Gadolinium dose reduction for brain MRI using conditional deep learning

TL;DR

The paper tackles gadolinium dose reduction in brain MRI by isolating the true contrast enhancement signal from subtraction images and training a conditional CNN to denoise and amplify this signal, rather than directly predicting standard-dose images. Ground-truth CE is obtained from a discriminative mask applied to standard-dose subtraction images, allowing the model to focus on enhancement without fabricating noise. A diffusion-inspired conditioning scheme incorporates acquisition physics—field strength $B$, administered dose $d_\mathrm{LD}$, relaxivity $r_B$, and noise level $\sigma_\mathrm{LD}$—by injecting an embedding into each conditional residual block of a 3D U-Net, enabling CE synthesis beyond the standard dose when adding predicted CE to the input image. Extensive evaluation on synthetic and real multi-site datasets shows improved lesion CE accuracy and robust generalization, with embedding analysis confirming that the conditioning captures meaningful variations across scanners and agents. This approach offers a principled path toward reducing GBCA exposure while preserving diagnostic information and potentially enabling enhanced contrast in follow-up imaging.

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.

Figures (9)

• 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 contrast-enhanced images $\mathbf{x}_{\{\mathrm{LD},\mathrm{SD}\}}$ are preprocessed to obtain subtraction images $\mathbf{z}_{\{\mathrm{LD},\mathrm{SD}\}}$. The pairing of low-dose subtraction and low-dose image $\{\tilde{\mathbf{z}}_\mathrm{LD},\mathbf{x}_\mathrm{LD}\}$ is then processed by a conditional CNN generating a contrast signal $\hat{\mathbf{y}}_\mathrm{SD}$ that approximates the target $\mathbf{y}_\mathrm{SD}$. This target is obtained by removing the noise from the standard-dose subtraction image $\mathbf{z}_\mathrm{SD}$. At inference the predicted contrast signal $\hat{\mathbf{y}}_\mathrm{SD}$ is added to the pre-contrast $\mathbf{x}_\mathrm{PC}$ or low-dose image $\mathbf{x}_\mathrm{LD}$ to obtain a synthesized standard-dose image $\hat{\mathbf{x}}_\mathrm{SD}$ or a further contrast enhanced image $\hat{\mathbf{x}}_{\mathrm{SD}+}$.
• 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 while the bottom images visualize the effects on the subtraction image.
• Figure 3: Visualization of the empirical densities of the different subtraction images of a prototypical case. The plot on the left depicts the densities of the initial subtraction images $\tilde{\mathbf{z}}_\mathrm{LD}$ and $\tilde{\mathbf{z}}_\mathrm{SD}$ along with the fitted Gaussian model plotted in pink. The associated densities of the normalized subtraction images $\mathbf{z}_\mathrm{LD},\mathbf{z}_\mathrm{SD}$ are shown on the right. The normalization ensures that both densities overlap more nicely on the right for small intensities. For larger intensities, the contrast-enhancing pixels of the standard-dose subtraction images dominate. The pink line on the right illustrates the decision of the discriminative CE signal extraction function \ref{['eq:CESignalExtraction']}.
• Figure 4: The top row depicts the contrast signal extraction from the standard-dose subtraction image $\mathbf{z}_\mathrm{SD}$. The pixel-wise multiplication with the CE mask $\mathbf{p}_\mathrm{SD}$ yields the target contrast signal image $\mathbf{y}_\mathrm{SD}$. The bottom row visualizes the differences between the standard-dose image $\mathbf{x}_\mathrm{SD}$ and the standard-dose image implicitly generated by adding the contrast signal to the pre-contrast image $\mathbf{x}_\mathrm{PC}+\mathbf{y}_\mathrm{SD}$. Note that both standard-dose images only differ by the noise in non-enhancing regions.
• Figure 5: Qualitative comparison for test samples of the SLD-METS dataset. The zooms highlight the metastasis location. By design, our approach adds the CE signal to the input image, thereby preserving its image quality. The input images at the top have the same resolution, while they strongly differ at the bottom. Thus, the image quality of our output $\hat{\mathbf{x}}_\mathrm{SD}$ is poor in the bottom row and decent in the top row.