Abstract
Deep learning (DL) based contrast dose reduction in MRI is gaining traction due to the potential risks associated with Gadolinium-based Contrast Agents (GBCAs). However, these algorithms are limited by the scarcity of high-quality low-dose datasets. We propose Gformer, a novel iterative transformer-based model for synthesizing MRI images with arbitrary contrast enhancement levels. By incorporating a sub-sampling based attention mechanism and a rotational shift module, Gformer captures global contextual features and irregular tumor shapes more effectively than existing methods. Quantitative evaluations demonstrate superior performance in dose simulation and downstream clinical tasks like tumor segmentation.
Introduction
Gadolinium-Based Contrast Agents (GBCAs) are essential for visualizing pathologies in MRI, yet they present risks including side effects and long-term accumulation in the body. Adhering to the ALARA (As Low As Reasonably Achievable) principle is critical. While deep learning can enhance low-dose images to match full-dose quality, training such models requires large, paired datasets of low-dose and full-dose scans—data that is difficult and expensive to acquire. Furthermore, different pathologies may require specific "minimum" doses for reliable diagnosis. There is a clear need for a flexible, data-driven method to simulate arbitrary dose levels to facilitate both AI training and clinical dose optimization.
Methodology
The proposed framework, the Iterative Global Transformer (Gformer), approaches contrast simulation as a gradual, multi-step process rather than a single-step prediction.
Iterative Design
The model treats the synthesis of low-dose images as the iterative removal of contrast enhancement over $k$ steps. This design allows the model to better interpolate between the sparse data points available in clinical datasets (typically only 0%, 10%, and 100% dose).
Gformer Architecture
The core of the model is the Gformer block, which introduces two key innovations:
- Sub-sampling Based Attention: Unlike traditional transformers that focus on local windows, Gformer uses a strided sub-sampling approach to create sub-images from the entire input. This enables the capture of global contextual information—crucial for understanding contrast distribution across the brain—with minimal computational cost.
- Rotational Shift Module: To account for the heterogeneous and irregular shapes of tumors, a rotational shift is applied. This helps the model learn features that are not strictly aligned to a standard grid.
Experimental Setup
The model was trained and validated using two distinct datasets:
- Site A: 126 clinical cases using Gadoterate meglumine, providing pre-contrast, 10% low-dose, and full-dose T1w MPRAGE scans.
- Site B: 159 cases using Gadobenate dimeglumine, used to evaluate the model's generalizability across different agents and scanners.
Results
Gformer significantly outperformed state-of-the-art architectures, including Rednet, Restormer, and SwinIR.
Quantitative Performance
On the 10% dose simulation task, Gformer achieved a PSNR of 42.29 dB and an SSIM of 0.98. These metrics indicate high structural fidelity and noise characteristics comparable to real scans.
Arbitrary Dose Simulation
A unique strength of the iterative approach is the ability to simulate any intermediate dose level. By adjusting the number of iterations or using the learned "soft labels," clinicians can visualize how a tumor might appear at 30%, 50%, or 80% of a standard dose.
Clinical Utility
The simulated images were validated through a downstream tumor segmentation task. An AI trained on Gformer-simulated low-dose images achieved a Mean Dice Score of 0.876, nearly matching the performance of a model trained on real low-dose data (0.889). This confirms that the simulated images preserve the diagnostic features necessary for clinical applications.
Conclusion
Gformer provides a robust solution for simulating arbitrary MRI contrast doses, effectively bridging the data gap for dose-reduction AI. By capturing global context and irregular features through its iterative transformer design, it offers a high-fidelity tool for both researchers and clinicians. Future work will focus on expert radiologist reader studies to establish diagnostic equivalence and the integration of physics-based tracer kinetics to further refine the simulation.
Summary updated