CT2MRI

Slice-Consistent 3D Volumetric Brain CT-to-MRI Translation with 2D Brownian Bridge Diffusion Model

Early Accepted @ MICCAI 2024

Kyobin Choo, Youngjun Jun, Mijin Yun^*, Seong Jae Hwang^*

Yonsei University

Examples of slice inconsistency and resolved outcomes

We introduce Style Key Conditioning (SKC) and Inter-Slice Trajectory Alignment (ISTA) to address the slice inconsistency issue that arises in 3D volumetric image-to-image translation using 2D models.

Abstract

In neuroimaging, generally, brain CT is more cost-effective and accessible imaging option compared to MRI. Nevertheless, CT exhibits inferior soft-tissue contrast and higher noise levels, yielding less precise structural clarity. In response, leveraging more readily available CT to construct its counterpart MRI, namely, medical image-to-image translation (I2I), serves as a promising solution. Particularly, while diffusion models (DMs) have recently risen as a powerhouse, they also come with a few practical caveats for medical I2I. First, DMs' inherent stochasticity from random noise sampling cannot guarantee consistent MRI generation that faithfully reflects its CT. Second, for 3D volumetric images which are prevalent in medical imaging, naively using 2D DMs leads to slice inconsistency, e.g., abnormal structural and brightness changes. While 3D DMs do exist, significant training costs and data dependency bring hesitation. As a solution, we propose novel style key conditioning (SKC) and inter-slice trajectory alignment (ISTA) sampling for the 2D Brownian bridge diffusion model. Specifically, SKC ensures a consistent imaging style (e.g., contrast) across slices, and ISTA interconnects the independent sampling of each slice, deterministically achieving style and shape consistent 3D CT-to-MRI translation. To the best of our knowledge, this study is the first to achieve high-quality 3D medical I2I based only on a 2D DM with no extra architectural models. Our experimental results show superior 3D medical I2I than existing 2D and 3D baselines, using in-house CT-MRI dataset and BraTS2023 FLAIR-T1 MRI dataset.

Methods

Pipeline

Training and sampling scheme of the proposed methods. (a) During the multi-slice BBDM training, a target histogram-based style key is injected into the U-Net. (b) Target volume sampling proceeds in the manner of the Predictor-Corrector method. During the co-prediction phase, multiple ${\epsilon}^{i,k}_{\theta,t}$ are employed to establish connections among the predicted slices within ${\bar{X}}_{t-1}$. In the subsequent correction phase, the co-predicted volume is refined through a score-guided deterministic process.

Inter-Slice Trajectory Alignment (ISTA) Sampling

Visualization of the latent space and algorithm for ISTA sampling. The trained U-Net produces inconsistent outputs for multi-slice inputs that include the $i^{th}$ slice. The co-prediction unifies the direction of these independent inferences, while the correction aligns the co-predicted $\bar{{x}}_{t}^i$ onto the manifold of ${x}^i_{t}$.

Main Results

In-house Dataset (CT$\rightarrow$MRI)

BraTS2023 Dataset (FLAIR$\rightarrow$T1)

Quantitative Results

Additional Studies

Various Style Keys for SKC (CT$\rightarrow$MRI)

The figure shows real MRIs of main subject A, another subject X, and Colin 27 Average Brain (public access), along with synthetic MRIs of subject A. Both $\text{Ours}_{\text{X}}$ and $\text{Ours}_{\text{colin}}$ present synthetic MRIs of subject A, each transformed into the respective styles of X and Colin. From the analysis of the intensity histogram presented above, we observe that $\text{Ours}_{\text{X}}$ closely matches the MRI histogram of subject X, while $\text{Ours}_{\text{colin}}$ aligns with that of Colin. This demonstrates that the style of MRIs can be effectively controlled using SKC.