Faithful Synthesis of Low-Dose Contrast-Enhanced Brain MRI Scans Using Noise-Preserving Conditional GANs

Parvin Mousavi, Thomas Pinetz, Erich Kobler, Robert Haase, Katerina Deike-Hofmann, Alexander Radbruch, Alexander EfflandMICCAI 2023 · DOI

Sure, here's a blog-style summary of the paper:

TL;DR

This paper introduces a novel conditional Generative Adversarial Network (cGAN) that synthesizes realistic low-dose contrast-enhanced brain MRI scans. By developing a noise-preserving content loss function, the model effectively captures subtle image details, crucial for accurate diagnosis, while reducing the need for expensive and potentially harmful contrast agents.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a vital tool for diagnosing various neurological conditions, often relying on contrast agents like Gadolinium-based contrast agents (GBCAs) to highlight abnormalities. However, GBCAs are costly and can accumulate in the body, raising concerns about potential side effects. Consequently, there's a strong push to reduce GBCA dosage while maintaining diagnostic image quality. A key challenge is understanding how much GBCA can be reduced without losing crucial information, especially in pathological regions. This is further complicated by the lack of publicly available datasets with varying GBCA dosages, hindering research and the development of robust AI models for this task.

Methods

The authors propose a conditional GAN (cGAN) framework for synthesizing low-dose contrast-enhanced (CE) brain MRI scans. The generator network, based on a U-Net architecture with local attention, takes a native MRI scan, a standard-dose CE scan, and metadata (including the desired low dose and field strength) as input. It predicts a residual image, which is then added to the native scan to produce the synthetic low-dose CE image.

The core innovation lies in their noise-preserving content loss function. Traditional content losses (like L1 or perceptual losses) tend to smooth out image details, including noise, which can be important for identifying subtle pathologies. Their proposed loss uses the Wasserstein distance to compare the empirical distributions of patches from real and synthesized images. This approach is designed to preserve local statistics and noise characteristics, leading to more realistic synthetic images. The discriminator network is trained to distinguish between real and synthesized residual images, also conditioned on the metadata.

Datasets

The study utilizes a newly collected dataset of 453 CE brain MRI scans with non-standard GBCA doses (10%, 20%, 33%) alongside corresponding standard-dose scans. This dataset was collected from two different scanner sites. For evaluation, they also used the BraTS dataset, a benchmark for brain tumor segmentation, to demonstrate the generalizability of their model.

Key Insights and Ideas

  • The importance of noise preservation: The paper highlights that simply smoothing out images to achieve a "clean" look can remove diagnostically relevant information, especially subtle contrast enhancements. Their noise-preserving loss is key to overcoming this.
  • Wasserstein distance for patch statistics: Using the Wasserstein distance to compare patch distributions is a powerful way to capture subtle differences in image texture and noise, going beyond simple pixel-wise comparisons.
  • Conditional GANs for dose interpolation: cGANs are shown to be effective for interpolating between known GBCA dose levels, enabling the synthesis of images at arbitrary fractional doses.
  • Generalizability across sites and contrast agents: The model demonstrates good performance even when tested on data from a different scanner site and using a different GBCA agent than what was used for training.

Key Results

  • The proposed noise-preserving content loss (NP) significantly improves the synthesis of realistic low-dose CE MRI scans compared to standard L1 and VGG-based perceptual losses.
  • Qualitative comparisons show that the NP-loss generates images with better preservation of noise texture and more accurate contrast enhancement, making it easier to identify pathological regions.
  • Quantitative evaluation shows that the NP-loss achieves competitive or superior performance in metrics like PSNR, SSIM, MAECE (mean absolute error on contrast-enhanced regions), and MAE (mean absolute error on noise standard deviation) compared to other methods.
  • When synthesized low-dose images are used to train virtual contrast models (VCMs), the NP-loss generated images lead to improved performance (higher PSNR and lower MAECE) compared to images synthesized with other loss functions.

Limitations and Future Work

  • Dataset limitations: While the authors collected a new dataset, it's still relatively small, and the number of distinct dose levels acquired is limited. Future work could involve expanding this dataset.
  • Generalizability to other pathologies: The current study focuses on contrast enhancement in general. Future work could explore the model's performance on specific types of pathologies and its ability to synthesize them accurately.
  • Clinical validation: While the results are promising, further clinical validation is needed to confirm the diagnostic utility of the synthesized images in real-world clinical settings.
  • Computational cost: Training GANs can be computationally intensive. Exploring more efficient training strategies or model architectures could be beneficial.

Summary updated

BibTeX
@inproceedings{mousavi2023faithful,
  author = {Parvin Mousavi and Thomas Pinetz and Erich Kobler and Robert Haase and Katerina Deike-Hofmann and Alexander Radbruch and Alexander Effland},
  title = {{Faithful Synthesis of Low-Dose Contrast-Enhanced Brain MRI Scans Using Noise-Preserving Conditional GANs}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43895-0\_57},
}
Figure