Abstract
A learning-based posterior distribution estimation method, Probabilistic Dipole Inversion (PDI), is proposed to solve the quantitative susceptibility mapping (QSM) inverse problem in MRI with uncertainty estimation. In PDI, a deep convolutional neural network (CNN) is used to represent the multivariate Gaussian distribution as the approximate posterior distribution of susceptibility given the input measured field. Such CNN is first trained on healthy subjects via posterior density estimation, where the training dataset contains samples from the true posterior distribution. Domain adaptations are then deployed on patient datasets with new pathologies not included in pre-training, where PDI updates the pre-trained CNN’s weights in an unsupervised fashion by minimizing the Kullback-Leibler divergence between the approximate posterior distribution represented by CNN and the true posterior distribution from the likelihood distribution of a known physical model and pre-defined prior distribution. Based on our experiments, PDI provides additional uncertainty estimation compared to the conventional MAP approach, while addressing the potential issue of the pre-trained CNN when test data deviates from training.
Introduction
Quantitative Susceptibility Mapping (QSM) is a non-invasive MRI technique used to quantify biomarkers such as iron, calcium, and gadolinium in tissue. However, reconstructing susceptibility maps from local magnetic field measurements requires solving an ill-posed inverse problem—specifically, undoing a dipole convolution. While deep learning (CNNs) has significantly improved reconstruction speed and accuracy, standard models often act as "black boxes," providing no measure of confidence and frequently failing when encountering pathologies (e.g., hemorrhages or lesions) not represented in the training data.
Methods: Probabilistic Dipole Inversion (PDI)
To address these limitations, we propose Probabilistic Dipole Inversion (PDI), which treats QSM as a Bayesian inference problem. Instead of a point estimate, PDI predicts a multivariate Gaussian posterior distribution $q_\psi(\chi|b)$ of susceptibility $\chi$ given the local field $b$.
The architecture utilizes a modified 3D U-Net with a dual-decoder: one branch outputs the posterior mean (the reconstructed QSM), and the other outputs the variance (the uncertainty map).
The training involves two stages:
- Posterior Density Estimation: The network is pre-trained on "gold standard" COSMOS data from healthy subjects to learn the general mapping between magnetic fields and tissue properties.
- Variational Inference (VI) Domain Adaptation: When applied to new pathologies, PDI performs unsupervised fine-tuning. It minimizes the Kullback-Leibler (KL) divergence between the network's predicted distribution and the true posterior derived from a physical dipole model. This physics-informed step ensures the model adapts to out-of-distribution data without requiring new labels.
Results and Discussion
PDI was evaluated on healthy subjects, Multiple Sclerosis (MS) patients, and Intracerebral Hemorrhage (ICH) patients.
Quantitative Performance and Uncertainty
On healthy test data, PDI achieved a pSNR of 47.77, outperforming clinical standards like MEDI (46.39). Crucially, the predicted standard deviation (SD) maps serve as effective proxies for reconstruction errors.
As shown in Figure 2, the uncertainty maps (STDs) highlight regions where the model is less confident, such as near large veins or high-susceptibility bleeds. This provides clinicians with a "map of doubt," a critical feature missing in standard CNN-based QSM.
Robustness to Pathologies
In patients with MS and ICH, standard pre-trained CNNs often underestimate susceptibility values. The PDI-VI adaptation successfully "recovers" these values by incorporating physical constraints during the inference phase. While subject-specific adaptation adds computational time, the amortized inference remains extremely fast (~0.61s), significantly quicker than classical iterative methods like MEDI (~17s).
Conclusion
PDI bridges the gap between fast deep learning and robust physics-based optimization. By providing uncertainty estimates and an unsupervised adaptation mechanism, it offers a more reliable and flexible tool for clinical QSM, particularly in complex pathological cases.
Reference: Zhang, J., et al. (2021). Probabilistic Dipole Inversion for Adaptive Quantitative Susceptibility Mapping. Journal of Machine Learning for Biomedical Imaging (MELBA).
Summary updated