Metacognitive physics-informed neural network for parameter estimation

Parameter estimation is critical for enabling intravoxel incoherent motion (IVIM) to provide accurate insights into biological functions. However, low signal‑to‑noise ratios (SNRs) in diffusion‑weighted (DW) data and the absence of ground‑truth labels present substantial challenges. To address this issue, we introduce a metacognitive physics-informed neural network (MC-PINN). This proposed model utilizes an iterative self‑feedback training strategy to achieve supervised learning using synthetic data informed by ground‑truth approximations while self‑assessing the reliability of its derivation. In comparative experiments using data from a magnetic resonance linear accelerator (MR‑Linac), MC‑PINN achieved robust parameter estimation even under low‑SNR conditions.