Q-Net: A Quantitative Susceptibility Mapping-based Deep Neural Network for Differential Diagnosis of Brain Iron Deposition in Hemochromatosis

Soheil Zabihi, Elahe Rahimian, Soumya Sharma, Sean K. Sethi, Sara Gharabaghi, Amir Asif, E. Mark Haacke, Mandar S. Jog, Arash MohammadiarXiv 2021 · arXiv

Introduction

Hereditary Hemochromatosis (HH) is a prevalent genetic disorder characterized by excessive systemic iron absorption. While its systemic manifestations are well-documented, recent clinical evidence indicates significant iron accumulation within the brain's deep gray matter. Detecting this deposition is clinically challenging; standard radiological assessment often fails to distinguish HH patients from healthy controls (HC) because the differences in magnetic susceptibility and relaxation rates are extremely subtle and non-linear. To address this diagnostic gap, researchers developed Q-Net, a deep learning framework designed to identify complex patterns of brain iron distribution that elude human visual inspection.

Methods

The Q-Net architecture employs a two-stage modeling approach to process multi-modal MRI data:

  1. Image-Level Modeling: A ResNet18 backbone extracts high-dimensional embeddings from individual 2D brain slices. To enhance model stability and generalization, the team utilized Stochastic Weight Averaging (SWA), which averages weights across multiple training epochs to find a more robust local optimum.
  2. Scan-Level Modeling: Recognizing that brain anatomy is inherently sequential, the model incorporates a Bidirectional Long Short-Term Memory (Bi-LSTM) network. This recurrent layer processes the sequence of slice embeddings, capturing spatial dependencies across the 3D volume to produce a final patient-level diagnosis.

The model utilizes a three-channel input strategy, combining Quantitative Susceptibility Mapping (QSM), R2, and T1-Weighted (T1W)* images. This multi-modal approach allows the network to correlate quantitative iron metrics with precise anatomical structures.

Figure 1: Comparison of QSM, R2*, and T1W MRI modalities, demonstrating the subtle visual differences between subjects that Q-Net is designed to analyze.

Results and Discussion

The study evaluated Q-Net on a dataset of 99 subjects (52 HH, 47 HC). A critical finding was the impact of spatial localization: when the input was cropped to the basal ganglia—the primary site of iron deposition—diagnostic performance improved significantly.

Key performance metrics include:

  • Differential Diagnosis Accuracy: 83.16% (scan-level) and 80.37% (image-level).
  • Area Under the Curve (AUC): 0.875 for scan-level classification.
  • Statistical Significance: DeLong tests confirmed that the integrated Q-Net (CNN + Bi-LSTM) significantly outperformed standard ResNet architectures.

These results suggest that while mean iron values may overlap between healthy and diseased populations, the specific texture and distribution of iron—features effectively captured by deep neural networks—are highly diagnostic.

Conclusion and Future Work

Q-Net demonstrates the potential of AI to enhance the diagnostic utility of quantitative MRI in metabolic brain disorders. While the current results are promising, future research will focus on expanding the dataset through multi-center collaborations and exploring 3D convolutional architectures or Vision Transformers to further refine spatial feature extraction. Additionally, determining the relative diagnostic weight of QSM versus R2* could help optimize clinical imaging protocols for HH.

Summary updated

BibTeX
@misc{zabihi2021q,
  author = {Soheil Zabihi and Elahe Rahimian and Soumya Sharma and Sean K. Sethi and Sara Gharabaghi and Amir Asif and E. Mark Haacke and Mandar S. Jog and Arash Mohammadi},
  title = {{Q-Net: A Quantitative Susceptibility Mapping-based Deep Neural Network for Differential Diagnosis of Brain Iron Deposition in Hemochromatosis}},
  year = {2021},
  eprint = {2110.00203},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.00203},
}
Figure