[Paper Review] Gadolinium dose reduction for brain MRI using conditional deep learning

Thomas Pinetz, Erich Kobler+8 more|arXiv (Cornell University)|Mar 6, 2024

Medical Imaging Techniques and ApplicationsMedicine3 citations

TL;DR

This paper proposes a conditional deep learning framework that reduces gadolinium-based contrast agent (GBCA) dose in brain MRI by learning to enhance contrast signal from low-dose subtraction images, avoiding artifact generation by training on noise-free standard-dose subtraction targets. The method achieves superior lesion visibility and enables synthetic images with contrast beyond standard dose, outperforming state-of-the-art methods in quantitative and qualitative evaluations.

ABSTRACT

Recently, deep learning (DL)-based methods have been proposed for the computational reduction of gadolinium-based contrast agents (GBCAs) to mitigate adverse side effects while preserving diagnostic value. Currently, the two main challenges for these approaches are the accurate prediction of contrast enhancement and the synthesis of realistic images. In this work, we address both challenges by utilizing the contrast signal encoded in the subtraction images of pre-contrast and post-contrast image pairs. To avoid the synthesis of any noise or artifacts and solely focus on contrast signal extraction and enhancement from low-dose subtraction images, we train our DL model using noise-free standard-dose subtraction images as targets. As a result, our model predicts the contrast enhancement signal only; thereby enabling synthesization of images beyond the standard dose. Furthermore, we adapt the embedding idea of recent diffusion-based models to condition our model on physical parameters affecting the contrast enhancement behavior. We demonstrate the effectiveness of our approach on synthetic and real datasets using various scanners, field strengths, and contrast agents.

Motivation & Objective

• To address the clinical challenge of reducing gadolinium-based contrast agent (GBCA) doses in brain MRI while preserving diagnostic accuracy.
• To overcome the limitations of existing deep learning methods that synthesize entire images, which can introduce artifacts or hallucinate pathologies.
• To disentangle contrast signal enhancement from image synthesis by focusing on subtraction images, thereby improving realism and diagnostic fidelity.
• To condition the deep learning model on physical acquisition parameters such as field strength, dose, and relaxivity to improve generalization and accuracy.
• To enable the generation of contrast-enhanced images with signal stronger than standard-dose images by adding predicted contrast signals to low-dose images.

Proposed method

• The method uses pre-contrast and low-dose contrast-enhanced MRI pairs to compute subtraction images, isolating the contrast signal from noise.
• A conditional convolutional neural network (CNN) is trained to predict the standard-dose contrast signal by minimizing deviation from a noise-suppressed target subtraction image derived from standard-dose scans.
• The target contrast signal is obtained by applying a statistical model to standard-dose subtraction images to suppress noise while preserving true contrast enhancement.
• The model is conditioned on physical parameters (e.g., field strength, dose, relaxivity) via learned embeddings, enabling accurate prediction across diverse acquisition settings.
• At inference, the predicted contrast signal is added to either the pre-contrast or low-dose image to synthesize a standard-dose or super-enhanced image, respectively.
• The approach avoids generative artifacts by focusing solely on contrast signal prediction rather than full image synthesis.

Figure 1: Visualization of our dose reduction and contrast enhancement DL approach. The inference process is illustrated at the top using solid arrows, while the additional processing steps for training are shown at the bottom. In both cases, pairs of pre-contrast $\mathbf{x}_{\mathrm{PC}}$ and cont

Experimental results

Research questions

• RQ1 Can a deep learning model effectively extract and enhance the true contrast signal from low-dose subtraction images without generating artifacts?
• RQ2 Does conditioning the model on physical acquisition parameters such as field strength, dose, and relaxivity improve the accuracy and generalization of contrast signal prediction?
• RQ3 Can the predicted contrast signal be used to generate images with contrast enhancement beyond the standard dose, improving lesion visibility?
• RQ4 How does the model’s performance compare to state-of-the-art methods in terms of quantitative metrics (PSNR, contrast enhancement) and qualitative image quality on both synthetic and real low-dose data?
• RQ5 To what extent do embedding vectors of the model reflect the underlying physical parameters of the MRI acquisition?

Key findings

• The proposed method achieved the highest PSNR scores in brain regions and for lesions on both synthetic (SLD-METS) and real low-dose (RLD) datasets, with statistically significant improvements over baselines.
• On the RLD dataset, the model significantly outperformed competitors in PSNR for lesion regions, demonstrating robustness to real-world acquisition variability.
• The model generated images with relative contrast enhancement exceeding standard-dose levels, as confirmed by mean and maximal contrast enhancement scores listed in Table 3.
• Embedding analysis using PCA and t-SNE revealed that the model’s latent space encodes physical parameters: field strength and dose show linear trends, while relaxivity and noise level form distinct clusters.
• Qualitative evaluation showed that the model preserves lesion details better than competing methods, even in cases where PSNR was not the highest, indicating improved diagnostic relevance.
• The use of noise-free standard-dose subtraction images as targets enabled artifact-free contrast signal prediction, avoiding the hallucination issues common in generative image synthesis.

Figure 2: Illustration of the preprocessing steps for extracting the initial low-dose subtraction image $\tilde{\mathbf{z}}_{\mathrm{LD}}$ from the pre-contrast $\mathbf{x}_{\mathrm{PC}}$ and low-dose $\mathbf{x}_{\mathrm{LD}}$ image. The top images in the steps show the transformed low-dose image w