Simulation of Arbitrary Level Contrast Dose in MRI Using an Iterative Global Transformer Model

Dayang Wang, Srivathsa Pasumarthi, Greg Zaharchuk, Ryan Chamberlain

TL;DR

Researchers have developed Gformer, an iterative vision transformer model that can simulate any level of MRI contrast enhancement from standard scans. By enabling the synthesis of high-quality "low-dose" images, this tool helps overcome data scarcity for training dose-reduction AI and allows clinicians to determine the minimum effective contrast dose for different pathologies.

What Problem Does This Paper Address?

Gadolinium-Based Contrast Agents (GBCAs) are the gold standard for highlighting tumors and lesions in MRI. However, they aren't without risks: they can cause side effects, are contraindicated for patients with kidney issues, and may even accumulate in the body over time. This has led to the ALARA principle—"As Low As Reasonably Achievable"—in medical imaging.

While deep learning models can help "boost" low-dose images to look like full-dose ones, training these models is a "chicken and egg" problem. To train an AI to reduce dose, you first need a massive dataset of paired low-dose and full-dose images. Collecting this data is expensive, time-consuming, and requires changing clinical protocols. Furthermore, different diseases might require different "minimum" doses to be visible. There is currently no flexible, data-driven way to simulate these arbitrary dose levels—until now.

Methods

The authors propose a novel framework called the Iterative Global Transformer (Gformer). Instead of trying to predict a low-dose image in one giant leap, the model uses an iterative learning design. It treats the process as a gradual removal of contrast enhancement over $k$ steps.

Key Innovations:

  1. Gformer Backbone: Traditional Transformers (like Swin) often look at local "windows" of pixels. Gformer uses a sub-sampling based attention mechanism. It creates sub-images by taking pixels at strided intervals across the entire image. This allows the model to capture global contextual information (how contrast looks across the whole brain) with very low computational overhead.
  2. Rotational Shift Module: To better capture the irregular, heterogeneous shapes of tumors and contrast uptake regions, the authors introduced a rotational shift. This helps the model learn features that aren't just aligned to a horizontal/vertical grid.
  3. Soft Labels: Since the model is iterative, it needs a target for each intermediate step. The researchers used a linear scaling of the estimated contrast uptake to create "soft labels," guiding the model through the "de-enhancement" process during training.

Datasets

The study utilized two primary datasets:

  • Site A (Internal): 126 clinical cases (113 for training, 13 for testing) using the contrast agent Gadoterate meglumine. This included pre-contrast, 10% low-dose, and full-dose T1w MPRAGE scans.
  • Site B (External): 159 patient studies using a different agent (Gadobenate dimeglumine) to test how well the model generalizes to different chemicals and scanners.
  • Preprocessing: All images were skull-stripped, co-registered, and mean-normalized.

Key Insights and Ideas

  • Iterative is Better for Imbalance: The authors found that iterative models are better at interpolating between data points when the training data is highly imbalanced (e.g., having only 0%, 10%, and 100% dose images).
  • Global Context Matters: Contrast enhancement isn't just local; the way a tracer moves through the brain depends on the overall anatomy. The sub-sampling attention allows the model to "see" the whole brain at once without the massive memory cost of standard global attention.
  • Clinical Utility via Downstream Tasks: The paper doesn't just stop at "the images look good." They proved the simulated images are useful by showing that an AI trained on simulated low-dose images performs almost as well at tumor segmentation as one trained on real low-dose images.

Key Results

The Gformer outperformed several state-of-the-art (SOTA) architectures, including Rednet, Restormer, and SwinIR.

  • Quantitative Accuracy: On the 10% dose simulation task, Gformer achieved a PSNR of 42.29 dB and an SSIM of 0.98, significantly higher than competing methods.
  • Tumor Segmentation: When using simulated images for tumor segmentation, the model achieved a Mean Dice Score of 0.876, which was statistically similar to the performance using real low-dose images (0.889).
  • Efficiency: Despite its complexity, the model maintains a reasonable throughput of 0.65 images per second, making it viable for research and clinical pipelines.

Limitations and Future Work

While the results are promising, the authors note a few areas for growth:

  • Reader Studies: The next step is to have expert radiologists evaluate the simulated images to ensure "diagnostic equivalence"—checking if a human would make the same call on a simulated image as a real one.
  • Physics Integration: The model is currently purely data-driven. Future versions could be guided by physics-based models that estimate contrast enhancement based on signal intensity and tracer kinetics.
  • Expansion: The authors plan to extend this technique to other parts of the body (anatomies) and different types of contrast agents.

Summary updated

Abstract

Deep learning (DL) based contrast dose reduction and elimination in MRI imaging is gaining traction, given the detrimental effects of Gadolinium-based Contrast Agents (GBCAs). These DL algorithms are however limited by the availability of high quality low dose datasets. Additionally, different types of GBCAs and pathologies require different dose levels for the DL algorithms to work reliably. In this work, we formulate a novel transformer (Gformer) based iterative modelling approach for the synthesis of images with arbitrary contrast enhancement that corresponds to different dose levels. The proposed Gformer incorporates a sub-sampling based attention mechanism and a rotational shift module that captures the various contrast related features. Quantitative evaluation indicates that the proposed model performs better than other state-of-the-art methods. We further perform quantitative evaluation on downstream tasks such as dose reduction and tumor segmentation to demonstrate the clinical utility.
BibTeX
@inproceedings{wang2023simulation,
  author = {Dayang Wang and Srivathsa Pasumarthi and Greg Zaharchuk and Ryan Chamberlain},
  title = {{Simulation of Arbitrary Level Contrast Dose in MRI Using an Iterative Global Transformer Model}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43993-3\_9},
}

Loading related content...

Figure