Quantitative susceptibility mapping via deep neural networks with iterative reverse concatenations and recurrent modules

Min Li, Chen Chen, Zhuang Xiong, Yin Liu, Pengfei Rong, Shanshan Shan, Feng Liu, Hongfu Sun, Yang GaoMedical 2025 · DOI

Introduction

Quantitative Susceptibility Mapping (QSM) is an advanced MRI technique that quantifies the magnetic susceptibility of biological tissues. This capability is essential for clinical applications such as monitoring iron deposition in neurodegenerative diseases like Parkinson’s or identifying cerebral hemorrhages. However, the fundamental challenge in QSM is the "dipole inversion" problem. The relationship between the measured MRI phase and the underlying tissue susceptibility is mathematically ill-posed because the magnetic dipole kernel contains zeros at specific angles in the frequency domain (k-space). This often results in "streaking" artifacts or necessitates multiple scans at different head orientations. While deep learning has shown promise, standard architectures often fail to capture the complex, non-local features required for high-fidelity reconstruction.

Proposed Method: IR²U-net

To address these challenges, the authors introduce IR²QSM, powered by a novel backbone called IR²U-net. The architecture is built upon three core innovations designed to enhance feature extraction and refinement:

  1. Iterative Refinement: The network does not rely on a single-pass reconstruction. Instead, it iterates through a specialized U-Net structure four times ($T=4$), with each step progressively refining the latent features from the previous iteration.
  2. Reverse Concatenation (RC): Unlike standard U-Nets that only use forward skip connections, IR²U-net incorporates "reverse" connections. These feed high-level semantic features from the decoder back into the encoder of the subsequent iteration, allowing low-level layers to benefit from global contextual information.
  3. Middle Recurrent Module (RM): A recurrent module, inspired by RNN architectures, is placed at the U-Net bottleneck. This module is specifically designed to capture long-distance spatial dependencies, which are critical in QSM because local magnetic fields are influenced by susceptibility sources throughout the brain.
The IR²U-net architecture featuring iterative refinement, reverse concatenation, and the recurrent module.

Experimental Setup and Datasets

The model was trained on a large synthetic dataset consisting of 14,400 3D patches derived from 96 simulated QSM volumes. To ensure robustness against real-world noise, the training process included a noise-adding module and a 5% dropout rate. Evaluation was conducted across several benchmarks:

  • Simulation Testing: 10 brain datasets simulated from COSMOS labels.
  • Public Benchmark: The 2019 QSM Reconstruction Challenge 2.0 dataset.
  • In Vivo Testing: Clinical data from healthy volunteers and patients with Cerebral Amyloid Angiopathy (CAA), Intracranial Hemorrhage (ICH), and Multiple Sclerosis (MS).

Results and Discussion

IR²QSM demonstrated superior performance across both quantitative and qualitative metrics. In simulated experiments, the method achieved an average Normalized Root Mean Square Error (NRMSE) of 27.59%, significantly outperforming traditional methods like iLSQR (by 15.48%) and MEDI (by 7.86%), as well as previous state-of-the-art deep learning models like xQSM (by 9.26%).

Qualitative comparison of IR²QSM against ground truth and ablation variants, highlighting artifact suppression and vessel detail.

Ablation studies revealed that Reverse Concatenation was the most critical design element for error reduction. Furthermore, the researchers identified $T=4$ as the optimal iteration count, balancing reconstruction accuracy with computational efficiency. Qualitatively, IR²QSM showed exceptional artifact suppression, particularly in challenging sinus regions near air-tissue interfaces, and maintained sharp details in small vessels without the over-smoothing common in other deep learning approaches like LPCNN.

Conclusion

The IR²QSM framework represents a significant advancement in susceptibility mapping by integrating iterative refinement with feedback loops and recurrent memory. While the iterative nature of the model increases computational cost compared to single-pass networks, the gains in accuracy and clinical robustness are substantial. Future work will focus on optimizing the architecture for different imaging resolutions and hardware configurations to facilitate real-time clinical deployment.

Summary updated

Abstract

Overall, the proposed IR<sup>2</sup>QSM showed superior QSM results compared to iterative and deep learning-based methods, offering a more accurate QSM solution for clinical applications.
BibTeX
@inproceedings{li2025quantitative,
  author = {Min Li and Chen Chen and Zhuang Xiong and Yin Liu and Pengfei Rong and Shanshan Shan and Feng Liu and Hongfu Sun and Yang Gao},
  title = {{Quantitative susceptibility mapping via deep neural networks with iterative reverse concatenations and recurrent modules}},
  booktitle = {Medical},
  year = {2025},
  doi = {10.1002/mp.17747},
}

Loading related content...

Figure