DRIFT: Difficulty-aware Rectified Flows for Through-plane MRI Super-Resolution

Conference: ECCV 2026
Paper: ECCV 2026 Poster 3862
Code: https://yoonseokchoi-ai.github.io/drift-eccv2026/
Area: Medical Imaging
Keywords: MRI Super-Resolution, Rectified Flow, Through-plane Super-Resolution, Physics-aware Scheduling, Manifold Projection

TL;DR

To tackle stair-step artifacts and continuous slice-thickness degradation in anisotropic MRI, DRIFT introduces a two-stage framework combining an Anatomical Projection Network with residual rectified flow, physics-aware difficulty (PAD) step scheduling, and consistent endpoint trajectory alignment (CETA), delivering high-fidelity isotropic reconstructions with substantially reduced function evaluations.

Background & Motivation

High-resolution isotropic Magnetic Resonance Imaging (MRI) is essential for neuroanatomical volumetry and clinical diagnostic accuracy. Nonetheless, intrinsic imaging trade-offs governed by signal-to-noise ratio (SNR), scan duration, and specific absorption rate constraints often force clinical routines to adopt anisotropic acquisition protocols. These scans capture fine in-plane details but feature coarse through-plane slice spacing, generating pronounced stair-step artifacts when 3D volumes are reformatted along non-acquisition planes.

Standard feed-forward super-resolution architectures are typically trained on discrete, fixed magnification factors, which severely restricts generalization across the continuous slice-thickness spectrum encountered across clinical scanner protocols. When presented with severe thick-slice degradations, pixel-wise regression baselines over-penalize structural uncertainties and revert to overly smooth anatomical boundaries. Conversely, coordinate-based Implicit Neural Representations (INRs) support continuous-scale query coordinates but suffer from spectral bias, inherently attenuating the high-frequency structural textures essential for clinical assessment while relying on unrealistic isotropic downsampling assumptions. Multi-step diffusion models can synthesize realistic high-frequency textures, but their high computational cost and black-box perceptual difficulty modules impose heavy latency and memory burdens.

The angle of attack in DRIFT stems from physical imaging principles: through-plane information degradation is fundamentally governed by RF slice-selection profile integration, meaning the missing bandwidth deficit can be directly derived from acquisition metadata without auxiliary neural networks. Core idea: formulate through-plane MRI super-resolution as a two-stage manifold projection-to-transport process, where an Anatomical Projection Network deterministically maps inputs onto a coarse high-resolution manifold to shorten transport, and acquisition metadata directly drives Physics-Aware Difficulty (PAD) ODE step scheduling regularized by Consistent Endpoint Trajectory Alignment (CETA).

Method

Overall Architecture

DRIFT processes anisotropic scans via a 2D slice-wise pipeline: for each input low-resolution slice \(\mathbf{x}_{p,i}\) acquired with slice thickness \(T_i\), the system reconstructs an isotropic high-resolution slice at target thickness \(T_{\mathrm{hr}}\). Slices are processed orthogonal to the degraded through-plane axis and composed back into an isotropic 3D volume.

In Stage 1, an Anatomical Projection Network (APN) conditions on the inverse slice thickness \(\tau_i = 1/T_i\) and projects the thick-slice input onto a coarse high-resolution anatomical manifold \(\mathbf{z}_{p,i}\), resolving stair-step artifacts and establishing a deterministic reference state. In Stage 2, the APN is frozen, and a Rectified Flow velocity network learns a straight-line vector field from the coarse estimate to the ground-truth high-resolution manifold. During inference, the ODE step count is dynamically determined via Physics-Aware Difficulty computed directly from slice-thickness metadata, dispensing with auxiliary perceptual predictor networks.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Anisotropic Input Slice x and Inverse Thickness τ"] --> B["Anatomical Manifold Projection<br/>APN deterministic mapping to coarse HR z"]
    B --> C["Physics-Aware Difficulty Assessment<br/>Metadata-derived bandwidth deficit PAD"]
    C --> D["Adaptive Integration Scheduler AIS<br/>Dynamic ODE step budget allocation N"]
    D --> E["Rectified Flow Residual Refinement<br/>Straight-line velocity field synthesizes fine texture"]
    E --> F["Output Isotropic Slices assembled into 3D Volume"]

Key Designs

1. Anatomical Manifold Projection: Deterministic Initialization to Shorten Flow Transport Standard generative super-resolution frameworks initiate reverse sampling trajectories from uninformative Gaussian noise. In 2D slice-wise medical volume processing, independent noise draws introduce cross-slice stochastic variance, resulting in high-frequency banding artifacts along reformatted orthogonal planes. DRIFT designs the Anatomical Projection Network (APN) \(f_\phi\), trained with Shinnar-Le Roux (SLR) RF-pulse degradation simulation. Conditioning on input slice \(\mathbf{x}_{p,i}\) and inverse-thickness embedding vector \(\mathbf{c}_i\), the APN outputs a deterministic coarse estimate \(\mathbf{z}_{p,i} = f_\phi(\mathbf{x}_{p,i}, \mathbf{c}_i)\). This projection anchors the intermediate state close to the ground-truth high-resolution distribution, eliminating inter-slice stochastic jitter and converting the generative task into a localized high-frequency residual transport that requires far fewer integration steps.

2. Physics-Aware Adaptive Inference: Zero-Overhead Metadata-Driven Step Scheduling Existing adaptive diffusion acceleration schemes incorporate auxiliary image-difficulty estimators, incurring noticeable runtime and memory overhead. DRIFT instead leverages the physics of MRI slice excitation, where through-plane information attenuation corresponds to frequency truncation determined by slice thickness. DRIFT quantifies degradation difficulty via the Physics-Aware Difficulty (PAD) metric, defined as the normalized through-plane bandwidth deficit: $\(\mathrm{PAD}(T_i, T_{\mathrm{hr}}) = 1 - \frac{T_{\mathrm{hr}}}{T_i}\)$ The Adaptive Integration Scheduler (AIS) maps this physical parameter directly to the ODE step budget: \(N = \mathrm{clamp}\left(\lfloor N_{\max} \cdot \mathrm{PAD}(T_i, T_{\mathrm{hr}}) \rceil, N_{\min}, N_{\max}\right)\), with \(N_{\min}=0, N_{\max}=15\). For mildly degraded thin-slice scans, AIS allocates minimal NFEs (e.g., 4–8 steps) or utilizes the APN output directly; for thick-slice scans with severe bandwidth deficits, AIS increases the step budget to restore delicate microstructures. Because PAD relies solely on scan metadata, scheduling incurs virtually zero computational overhead.

3. CETA Loss and U-shaped Timestep Sampling: Local Trajectory Regularization To ensure consistent reconstructions across continuous slice thicknesses, DRIFT introduces Consistent Endpoint Trajectory Alignment (CETA). During Stage 2 training, proximal thickness pairs \(T_j = T_i - \Delta T\) (with fixed interval \(\Delta T = 1.0\text{ mm}\)) are synthesized from the same underlying high-resolution anatomy. DRIFT computes endpoint proxies \(\tilde{\mathbf{y}}_{p,k}(t) = \mathbf{z}_{p,k} + \mathbf{v}_{p,k}(t)\) for \(k \in \{i, j\}\) and minimizes their discrepancy: $\(\mathcal{L}_{\mathrm{CETA}} = \left\|\tilde{\mathbf{y}}_{p,i}(t) - \tilde{\mathbf{y}}_{p,j}(t)\right\|_2^2\)$ This formulation constructs a chain of localized constraints across the continuous thickness manifold, avoiding optimization instability seen in random-thickness pairing. Furthermore, recognizing that residual transport is most sensitive near boundary trajectories, Stage 2 adopts a U-shaped timestep distribution \(p_\alpha(t)\) (\(\alpha=2.0\)) that concentrates training loss at \(t \approx 0\) and \(t \approx 1\). Combined with zero-initialized output layers, this stabilizes velocity field training and enforces smooth structural transitions.

Loss & Training

DRIFT follows a decoupled two-stage training scheme: - Stage 1 (APN): Trained via Charbonnier loss combined with structural similarity (SSIM): $\(\mathcal{L}_{\mathrm{recon}} = \mathcal{L}_{\mathrm{Char}}(\mathbf{z}_{p,i}, \mathbf{y}_p) + \lambda_{\mathrm{ssim}}\mathcal{L}_{\mathrm{SSIM}}(\mathbf{z}_{p,i}, \mathbf{y}_p)\)$ with \(\lambda_{\mathrm{ssim}} = 0.5\). - Stage 2 (Rectified Flow): With the APN frozen, the velocity network \(v_\theta\) learns straight-line flow targets \(\mathbf{u}_{p,i} = \mathbf{y}_p - \mathbf{z}_{p,i}\) using a Huber loss (\(\delta=0.1\)) under the U-shaped sampling distribution \(t \sim p_\alpha(t)\), regularized by CETA: $\(\mathcal{L}_{\mathrm{Stage2}} = \mathbb{E}_{t \sim p_\alpha(t)}\left[\mathrm{Huber}_\delta(\mathbf{v}_{p,i}(t), \mathbf{u}_{p,i})\right] + \lambda_{\mathrm{ceta}}\mathcal{L}_{\mathrm{CETA}}\)$ with \(\lambda_{\mathrm{ceta}} = 1.0\).

Key Experimental Results

Main Results

Evaluated across three public benchmark datasets with isotropic ground truth (HCP, MIND, and IDEAS), DRIFT consistently outperforms state-of-the-art fixed-scale regressors, coordinate-based INRs, and diffusion models under both in-distribution fixed scales and continuous out-of-distribution slice thicknesses up to \(6.5\text{ mm}\).

Performance comparison from Table 1 of the paper (PSNR [dB] / SSIM):

Dataset	Scale / Input Thickness	SwinIR [16]	AFCM [25]	ResShift [36]	TPDM [14]	DRIFT (Ours)
HCP	\(\times 6\) (\(4.2\text{ mm}\))	34.10 / 0.901	27.11 / 0.642	32.89 / 0.888	29.31 / 0.798	37.64 / 0.952
HCP	\(\times 8\) (\(5.6\text{ mm}\))	33.13 / 0.882	26.08 / 0.501	30.34 / 0.853	28.24 / 0.755	35.93 / 0.936
MIND	\(\times 5\) (\(4.5\text{ mm}\))	30.15 / 0.874	27.74 / 0.728	29.48 / 0.839	22.50 / 0.452	32.74 / 0.904
MIND	\(\times 6\) (\(5.4\text{ mm}\))	29.01 / 0.838	26.02 / 0.713	29.00 / 0.833	22.11 / 0.413	30.02 / 0.886
IDEAS	\(\times 5\) (\(5.0\text{ mm}\))	31.64 / 0.924	23.89 / 0.825	31.93 / 0.910	26.64 / 0.744	32.75 / 0.946
IDEAS	\(\times 6\) (\(6.0\text{ mm}\))	31.62 / 0.914	22.55 / 0.796	30.40 / 0.899	26.11 / 0.720	32.01 / 0.938

Under continuous arbitrary-scale evaluations (e.g., \(6.0\text{ mm}\) and out-of-distribution \(6.5\text{ mm}\)), DRIFT outperforms the strongest arbitrary-scale baselines (LTE, SA-INR, ArSSR) by \(3.29\sim 6.29\text{ dB}\) PSNR, maintaining anatomical stability where coordinate MLPs suffer from severe spectral blurring.

Ablation Study

Ablations on Stage 2 core components and the CETA proximal gap \(\Delta T\) are reported in Table 2 and Table 3 on HCP (averaged over \(T_i \in \{1.5, 3.0, 5.0\}\text{ mm}\) with \(T_{\mathrm{hr}} = 0.7\text{ mm}\)):

Stage 2 Component Ablation (Table 2 in original paper):

Thickness Cond.	U-shaped Sampling	CETA Loss (\(\lambda_{\mathrm{CETA}}=1\))	PSNR (dB) ↑	SSIM ↑	LPIPS ↓
✓	✓	✗	36.92	0.9416	0.0308
✗	✓	N/A	34.82	0.8986	0.0491
✓	✗	✓	38.12	0.9458	0.0342
✓	✓	✓	40.85	0.9634	0.0229

CETA Proximal Gap \(\Delta T\) Ablation (Table 3 in original paper):

Configuration / Gap \(\Delta T\)	Coverage	PSNR (dB) ↑	SSIM ↑	LPIPS ↓	Note
Random Pairs	65%	38.23	0.9460	0.0290	Lacks localized neighborhood topology
\(0.5\text{ mm}\)	91%	37.24	0.9430	0.0301	Overly narrow gap acts as identity mapping
\(1.0\text{ mm}\) (Ours)	81%	40.85	0.9634	0.0229	Optimal balance of constraint strength & coverage
\(2.0\text{ mm}\)	62%	39.10	0.9540	0.0265	Wider gap increases optimization tension
\(4.0\text{ mm}\)	25%	37.58	0.9445	0.0338	Dissimilar pairs destabilize learning

Key Findings

• Acquisition conditioning is paramount: Omitting slice-thickness conditioning causes PSNR to drop sharply from 40.85 dB to 34.82 dB, demonstrating that explicitly conditioning on physical slice parameters is essential for multi-thickness SR.
• Proximal pairing outmatches random regularization: Fixed proximal spacing at \(\Delta T = 1.0\text{ mm}\) surpasses naive random pair sampling by 2.62 dB in PSNR and achieves a superior LPIPS of 0.0229, validating the benefit of chained local manifold constraints.
• Unprecedented trade-off between throughput and perceptual fidelity: On HCP \(\times 8\), DRIFT reconstructs an entire brain volume in \(\sim 20\) seconds with an LPIPS of 0.043. It is \(14\times\) faster than 3D INR model ArSSR (0.112 LPIPS, \(\sim 280\) s) and \(450\times\) faster than dual-diffusion baseline TPDM (0.260 LPIPS, \(\sim 9000\) s).

Highlights & Insights

• Domain physics replaces auxiliary networks: Quantifying degradation through RF pulse slice integration allows ODE integration budgets to be derived directly from DICOM header metadata without auxiliary network evaluation.
• Deterministic projection anchors residual flow: The APN absorbs low-frequency recovery and stair-step artifact removal deterministically, allowing Rectified Flow to focus exclusively on localized residual textures, preventing slice-to-slice stochastic discrepancies.
• Local proximal alignment ensures manifold continuity: CETA's 1.0 mm proximal gap connects neighboring thickness conditions, preventing trajectory drift across the continuous resolution spectrum.

Limitations & Future Work

• Target resolution fixed to native isotropic spacing in primary benchmarks: While the architecture admits target-thickness embeddings, primary experiments held target resolution fixed to native scan spacing; extending continuous target sampling across heterogeneous protocols remains an active direction.
• Validation focused on brain neuroimaging: The methodology has been validated primarily on cerebral scans (T1w, T2w, FLAIR); extending the framework to abdominal and cardiac acquisitions with respiratory and cardiac motion requires additional motion-compensated mechanisms.

Related Work & Insights

• vs TPDM [14]: TPDM iterates over orthogonal 2D diffusion models requiring 2,000 sampling steps per volume and frequently generates hallucinated noise patterns; DRIFT achieves superior anatomical fidelity within 8–13 deterministic steps.
• vs ResShift [36]: ResShift reduces sampling steps via residual shifting but enforces a static 15-step schedule; DRIFT modulates step count dynamically via acquisition-derived PAD, accelerating thin-slice processing.
• vs ArSSR / SA-INR [32, 34]: INR architectures suffer from spectral bias that attenuates fine cortical boundaries; DRIFT leverages flow-based generative priors to restore crisp tissue contrast and fine gray/white matter boundaries.

Rating

• Novelty: ⭐⭐⭐⭐ [Principled fusion of MRI RF slice excitation physics, residual rectified flow, and zero-overhead metadata-guided scheduling]
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Evaluated across three gold-standard isotropic benchmarks, zero-shot real thick-slice clinical scans, extensive component ablations, and 3D reformat analyses]
• Writing Quality: ⭐⭐⭐⭐⭐ [Clear structural exposition, rigorous mathematical and physical grounding, and transparent empirical trade-off evaluations]
• Value: ⭐⭐⭐⭐⭐ [Directly resolves the throughput-fidelity bottleneck in volumetric reformats of rapid clinical thick-slice MRI scans]

Related Papers

• [ECCV 2026] MArFE: Multi-Contrast MRI Arbitrary Scale Super-Resolution with Fourier Enhancement
• [ECCV 2026] PyraE2E: Enhancing End-to-End WSI Analysis via Cross-Scale Super-Resolution
• [AAAI 2026] PINGS-X: Physics-Informed Normalized Gaussian Splatting with Axes Alignment for Efficient Super-Resolution of 4D Flow MRI
• [ECCV 2026] Dual-Prior Guided Null-Space Learning with Mixture-of-Splines for Arbitrary Medical Slice Super-Resolution
• [CVPR 2026] Adaptive Anisotropic Gaussian Splatting for Multi-contrast MRI Arbitrary-Scale Super-Resolution with Anatomy Guidance