Conditional Generative Models for Contrast-Enhanced Synthesis of T1w and T1 Maps in Brain MRI

TL;DR

This work addresses enhancement prediction by neural networks with two new contributions, and studies the potential of generative models, more precisely conditional diffusion and flow matching, for uncertainty quantification in virtual enhancement.

Abstract

Contrast enhancement by Gadolinium-based contrast agents (GBCAs) is a vital tool for tumor diagnosis in neuroradiology. Based on brain MRI scans of glioblastoma before and after Gadolinium administration, we address enhancement prediction by neural networks with two new contributions. Firstly, we study the potential of generative models, more precisely conditional diffusion and flow matching, for uncertainty quantification in virtual enhancement. Secondly, we examine the performance of T1 scans from quantitive MRI versus T1-weighted scans. In contrast to T1-weighted scans, these scans have the advantage of a physically meaningful and thereby comparable voxel range. To compare network prediction performance of these two modalities with incompatible gray-value scales, we propose to evaluate segmentations of contrast-enhanced regions of interest using Dice and Jaccard scores. Across models, we observe better segmentations with T1 scans than with T1-weighted scans.

Figures (5)

• Figure 1: Illustration of E2E network (top) versus conditional generative one (bottom). While E2E networks provide only one contrast enhanced prediction for a pre-contrast image $y$, generative models produce many samples (three depicted) from the distribution of post-contrast images conditioned to $y$. The sample mean gives a contrast prediction and their standard deviation shows areas of uncertainty.
• Figure 2: Two zoomed-in examples of HGG enhancement prediction with bad end-to-end (E2E) network performance. E2E does not enhance the tumor region at all (first example) or only a small part (second example). Flow matching (FM) mean shows blurred enhancement, but the standard deviation (StdDev) clearly indicates the uncertainty in the tumor area.
• Figure 3: Two examples of HGG enhancement prediction by FM, resp. DM and segmentation for T1 and T1w scans. Based on the difference between ground-truth pre- and post-contrast slices, the regions of interest and thresholding, we get comparable T1 and T1w ground-truth segmentations (3rd column, 20% outlier threshold) which can be compared with FM (5th column) and DM segmentations (7th column) via the Dice and Jaccard score.
• Figure 4: Segmentation comparison (Dice and Jaccard $\uparrow$) of contrast-enhanced region estimated by applying a voxel-wise difference outlier threshold between 0% and 30% to the ROI. T1 (blue dashed) gives better results than T1w (orange full).
• Figure 5: Visualization of FM, DM, and E2E prediction for T1 and T1w for HGG and MET test slices, and HGG test slice without visible tumor (marked in white) (top to bottom).