Spectral Consistent Flow for One-step 3D Medical Image Translation¶

Conference: ECCV 2026
Paper: ECCV Official
Area: Medical Imaging
Keywords: 3D Medical Image Translation, Flow Matching, Brownian Bridge, Power Spectral Density, One-step Generation

TL;DR¶

Presents Spectral Consistent Flow (SC-Flow), an efficient 1-NFE 3D medical image translation framework that combines a stochastic Brownian bridge with a Spectral Consistency Corrector (SCC) to resolve high-frequency attenuation, over-smoothing, and modality entanglement in latent average velocity fields.

Background & Motivation¶

3D medical image translation (MIT)—such as synthesizing PET from MRI or generating CT from MRI—plays a pivotal role in multi-modal clinical diagnostic workflows. By synthesizing missing imaging sequences or functional metabolic maps from accessible scans, it substantially reduces patient radiation dose, scanning duration, and economic cost. However, volumetric medical images possess intricate spatial geometries and rich anatomical textures. Prior translation architectures grounded in voxel-space diffusion models or multi-step iterative ODE integrators incur prohibitive memory footprints and inference latencies (often taking tens of seconds per volume), preventing real-time intraoperative deployment.

Flow Matching (FM) and MeanFlow recently emerged as prominent single-function evaluation (1-NFE) paradigms by directly learning average velocity fields in latent space. Nonetheless, adapting continuous normalizing flows to paired 3D MIT introduces severe physical and statistical bottlenecks. Because paired 3D training scans are scarce, data distributions reside on sparse, lower-dimensional submanifolds; deterministic straight paths are forced across zero-density voids, destabilizing velocity field gradients. Furthermore, the time-integrated average velocity acts as an implicit low-pass filter over spatial frequencies. When compounded with the low-frequency compression bias of 3D latent encoders, reconstructed volumes suffer steep Power Spectral Density (PSD) decay, resulting in smoothed anatomical boundaries and loss of clinical texture fidelity. Meanwhile, bridge flows lacking explicit target-modality spectral supervision frequently lead to feature entanglement between source and target domains.

This paper tackles these challenges by transforming the linear transport path into an endpoint-conditioned stochastic diffusion process and closing the energy conservation loop in the frequency domain. Core idea: formulate latent flow matching as a stochastic Brownian bridge process to broaden manifold support overlap, coupled with an end-to-end Spectral Consistency Corrector (SCC) that dynamically modulates radial frequency gains to counteract high-frequency decay and preserve fine anatomical realism.

Method¶

Overall Architecture¶

SC-Flow operates within a compact latent space parameterized by a 3D autoencoder (VAE) to overcome the extreme memory demands of volumetric data. Given source volume $x_1$ encoded into latent $z_1$, the system predicts the target latent $\hat{z}_0$ in a single neural evaluation ($1$-NFE) and reconstructs high-fidelity target volume $\hat{x}_0$ via the decoder. The end-to-end architecture tightly coordinates two innovative components: the Support Regularized Flow and the Spectral Consistency Corrector (SCC).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Source Volume x1"] --> B["Pre-trained Encoder E<br/>Extract Latent z1"]
    B --> C["Support Regularized Flow<br/>Brownian bridge & average velocity u_theta"]
    C --> D["One-Step Sampling (1-NFE)<br/>Predict target latent z0 & decode initial volume"]
    D --> E["Spectral Consistency Corrector (SCC)<br/>3D FFT radial binning & adaptive gain modulation"]
    E --> F["Inverse 3D FFT (IFFT)<br/>Output restored high-fidelity volume x0"]

Key Designs¶

1. Support Regularized Flow: stochastic Brownian bridge for sparse manifold overlap In small-sample 3D medical cohorts, distributions $p_A$ and $p_B$ share minimal support. Conventional Deterministic Flow Matching (DFM) forces interpolation trajectories through low-density manifolds, causing unstable velocity gradients. SC-Flow extends the deterministic MeanFlow trajectory into an endpoint-constrained Brownian bridge stochastic process: $$dz_t = v_\theta(z_t, t)dt + \sigma(t)dW_t$$ where $W_t$ is a standard Wiener process and the time-dependent diffusion coefficient satisfies $\sigma(0)=\sigma(1)=0$. The marginal state at any time $t \in [0, 1]$ is analytically parameterized as: $$z_t = z_0(1-t) + z_1 t + \sqrt{t(1-t)}\epsilon$$ The diffusion term injects Gaussian perturbations into the average velocity field, expanding the empirical support and functioning as an entropy regularizer that stabilizes cross-modal field estimation. Neural velocity network $u_\theta(z_t, r, t)$ models this average velocity field. During single-step inference ($r=0, t=1$), the noise term $\sqrt{t(1-t)}\epsilon$ vanishes identically, yielding an exact, noise-free single-step formulation: $\hat{z}_0 = z_1 - u_\theta(z_1, 0, 1)$.

2. Spectral Consistency Corrector: radial binning and learnable gain field Average velocity modeling implicitly attenuates high frequencies, causing severe drop-offs in the power spectral density (PSD). By Parseval's theorem, spatial texture preservation is isomorphic to frequency-domain energy retention. Directly computing PSD across 3D voxels is noisy and computationally prohibitive. SC-Flow performs isotropic radial binning over the 3D Fast Fourier Transform (FFT) coefficients, partitioning frequency coordinates into $B$ radial bins via soft masks $m_b(\mathbf{k})$ according to radius $r = |\mathbf{k}|$: $$P_{x,b} = \sum_{\mathbf{k}} m_b(\mathbf{k}) |\mathcal{F}_x(\mathbf{k})|^2$$ To account for patient-specific anatomical variations, a lightweight encoder $\phi(\cdot)$ aggregates a joint conditioning vector $[\phi(\hat{x}_0), \phi(x_1)]$. An MLP then predicts band-wise gain coefficients $g_b$, synthesizing a continuous frequency gain map $g_{map}$ that modulates Fourier amplitudes: $$\tilde{P}_x = \left(\sum_{b=1}^B g_b m_b(\mathbf{k})\right)^2 \hat{P}_x$$ Applying the 3D Inverse FFT (IFFT) to this gain-corrected spectrum completes the translation, simultaneously recovering high-frequency tissue textures and preventing modality feature leakage without adding multi-step sampling overhead.

Loss & Training¶

The framework is trained end-to-end by jointly optimizing the latent generator and the volumetric decoder across spatial and spectral domains: $$\mathcal{L} = \lambda_1 \mathcal{L}_{mf} + \lambda_2 \mathcal{L}_{voxel} + \lambda_3 \mathcal{L}_{spec}$$ The mean-flow loss $\mathcal{L}_{mf}$ supervises average velocity predictions: $$\mathcal{L}_{mf}(z_t, r, t) = \mathbb{E} \|u_\theta(z_t, r, t) - u(z_t, r, t)\|_2^2$$ The spatial voxel loss $\mathcal{L}_{voxel}$ enforces macro-structural alignment via $L_1$ and SSIM objectives. The spectral objective $\mathcal{L}_{spec}$ minimizes the weighted log mean squared error across radial bins: $$\mathcal{L}_{spec} = \sum_b w_b (\log \hat{P}_{x,b} - \log P_{x,b})$$ Using logarithmic energy converts multiplicative spectral attenuation into an additive objective, providing robust gradient updates across both low- and high-frequency regimes.

Key Experimental Results¶

Main Results¶

SC-Flow was validated across four comprehensive datasets (ADNI, A4, NACC, synthRAD2025) spanning 8 cross-modal translation setups including MRI to Aβ/FDG/Tau-PET, CT to MRI, and CT to CBCT.

Paper Table 1: Quantitative comparison of MRI to Aβ-PET, FDG-PET, and Tau-PET translation

Translation Task (Dataset)	Method	NFE	MAE (%) ↓	RMSE (%) ↓	PSNR ↑	SSIM (%) ↑
MRI → Aβ-PET (ADNI)	LDM [53]	1000	6.35	11.57	19.03	63.60
	IL-CLDM [47]	1000	4.13	8.34	22.46	77.34
	PASTA [37]	100	3.63	7.61	23.04	85.18
	CycleGAN [71]	1	3.21	6.91	23.69	86.18
	MeanFlow [17]	1	3.74	8.05	22.16	78.32
	Causal-PETs [38]	1	3.97	7.13	23.57	87.08
	SC-Flow (Ours)	1	2.98	6.32	24.72	88.71
MRI → FDG-PET (NACC)	LDM [53]	1000	18.35	26.06	11.81	43.09
	PASTA [37]	100	3.38	7.75	22.43	84.64
	CycleGAN [71]	1	2.17	5.24	25.70	90.10
	MeanFlow [17]	1	3.38	7.99	21.96	78.18
	Causal-PETs [38]	1	2.03	4.66	26.70	92.47
	SC-Flow (Ours)	1	1.77	4.34	27.41	93.30
MRI → Tau-PET (A4)	LDM [53]	1000	14.41	22.30	13.07	49.32
	PASTA [37]	100	2.84	5.90	24.99	88.48
	MeanFlow [17]	1	3.27	7.32	22.81	79.67
	Causal-PETs [38]	1	2.63	5.40	25.83	89.32
	SC-Flow (Ours)	1	2.38	5.09	26.43	90.33

Paper Table 2: Quantitative comparison of CT to MRI / CBCT on synthRAD2025

Task	Method	MAE (%) ↓	RMSE (%) ↓	PSNR ↑	SSIM (%) ↑
CT → MRI	CycleGAN [71]	7.02	12.24	19.00	70.46
	MeanFlow [17]	4.29	9.76	20.71	72.20
	LBM [6]	4.08	9.40	21.05	74.59
	SC-Flow (Ours)	3.77	8.53	21.84	78.40
CT → CBCT	CycleGAN [71]	4.41	7.69	22.85	82.39
	MeanFlow [17]	3.62	6.71	23.99	82.97
	LBM [6]	3.01	5.63	25.91	88.37
	SC-Flow (Ours)	2.92	5.53	26.10	87.69

Ablation Study¶

The authors systematically evaluated the contribution of the Brownian bridge diffusion term ($dW_t$), the Spectral Consistency Corrector architecture (SCC), and the spectral loss objective ($\mathcal{L}_{spec}$) on the NACC benchmark (MRI → FDG-PET).

Paper Table 4: Component ablation on NACC dataset

Config	$dW_t$ (Brownian Bridge)	SCC (Corrector Module)	$\mathcal{L}$ (Spectral Loss)	MAE (%) ↓	RMSE (%) ↓	PSNR ↑	SSIM (%) ↑	Description
1	✗	✗	✗	3.38	7.99	21.97	78.18	Vanilla MeanFlow baseline
2	✓	✗	✗	2.47	5.96	24.59	86.54	Brownian bridge regularization only
3	✗	✓	✗	2.57	6.19	24.25	84.33	SCC module only
4	✗	✓	✓	2.50	6.03	24.47	85.69	SCC with spectral loss
5	✓	✓	✗	2.42	5.83	26.14	87.72	Bridge + SCC without loss
6	✓	✗	✓	2.11	5.15	25.86	89.11	Bridge + spectral loss
7 (Full)	✓	✓	✓	1.77	4.34	27.41	93.30	Complete SC-Flow model

In the radial bin sensitivity analysis (Paper Table 6), $B=24$ bins achieved the optimal balance (PSNR 27.24 dB, SSIM 92.77%). Finer discretization (e.g., $B=96$) induced spectral variance and overfitted noise (PSNR dropping to 25.97 dB), while overly coarse binning ($B=12$) failed to distinguish subtle frequency bands.

Key Findings¶

Synergy between stochastic bridging and spectral correction: Adding Brownian bridge perturbations alone boosted PSNR from 21.97 dB to 24.59 dB by resolving support mismatch; layering the SCC and spectral objective further pushed PSNR to 27.41 dB, demonstrating complementary spatial and frequency domain regularizations.
Superior throughput and low memory footprint: On an NVIDIA Tesla A100 GPU with $160 \times 192 \times 160$ volumes, SC-Flow requires only 28.5 GB VRAM (nearly half of voxel-space Causal-PETs at 52.8 GB) and executes inference in 0.304 seconds (a $>110\times$ speedup over multi-step diffusion IL-CLDM at 34.853 seconds). The lightweight SCC module contributes only 0.0954M parameters and 32.8 ms of latency.
Direct clinical diagnostic benefits: When evaluating synthetic volumes on Alzheimer's Disease classification (Paper Table 3), SC-Flow achieved 71.77% F1 and 81.25% AUC with ResNet50 on NACC-FDG (outperforming baseline MeanFlow's 43.29% F1 by +28.48%), proving that spectral fidelity translates directly into downstream pathological discrimination.

Highlights & Insights¶

Exact zero-noise single-step collapse: Ingeniously exploits Brownian bridge boundary conditions where the stochastic diffusion coefficient $\sigma(t)$ naturally vanishes at endpoints $t \in \{0, 1\}$, allowing rich stochastic training while maintaining deterministic, artifact-free 1-NFE inference.
Physical frequency prior against latent over-smoothing: Reconnects generative latent flow modeling with Wiener-Khinchin spectral theory, showing that high-frequency texture decay can be rectified via learnable radial gain modulation without costly iterative deblurring.
Unsupervised cross-task generalization: Eliminates handcrafted anatomical or tracer-specific priors, smoothly generalizing across diverse clinical tasks including functional-to-structural reverse translation (PET to MRI) and radiotherapy CT-to-CBCT synthesis.

Limitations & Future Work¶

Ceiling imposed by pre-trained autoencoder: Since all flow transport occurs in compressed latent space, subtle micro-calcifications or fine vascular networks lost during initial VAE encoding cannot be recovered by the spectral corrector.
Sensitivity to severe spatial misalignment: In cases with substantial motion artifacts or unregistered source-target pairs, high-frequency spectral gain amplification might inadvertently accentuate artifacts.
Future directions: Integrating spatiotemporal attention priors for progressive longitudinal disease monitoring and developing joint frequency-aware latent codebook representations.

vs MeanFlow [17]: MeanFlow pioneered average velocity modeling for 1-NFE generation but suffers from severe high-frequency attenuation in medical volumes. SC-Flow introduces Brownian bridge support expansion and radial spectral gain correction, cutting MAE by $>20\%$ and boosting PSNR by $>2.5$ dB.
vs Causal-PETs [38] / PASTA [37]: These specialized medical architectures depend on heavy 3D voxel operations or heuristic disease priors, requiring $>50$ GB VRAM. SC-Flow operates in latent space with a lightweight frequency module, achieving higher fidelity with half the memory footprint.
vs LBM (Latent Bridge Matching) [6]: LBM models latent trajectories but lacks explicit target spectral guidance, leading to modality entanglement. SC-Flow's spectral consistency constraints decouple domain-specific frequencies and enforce clean anatomical transfer.

Rating¶

Novelty: ⭐⭐⭐⭐☆ [Combines Brownian bridge stochastic support regularization with Wiener-Khinchin spectral energy compensation to resolve latent flow over-smoothing]
Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive validation across 4 benchmark datasets, 8 cross-modal setups, component ablations, frequency bin analyses, and downstream diagnostic tasks]
Writing Quality: ⭐⭐⭐⭐⭐ [Rigorous mathematical formulation, clear physical motivation, and transparent presentation of experimental evidence]
Value: ⭐⭐⭐⭐⭐ [Enables high-fidelity 3D medical synthesis in 0.3 seconds on a single GPU, offering immediate utility for clinical acceleration and radiation minimization]