Quantitative Breast MRI Using Physics-Informed Deep Learning

The intravoxel incoherent motion (IVIM) model offers a noninvasive means to quantify both diffusion and perfusion characteristics in breast MRI, providing valuable insight into the tumor microenvironment without the need for contrast agents. However, its clinical utility remains limited due to high sensitivity to noise, which leads to inaccurate and unstable estimation of key parameters such as diffusivity (D), the perfusion fraction (f) and pseudo-diffusion coefficient (D*). To address this challenge, we introduce PIA-IVIM, a Physics-Informed Autoencoder that learns the inverse of the IVIM function within its encoder through self-supervised training-effectively serving as a solver even under Rician-noise-corrupted input. The PIA-IVIM encoder is a deep neural network composed of fully connected layers that outputs estimates of f, D, and D*, while the decoder implements the differentiable IVIM forward model. PIA-IVIM was trained on inputs corrupted with varying levels of Rician noise, optimizing until the reconstruction loss converged. To evaluate PIA-IVIM's performance, we conducted three sets of experiments: (1) using in silico simulated data within the physical limits of breast tissue, (2) using virtual patients from the VICTRE breast digital phantom dataset, and (3) using an in vivo IVIM image from a patient with confirmed breast cancer. In in silico and phantom experiments, where ground truth f, D, and D* values were known, PIA-IVIM demonstrated strong noise robustness, achieving significantly lower mean absolute error and higher Spearman's rank correlation than conventional nonlinear least squares (NLLS) fitting under several noise levels. Additionally, PIA-IVIM is over 10,000 times faster than NLLS fitting, reducing the analysis time from minutes to milliseconds. In the highly noisy in vivo patient image, PIA-IVIM's f-map exhibited a significantly improved contrast-to-noise ratio in the cancerous region, whereas NLLS failed to highlight the tumor due to noise interference. We introduce PIA-IVIM as a new tool for quantitative breast MRI that enables accurate, fast, and robust solutions, even under challenging noise conditions.