number of hidden layers, dropout, batch normalization, learning rate), by calculating the normalized root-mean-square error (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 an independent dataset of 23 patients with pancreatic ductal adenocarcinoma. Fourteen of the patients received no treatment between two repeated scan sessions and nine received chemoradiotherapy between the repeated sessions. Intersession within-subject standard deviations (wSD) and treatment-induced changes were assessed. Results: In simulations (SNR = 20), IVIM-NET optim outperformed IVIM-NET orig in accuracy (NRMSE(D) = 0.177 vs 0.196; NMRSE(f) = 0.220 vs 0.267; NMRSE(D*) = 0.386 vs 0.393), independence (ρ(D*, f) = 0.22 vs 0.74), and consistency (CV NET(D) = 0.013 vs 0.104; CV NET(f) = 0.020 vs 0.054; CV NET(D*) = 0.036 vs 0.110). IVIM-NET optim showed superior performance to the LS and Bayesian approaches at SNRs < 50. In vivo, IVIM-NET optim showed significantly less noisy 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, informative, and consistent IVIM fitting to DWI data.

Original languageEnglish
Pages (from-to)2250-2265
Number of pages16
JournalMagnetic resonance in medicine
Volume86
Issue number4
Early online date2021
DOIs
Publication statusPublished - Oct 2021

Keywords

  • IVIM
  • deep neural network
  • diffusion-weighted magnetic resonance imaging
  • intravoxel incoherent motion
  • pancreatic cancer
  • unsupervised physics-informed deep learning

Cite this