SC-BBDM: Slice-Consistent 3D Image Diffusion Model

• SC-BBDM is a diffusion-based framework that integrates slice-consistency with 2D synthesis for volumetric medical image translation.
• It leverages Style Key Conditioning (SKC) and Inter-Slice Trajectory Alignment (ISTA) to enforce global style and local continuity in image synthesis.
• Experimental results demonstrate improved PSNR, SSIM, and anatomical fidelity in CT-to-MRI and CTA translations compared to baseline models.

Searching arXiv for the cited SC-BBDM and foundational BBDM papers to ground the article. Slice-Consistent Brownian Bridge Diffusion Model (SC-BBDM) denotes a slice-consistent extension of Brownian Bridge Diffusion Models for volumetric medical image-to-image translation. In the literature, the formulation is most directly associated with the method introduced in “Slice-Consistent 3D Volumetric Brain CT-to-MRI Translation with 2D Brownian Bridge Diffusion Model” (Choo et al., 2024), where a 2D Brownian Bridge Diffusion Model is augmented with Style Key Conditioning (SKC) and Inter-Slice Trajectory Alignment (ISTA) to obtain style- and shape-consistent 3D synthesis without training a full 3D diffusion model. A later thoracic imaging study explicitly uses the full name “Slice-Consistent Brownian Bridge Diffusion Model (SC-BBDM)” for synthetic contrast-enhanced chest CT generation from non-contrast CT, extending the same conceptual machinery to synthetic CTA synthesis (Shiri et al., 23 Aug 2025).

1. Origins, terminology, and conceptual basis

SC-BBDM inherits its generative core from Brownian Bridge Diffusion Models (BBDMs), which recast image-to-image translation as a diffusion process bridging two domain endpoints rather than as conditional generation from pure Gaussian noise. In “BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models” (Li et al., 2022), the forward process is anchored at both source and target, so translation is modeled as direct stochastic interpolation between paired domains.

The original 2024 brain CT-to-MRI paper does not present “SC-BBDM” as a formal acronym; the method is described by its full title and by its two added components, SKC and ISTA (Choo et al., 2024). The shorthand is nevertheless precise enough to denote the method family: a slice-consistent extension of 2D BBDM for 3D volumetric translation. By contrast, the 2025 chest CTA paper explicitly names the method “Slice-Consistent Brownian Bridge Diffusion Model (SC-BBDM)” and positions it as a bridge-diffusion framework for paired non-contrast to arterial-phase CT translation (Shiri et al., 23 Aug 2025).

The central motivation is twofold. First, ordinary diffusion models are treated as suboptimal for faithful medical image-to-image translation because their stochasticity can weaken source-faithful reconstruction. Second, naively applying 2D generative models slice-by-slice to 3D volumes introduces global style inconsistency and local inter-slice discontinuity. SC-BBDM is therefore designed to preserve 3D anatomical integrity while retaining the memory advantages of high-resolution 2D synthesis (Choo et al., 2024).

2. Brownian-bridge formulation for volumetric translation

The base SC-BBDM formulation follows the BBDM bridge between a target image $\mathbf{x}_0$ and a source image $\mathbf{y}$. In the 2024 brain CT-to-MRI setting, $\mathbf{x}_0$ denotes target MRI and $\mathbf{y}$ denotes source CT. The forward bridge marginal is

$$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$

with

$m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$

This differs from DDPM-style corruption because the latent state is centered on a time-dependent interpolation between paired endpoints rather than on a monotone trajectory toward isotropic noise. The reverse process is written as

$$p_\theta(\mathbf{x}_{t-1}\mid \mathbf{x}_t,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_{t-1};\, \mu_\theta(\mathbf{x}_t,t), \tilde{\delta}_t \mathbf{I} \right).$$

The training target is likewise bridge-specific rather than pure noise. The denoiser is optimized with

$$\mathbb{E}_{\mathbf{x}_0,\mathbf{y},\boldsymbol{\epsilon}} \left[ c_{\epsilon t} \left\| m_t(\mathbf{y}-\mathbf{x}_0)+\sqrt{\delta_t}\boldsymbol{\epsilon} - \boldsymbol{\epsilon}_{\theta}(\mathbf{x}_t,t) \right\|_2^2 \right].$$

A critical extension for volumetric synthesis is the move from single slices to local sub-volumes. The notation in the original SC-BBDM paper distinguishes whole volumes $$\mathbf{X}, \mathbf{Y}, \mathbf{X}_t \in \mathbb{R}^{Z\times H\times W}$$, individual slices $x^i, y^i, x_t^i \in \mathbb{R}^{H\times W}$, and local sub-volumes $\mathbf{y}$0. The method sets $\mathbf{y}$1, so each training instance is effectively a 3-slice stack (Choo et al., 2024).

This multi-slice bridge formulation does not convert the model into a full 3D diffusion process. Instead, it preserves a 2D backbone while exposing the denoiser to immediate inter-slice context. A plausible implication is that SC-BBDM should be read as a 2.5D bridge model with slice-coupled training and slice-coupled sampling, rather than as volumetric diffusion in the strict voxel-wise sense.

3. Slice consistency via SKC and ISTA

SC-BBDM’s defining contribution is the explicit decomposition of slice consistency into a global style problem and a local trajectory-alignment problem. SKC addresses the former; ISTA addresses the latter (Choo et al., 2024).

Component	Function	Technical object
SKC	Volume-wide style consistency	Histogram, cumulative histogram, histogram differential
ISTA	Local inter-slice coherence	Co-prediction and deterministic score-based correction

SKC conditions the denoiser on a histogram-derived style key

$\mathbf{y}$2

with $\mathbf{y}$3, composed of three 1D descriptors of the target MRI volume: histogram, cumulative histogram, and histogram differential. The conditional training objective becomes

$\mathbf{y}$4

The role of SKC is to force all slices in a volume toward a coherent target appearance, mitigating abrupt brightness and contrast variation across slices. During inference, the default style key is the average histogram over the training set; the paper also evaluates a Colin 27 template histogram and a “best” target-proximal histogram (Choo et al., 2024).

ISTA couples neighboring slices during reverse sampling. Because each slice is present in multiple overlapping 3-slice windows, SC-BBDM generates multiple predictions per slice and averages them. For slice $\mathbf{y}$5,

$\mathbf{y}$6

and the co-predicted volume is

$\mathbf{y}$7

Co-prediction alone soft-couples neighboring slice trajectories. ISTA then adds a deterministic correction step,

$\mathbf{y}$8

with score

$\mathbf{y}$9

The paper explicitly removes Langevin noise from this corrector, so ISTA is fully deterministic. The intended effect is to align neighboring reverse trajectories with the same latent manifold and suppress slice-to-slice flicker or geometric discontinuity (Choo et al., 2024).

4. Architecture, training regime, and inference behavior

The original SC-BBDM implementation retains a 2D U-Net-based BBDM backbone and adds no extra architectural models. The denoiser operates on axial 3-slice stacks, conditions on the style key, and performs deterministic slice-coupled inference through ISTA. The method is explicitly presented as a way to obtain high-quality 3D medical image-to-image translation “based only on a 2D DM with no extra architectural models” (Choo et al., 2024).

For the brain CT-to-MRI study, both in-house CT-MRI and BraTS FLAIR-T1 data are resampled to $\mathbf{x}_0$0. The in-house CT-MRI volumes are cropped to $\mathbf{x}_0$1, and BraTS volumes to $\mathbf{x}_0$2. The preprocessing includes co-registration for in-house CT/MRI using SPM12, skull stripping with SynthStrip, and min-max normalization to $\mathbf{x}_0$3. Training uses $\mathbf{x}_0$4, diffusion time steps $\mathbf{x}_0$5, sampling steps $\mathbf{x}_0$6 for the standard BBDM sampler, and an ISTA fair-comparison setting of $\mathbf{x}_0$7 steps with $\mathbf{x}_0$8. Batch sizes are $\mathbf{x}_0$9 for CT$\mathbf{y}$0MRI and $\mathbf{y}$1 for FLAIR$\mathbf{y}$2T1; training iterations are $\mathbf{y}$3 and $\mathbf{y}$4, respectively; the reported hardware is an NVIDIA RTX A6000 (Choo et al., 2024).

The later thoracic CTA implementation is more explicit about U-Net configuration. It uses OpenAI’s UNet architecture with image size $\mathbf{y}$5, input/output channels $\mathbf{y}$6, base channels $\mathbf{y}$7, channel multipliers $\mathbf{y}$8, two residual blocks per level, attention resolutions $\mathbf{y}$9, and eight attention heads with 64 channels each. That study trains for 100 epochs on an NVIDIA RTX 3090 with 24 GB memory, uses 1000 diffusion steps, a linear schedule for $$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$0 from 0.001 to 0.999, and 50 DDIM ISTA sampling steps with correction factor $$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$1; inference time is about 23 minutes per case (Shiri et al., 23 Aug 2025).

The architectural distinction from full 3D diffusion is deliberate. SC-BBDM does not learn a full volumetric denoiser; instead, it combines local multi-slice context at training time with overlapping-window aggregation and deterministic trajectory correction at inference time. This suggests a design philosophy in which volumetric consistency is imposed by conditioning and sampler geometry rather than by replacing the 2D backbone with a 3D network.

5. Empirical performance and medical applications

The original 2024 study evaluates SC-BBDM on two paired volumetric translation tasks: brain CT$$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$2MRI on an in-house dataset and FLAIR$$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$3T1 on BraTS2023. On the in-house CT$$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$4MRI task, the reported scores are NRMSE $$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$5, PSNR $$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$6, and SSIM $$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$7, outperforming RevGAN, ALDM, MaskGAN, and Palette. On BraTS FLAIR$$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$8T1, the method reports NRMSE $$q_{BB}(\mathbf{x}_t \mid \mathbf{x}_0,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_t;\, (1-m_t)\mathbf{x}_0 + m_t\mathbf{y}, \delta_t \mathbf{I} \right),$$9, PSNR $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$0, and SSIM $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$1, again exceeding the same baselines (Choo et al., 2024).

The ablations clarify the relative roles of SKC and ISTA. Pure BBDM is already competitive on in-house CT$m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$2MRI but substantially weaker on BraTS. Adding SKC improves performance markedly, especially with the average training histogram. Co-prediction alone yields most of ISTA’s gain, while the deterministic correction adds a smaller but positive improvement. The best results arise from SKC plus ISTA, and the “best” style histogram, when available, yields the strongest quantitative scores. This supports the paper’s interpretation that style consistency and local trajectory alignment are complementary rather than interchangeable (Choo et al., 2024).

A later thoracic imaging study transfers SC-BBDM to synthetic arterial-phase chest CT generation from native non-contrast CT. It uses the Coltea-Lung dataset, filters out 38 low-quality cases, and evaluates two anatomical settings: Aortic Volume (AV) and Cardiac-Aortic Volume (CAV). The preprocessing is unusually elaborate: segmentation-based cropping removes spine and abdominal organs; native and arterial scans are aligned with SyN registration; masks are obtained with TotalSegmentator; AV uses 10-voxel aortic dilation, whereas CAV uses 20-voxel dilation including cardiac chambers; scans are resampled to $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$3 with voxel spacing $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$4 mm, clipped to HU range $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$5, and normalized to $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$6 (Shiri et al., 23 Aug 2025).

The thoracic results are mixed in a way that is informative rather than contradictory. On AV, SC-BBDM is strongest in voxel fidelity: GT-Test SC-BBDM achieves NRMSE $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$7, PSNR $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$8, SSIM $m_t=\frac{t}{T}, \qquad \delta_t = 2s(m_t-m_t^2).$9, NRMSE (NZ) $$p_\theta(\mathbf{x}_{t-1}\mid \mathbf{x}_t,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_{t-1};\, \mu_\theta(\mathbf{x}_t,t), \tilde{\delta}_t \mathbf{I} \right).$$0, and PSNR (NZ) $$p_\theta(\mathbf{x}_{t-1}\mid \mathbf{x}_t,\mathbf{y}) = \mathcal{N}\!\left( \mathbf{x}_{t-1};\, \mu_\theta(\mathbf{x}_t,t), \tilde{\delta}_t \mathbf{I} \right).$$1, outperforming CyTran and Pix2Pix in PSNR and NRMSE. On CAV, CyTran is slightly better numerically in PSNR and SSIM, while SC-BBDM remains competitive and is reported by radiologist review to preserve vascular structures and intensity distributions better than the baselines. The paper therefore supports SC-BBDM most strongly in the aorta-focused setting and in qualitative vascular fidelity, not as a uniformly dominant method across every metric and anatomical context (Shiri et al., 23 Aug 2025).

6. Scope, misconceptions, and limitations

A frequent misconception is to treat any Brownian-bridge medical translation model as slice-consistent. The literature does not support that equivalence. The CBCT-to-pCT work “Improving Cone-Beam CT Image Quality with Knowledge Distillation-Enhanced Diffusion Model in Imbalanced Data Settings” uses BBDM as the core translation model in a slice-based self-training framework, but explicitly introduces no slice-consistency, 3D, or volumetric continuity mechanism (Hwang et al., 2024). Likewise, the retrieval-augmented ReBrain framework employs 2D BBDM for sparse CT-to-MRI reconstruction and improves through-plane continuity with retrieval-conditioned ControlNet guidance and SLERP fallback, but it is not slice-consistent by construction in the SC-BBDM sense (Liu et al., 21 Nov 2025).

A second misconception is to identify SC-BBDM with generic trajectory-consistency ideas from other modalities. Speech enhancement with SE-Bridge uses Brownian-bridge trajectories plus consistency training so that nearby PF-ODE states map to the same clean endpoint, but that is local temporal consistency on bridge trajectories, not inter-slice volumetric consistency (Qiu et al., 2023). Frame interpolation with consecutive Brownian Bridge diffusion similarly addresses deterministic interpolation and low cumulative variance, but in an ordered video-latent setting rather than a slice-consistent volumetric one (Lyu et al., 2024). These works are conceptually adjacent, not equivalent.

The main limitations of SC-BBDM are also consistent across its current applications. It still relies on paired and well-registered data; the style key is a coarse histogram descriptor rather than a full appearance model; and the 2024 paper does not provide a dedicated numerical metric specifically designed for slice continuity or 3D smoothness, even though its qualitative analysis emphasizes sagittal and coronal consistency (Choo et al., 2024). In the 2025 chest CTA study, performance degrades on the more complex cardiac-aortic setting, the post-filtering dataset is small, all scans are from female patients, and no pathology labels are provided. The reported failure modes include blurring of anatomical borders, hyperenhancement, artifactual vessel narrowing, loss of cardiac chamber definition, and branch-vessel narrowing (Shiri et al., 23 Aug 2025).

Taken together, the literature supports a precise characterization: SC-BBDM is not merely BBDM applied slice-by-slice, and it is not a full 3D diffusion model. It is a 2D Brownian-bridge translation framework augmented with explicit style conditioning and deterministic inter-slice trajectory alignment so that high-resolution slice synthesis can behave as a volumetrically coherent generator (Choo et al., 2024).