Abstract
Quantitative susceptibility mapping (QSM) provides a valuable tool for quantifying susceptibility distributions in human brains; however, two types of opposing susceptibility sources (i.e., paramagnetic and diamagnetic), may coexist in a single voxel, and cancel each other out in net QSM images. Susceptibility source separation techniques enable the extraction of sub-voxel information from QSM maps. This study proposes a novel SUSEP-Net for susceptibility source separation by training a dual-branch U-net with a simulation-supervised training strategy. In addition, a contrastive learning framework is included to explicitly impose similarity-based constraints between the branch-specific guidance features in specially-designed encoders and the latent features in the decoders. Comprehensive experiments were carried out on both simulated and in vivo data, including healthy subjects and patients with pathological conditions, to compare SUSEP-Net with three state-of-the-art susceptibility source separation methods (i.e., APART-QSM, χ-separation, and χ-sepnet). SUSEP-Net consistently showed improved results compared with the other three methods, with better numerical metrics, improved high-intensity hemorrhage and calcification lesion contrasts, and reduced artifacts in brains with pathological conditions. In addition, experiments on an agarose gel phantom data were conducted to validate the accuracy and the generalization capability of SUSEP-Net.
Introduction
Quantitative Susceptibility Mapping (QSM) is a powerful MRI technique used to measure the magnetic properties of brain tissue, serving as a critical biomarker for neurodegenerative diseases like Parkinson's and Multiple Sclerosis. However, QSM suffers from a "blind spot" known as sub-voxel cancellation. Because a single voxel often contains both paramagnetic sources (e.g., iron) and diamagnetic sources (e.g., myelin), their opposing magnetic signals can cancel each other out, leading to inaccurate measurements. Existing source separation algorithms often rely on unrealistic physical assumptions or produce significant reconstruction artifacts.
Methods
To address these challenges, the authors introduce SUSEP-Net, a deep learning approach utilizing a dual-branch U-Net architecture. The network processes standard MRI inputs—local field maps, R2' relaxometry, and initial QSM images—to produce separate maps for positive ($χ_{pos}$) and negative ($χ_{neg}$) susceptibility sources.
Two major innovations drive the performance of SUSEP-Net:
- Simulation-Supervised Training: Due to the lack of in vivo ground truth, the model is trained on synthetic datasets. By digitally injecting high-intensity lesions (hemorrhages and calcifications) into baseline maps and using a forward physical model to simulate MRI acquisition, the network learns the underlying physics of magnetic field perturbation.
- Contrastive Learning: The architecture includes specialized encoders to generate "guidance features." A contrastive loss function forces features within the same branch (e.g., positive) to be mathematically similar while pushing features from opposing branches apart, effectively reducing signal "leakage" between the reconstructed maps.
Results and Discussion
SUSEP-Net demonstrated superior performance across simulated, phantom, and in vivo datasets:
- Numerical Accuracy: In simulated healthy brains, SUSEP-Net achieved a Normalized Root Mean Squared Error (NRMSE) of 5.08% on positive maps, significantly outperforming competing methods like χ-sepnet (10.27%) and APART-QSM (11.52%).
- Ablation Studies: The inclusion of the contrastive learning module alone improved NRMSE by approximately 3.2%.
- Pathology and Generalization: On simulated pathological data, SUSEP-Net maintained a low NRMSE of 9.66% for negative maps, whereas traditional methods like APART-QSM reached 24.72% error. Real-world validation using an agarose gel phantom showed a near-perfect linear correlation ($R^2 \approx 0.97$) with actual chemical concentrations.
- In Vivo Performance: In patients with conditions such as Carbon Monoxide poisoning, SUSEP-Net successfully highlighted lesions while eliminating visual artifacts that typically plague older algorithms.
While the dual-branch design increases computational cost slightly (17.96s vs 14.31s for inference), the gain in accuracy and robustness makes SUSEP-Net a significant advancement in the field. Future work aims to explore fully unsupervised algorithms to eliminate dependency on bootstrapped labels.
Summary updated