Radiation Physics and Engineering
A peer-reviewed journal published by K. N. Toosi University of Technology

PKest-LSnet: A robust LSTM-Based deep learning framework for accurate pharmacokinetic parameter estimation in DCE-MRI with clinical validation

Articles in Press

Document Type : Research Article

Authors

Azimeh NV Dehkordi 1
Hossein Malekmohammadi 2
Sedigheh Sina 2, 3

1 Department of Physics, Na. C, Islamic Azad University, Najafabad, Iran

2 Mechanical Engineering Department, Shiraz University, Shiraz, Iran

3 Radiation Research Center, Shiraz University, Shiraz, Iran

https://doi.org/10.22034/rpe.2026.579118.1347
Abstract
Dynamic contrast-enhanced MRI (DCE-MRI) plays a pivotal role in quantifying tissue hemodynamics, yet accurate pharmacokinetic (PK) parameter estimation remains challenging. We present PKest-LSnet, a novel Long Short-Term Memory (LSTM)-based deep learning framework for robust PK analysis. The network was trained on comprehensive simulated data generated using physiologically plausible ranges of PK parameters (parameetrs of Tofts equation: vp, Ktrans, kep) and a population-based arterial input function (AIF), enabling precise ground-truth validation. Rigorous testing on clinical data from 19 glioblastoma patients demonstrated strong agreement with reference method maximum likelihood estimation (MPE <12% for all parameters), with 99.42% accuracy in classifying tumor signals. The model maintained stability across variable acquisition protocols (time intervals: 2-6 sec; SNR: 5–100), proving its adaptability to real-world clinical variability. While excelling in tumor characterization (DSC=77.27%), leaky vasculature classification challenges (DSC=36.76%) revealed opportunities for architectural refinements. PKest-LSnet eliminates dependency on initial parameter estimates, reduces computational time by orders of magnitude compared to conventional methods, and offers a robustness method to the tested variations in temporal sampling intervals and SNRs-critical for multicenter studies. This work bridges the gap between simulated training and clinical deployment, providing a validated tool for precision DCE-MRI analysis with direct applications in oncology and therapeutic monitoring. Future directions include hybrid architectures for improved intermediate-tissue classification and patient-specific AIF integration.

Highlights

  • The model maintained stability across variable acquisition protocols.
  • Excelling in tumor characterization (DSC=77%) leaky vasculature classification challenges (DSC=36%).
  • This work bridges the gap between simulated training and clinical deployment.

Keywords

DCE-MRI
Pharmacokinetic modeling
LSTM networks
Glioblastoma
Deep learning
Subjects

Copyright
RPE is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Conflict of Interest
The authors declare no potential conflict of interest regarding the publication of this work‎.

Funding
‎The authors declare that no funds‎, ‎grants‎, ‎or other financial support were received during the preparation of this manuscript‎.

Radiation Physics and Engineering

Articles in Press, Accepted Manuscript
Available Online from 09 September 2026

  • Receive Date 19 April 2026
  • Revise Date 29 August 2026
  • Accept Date 06 September 2026

Home

Submit Manuscript

Reviewers

Contact Us