Résumé
Quantitative MRI (qMRI) extracts tissue parameters as reproducible biomarkers. However, their estimation is limited by existing methods: traditional least-squares (LSQ) fitting is noise-sensitive and slow at inference, while deep-learning (DL) methods often require fixed-size inputs, which is impractical for variable-length clinical acquisitions. In this work, we investigate Neural Controlled Differential Equations (NCDE), a model that natively handles variable-length sequences, for qMRI parameter estimation. We compare the NCDE against a LSQ baseline and two fixed-size DL models on both synthetic and real pancreatic data, across intravoxel incoherent motion (IVIM) and true transverse relaxation rate (R2*) parameters. On synthetic signals, the NCDE remained robust to variable-length input, degrading far less than fixed-size models, while keeping inference fast and clinically viable. On real data, its estimates agreed with reference values from the literature and yielded competitive repeatability, confirming that the accuracy of NCDE-based estimation is preserved in vivo. The code is available at: https://github.com/PINKcc-Lab/FlexQMRI