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

Abstract: Magnetic Resonance Imaging (MRI) is often acquired with anisotropic resolution to reduce scan time, producing stair-step artifacts along the through-plane direction. In through-plane MRI super-resolution, an efficiency-fidelity trade-off arises: feed-forward regressors are fast but oversmooth at large slice-thicknesses, while sampling-based methods improve fidelity at high inference cost. We propose DRIFT, a two-stage thickness-conditioned rectified flow framework for through-plane MRI super-resolution with continuous input slice-thickness. Stage 1 employs an Anatomical Projection Network (APN) to map low-resolution patches to a coarse high-resolution manifold, providing a deterministic anatomical initialization that shortens the residual transport of Stage 2 and stabilizes slice-wise refinement. Stage 2 refines details via rectified flow and introduces a Physics-Aware Difficulty (PAD) metric derived from slice-thickness induced through-plane bandwidth deficit to guide an Adaptive Integration Scheduler (AIS), allocating ODE steps by thickness. A Consistent Endpoint Trajectory Alignment (CETA) loss enforces thickness-consistent reconstructions. Experiments show that DRIFT outperforms super-resolution baselines while reducing inference cost. Code, models, and interactive demos are available at https://yoonseokchoi-ai.github.io/drift-eccv2026/.

• The paper introduces DRIFT, a two-stage MRI super-resolution framework that combines deterministic anatomical projection with thickness-conditioned rectified-flow refinement and physics-aware adaptive sampling.
• DRIFT achieves state-of-the-art results across HCP, MIND, and IDEAS, including 37.64 dB PSNR and 0.952 SSIM on HCP at 4.2 mm, while improving arbitrary-scale reconstruction by up to 6.29 dB over prior methods.
• The method uses 8–13 adaptive integration steps instead of up to 2,000 for diffusion models, reaching approximately 20 seconds per volume and supporting zero-shot reduction of stair-step artifacts in real thick-slice MRI scans.

DRIFT (Difficulty-aware RectifIed Flow for Through-plane SR) addresses through-plane MRI super-resolution (SR), where anisotropic acquisition produces thick-slice volumes with stair-step artifacts along the slice-selection direction. The paper identifies an efficiency–fidelity trade-off in existing approaches: feed-forward regressors are fast but oversmooth at large slice thicknesses, while diffusion-based samplers improve fidelity at prohibitive inference cost. DRIFT resolves this tension by combining a deterministic anatomical projection stage with a rectified-flow refinement stage, both conditioned on continuous input slice thickness, and by allocating sampling effort according to a physics-derived difficulty metric rather than image-dependent heuristics (2607.16649).

Motivation and positioning

The authors argue that most prior through-plane SR methods assume discrete degradation scales. Self-supervised frameworks such as SMORE and SIMPLE exploit in-plane/through-plane resolution asymmetry; AFCM targets aliasing artifacts; SynthSR handles heterogeneous clinical protocols. Implicit neural representation (INR) methods—LIIF, LTE, ArSSR, SA-INR—offer arbitrary-scale flexibility but suffer from spectral bias and are typically trained under isotropic downsampling that does not reflect slice-profile-governed through-plane integration. Generative approaches such as TPDM achieve sharp textures but require thousands of sampling steps; ResShift reduces steps but uses a fixed scheduler regardless of degradation severity; AdaDiffSR adapts dynamically but relies on image-dependent perception modules detached from the physical degradation. DRIFT's central premise is that for MRI, the relevant difficulty signal is available for free in acquisition metadata: slice thickness directly determines through-plane bandwidth loss.

Method

Slice-profile simulation. Training pairs are synthesized from isotropic HR volumes using an SLR-derived slice-profile kernel applied along a randomly sampled axis, followed by non-integer resampling to the thick-slice grid and nearest-neighbor mapping back to the HR grid. This reproduces the stair-step morphology of reformatted thick-slice MRI and is shared with retrained baselines for fairness.

Thickness conditioning. Both stages are conditioned on inverse thickness $\tau = 1/T$ for input and target, embedded via unshared MLPs and fused into a 512-dimensional conditioning vector injected through Adaptive Group Normalization. The paper notes that target-continuous operation requires sampling target thicknesses during training, which is deferred to supplementary experiments; main experiments fix $T_{\mathrm{hr}}$ at each dataset's native resolution.

Stage 1: Anatomical Projection Network (APN). A deterministic regressor maps LR patches to a coarse HR manifold using Charbonnier plus SSIM loss ($\lambda_{\mathrm{ssim}}=0.5$). Beyond shortening the subsequent transport trajectory, APN provides spatially correlated initialization across neighboring slices—a deliberate contrast to stochastic noise initialization in slice-wise generative reconstruction, which can introduce inter-slice inconsistency.

Stage 2: Rectified flow refinement. With APN frozen, a velocity network learns straight-line transport between $\mathbf{z}_{p,i}$ and the HR target, trained with a Huber objective under an endpoint-biased U-shaped timestep distribution ($\alpha=2.0$) inspired by RF++. The output layer is zero-initialized so Stage 2 begins as an identity refinement.

CETA loss. The Consistent Endpoint Trajectory Alignment loss regularizes endpoint proxies $$\tilde{\mathbf{y}}_{p,k}(t) = \mathbf{z}_{p,k} + \mathbf{v}_{p,k}(t)$$ across proximal thickness pairs separated by $\Delta T = 1$\,mm, chaining local consistency constraints across the sampled thickness range.

PAD/AIS inference. The Physics-Aware Difficulty metric $\mathrm{PAD} = 1 - T_{\mathrm{hr}}/T_i$ quantifies normalized through-plane bandwidth deficit purely from metadata. The Adaptive Integration Scheduler maps PAD to an Euler step budget clamped to $[0, 15]$ NFEs, requiring no auxiliary image-difficulty network and negligible overhead.

Quantitative results

On HCP, MIND, and IDEAS with subject-wise splits, DRIFT achieves the best PSNR/SSIM in all fixed-scale and arbitrary-scale settings evaluated:

Setting	Dataset / scale	Best baseline	DRIFT
Fixed	HCP ×6 (4.2 mm)	SwinIR: 34.10 dB	37.64 dB / 0.952
Fixed	HCP ×8 (5.6 mm)	SwinIR: 33.13 dB	35.93 dB / 0.936
Arbitrary	HCP ×8.57 (6.0 mm)	LTE: 26.68 dB	32.97 dB / 0.891
Arbitrary	MIND ×6.67 (6.0 mm)	SA-INR: 26.22 dB	29.51 dB / 0.882

Gains over the strongest fixed-scale baseline reach 3.54 dB on HCP, 2.59 dB on MIND, and 1.11 dB on IDEAS. Against arbitrary-scale baselines—which were largely designed for isotropic downsampling—the margins are larger: 3.29–6.29 dB at 6.0 mm. Notably, DRIFT remains robust at the out-of-distribution 6.5 mm thickness where INR methods degrade under stair-step artifacts. These results imply that acquisition-physics-aware training data and conditioning, rather than architectural capacity alone, drive the improvement over INR paradigms.

Zero-shot transfer and efficiency

Without retraining, DRIFT applied to IRB-approved in-house thick-slice scans (5 mm FLAIR and T2w) reduces stair-step artifacts and preserves anatomical continuity; TPDM shows weak cross-dataset generalization on these scans. On fastMRI brain T2w volumes against the foundation model BME-X, DRIFT yields higher sharpness (0.291 vs. 0.238) and substantially lower no-reference quality indices (NIQE 5.30 vs. 7.22; BRISQUE 21.16 vs. 62.46). Because these real-data evaluations lack isotropic ground truth, they rest on qualitative and no-reference evidence rather than full-reference metrics—an assumption the authors state plainly.

On the efficiency side, DRIFT attains the lowest LPIPS (0.043) on HCP ×8 at roughly 20 s per volume—3.3× better LPIPS than ArSSR while being 14× faster—and outperforms TPDM despite TPDM being ~450× slower. In NFE terms, DRIFT uses 8–13 steps adaptively versus ResShift's fixed 15 and TPDM's 2,000, achieving higher PSNR. A compact DRIFT-Small variant reduces parameters, FLOPs, and peak VRAM by 3.3×, 4.0×, and 5.3× respectively while remaining competitive, indicating the gains are not explained by model size.

Ablations

Component ablations on HCP (averaged over input thicknesses of 1.5, 3.0, and 5.0 mm) show that removing thickness conditioning causes the largest drop (40.85 → 34.82 dB PSNR), followed by removing CETA (→ 38.23 dB) and U-shaped timestep sampling (→ 38.12 dB). For the CETA gap, random pairing improves over no CETA but underperforms the fixed 1.0 mm gap (38.23 vs. 40.85 dB); gaps of 0.5 mm provide near-identity constraints while 4.0 mm reduce coverage to 25% and destabilize optimization. An AdaDiffSR-style image-dependent difficulty score shows negligible correlation with HF-PSNR within fixed-thickness groups on multi-vendor CC359, supporting metadata-based scheduling in this setting. Two-stage ablations confirm that Stage 1 alone oversmooths and Stage 2 alone lacks stable initialization at large thicknesses.

Limitations and open questions

The paper concedes that main experiments fix the target thickness at each dataset's native isotropic resolution; extending training to sampled target thicknesses for fully target-continuous operation is examined only in supplementary material. The zero-shot clinical evaluations rely on qualitative and no-reference metrics due to absent ground truth. The PAD metric assumes bandwidth deficit is the dominant difficulty factor; whether it suffices for pathologies or sequences where image content modifies effective difficulty is not established. The CETA gap of 1 mm is presented as an empirical balance rather than a principled choice, and its interaction with other clinical thickness ranges remains open.

Conclusion

DRIFT formulates through-plane MRI SR as a two-stage projection-to-transport problem conditioned on continuous slice thickness, with metadata-driven adaptive ODE scheduling and proximal-thickness trajectory regularization. It delivers consistent state-of-the-art fidelity across public datasets, strong zero-shot behavior on real thick-slice acquisitions, and order-of-magnitude inference savings relative to diffusion baselines, while leaving target-thickness generalization and broader clinical validation as open questions.