PaperlayerView exact revision on arXiv

Source-linked AI summary

PredRA: Fast Medical Image Translation by Deterministic Component Extraction and Controlled Stochastic Refinement

Jianhai Zhang, Pattarawut Charatpangoon, Donghao Zhang, Bijoy K. Menon, Wu Qiu, M. Ethan MacDonald, Aravind Ganesh

arXiv:2609.31912v1cs.CV

TL;DR

Strongly paired medical translation must balance deterministic predictions that can smooth detail against generative models that can add unnecessary variation. PredRA keeps a deterministic reference and globally controls how much generated residual refinement is added. Across three translation tasks, this improves paired fidelity over matched full residual proposals while retaining much of their perceptual benefit.

  • Problem

    In strongly paired medical translation, deterministic prediction can smooth fine detail, whereas generative refinement can introduce unnecessary variation.

    Source: §I · p. 1

  • Method

    PredRA keeps a deterministic prediction as reference, models the remaining residual, and globally controls the amount of refinement accepted.

    Source: §I · p. 2, §III · p. 3

  • Results

    Across CT+CTA→CTP, CT+CTA→DWI+ADC, and SynthRAD CT→MRI, controlled refinement improves paired fidelity over matched full residual proposals while keeping much of their perceptual benefit.

    Source: §VIII · p. 11

  • Takeaways & Limitations

    The practical conclusion is to predict the source-supported part first, then add only the residual refinement validation shows is useful.

    Source: §VIII · p. 11

    The stable-versus-remaining residual decomposition is an interpretation, not a calibrated uncertainty model; with one paired target per input, model error cannot be cleanly separated from true conditional variability.

    Source: §VII · p. 10

Abstract

from arXiv · show

Strongly paired medical image translation contains a substantial component that can be predicted directly from the source. The current reality is that pure deterministic prediction can smooth away fine detail, while generative models can recover detail but may also introduce unnecessary or potentially harmful variation. We propose PredRA, a fast framework that uses the deterministic prediction as a stable reference and further extracts additional deterministic components from the residual during generative refinement, thereby improving fidelity while maintaining perceptual quality. The goal is simple: we view the entire residual-based generative process as an optimization problem and derive a practical solution that recovers useful residual detail through controlled refinement while keeping the prediction close to the paired target. Mechanism studies further show that useful residual information follows structured patterns, but its usefulness is difficult to estimate reliably at the voxel level. We therefore globally control how much of the residual refinement is added to the deterministic prediction, thereby reducing the accumulation of unnecessary uncertainty. PredRA therefore combines deterministic component extraction with controlled stochastic refinement for fast, fidelity-preserving, and perceptually strong medical image translation. We validate the approach across multiple real-world datasets, showing that controlled refinement improves fidelity to the paired target compared with full residual refinement while retaining much of the perceptual benefit of generative modeling. PredRA achieves competitive or superior performance to substantially larger state-of-the-art models with 1.4-11.9x fewer total parameters and 3.1-26.2x fewer trainable parameters, while its 32-step flow sampler requires 31.25x fewer sampling steps than the matched 1000-step DDPM.

I. INTRODUCTION

In strongly paired medical translation, deterministic models preserve source-supported structure but can smooth detail, while generation can recover detail and add unnecessary variation. PredRA keeps a deterministic reference, generates a residual correction, and selects its strength on validation; it tests this approach with conditional DDPM and Flow.

I. INTRODUCTION

  • Strongly paired translation creates a fidelity–perceptual-quality trade-off: deterministic models can smooth fine detail, while stronger generation can introduce unnecessary variation.PredRA treats the deterministic prediction as a stable reference and adds only validation-supported residual refinement.Source: §I · p. 1
  • PredRA reuses one residual proposal across candidate strengths and selects the amount of refinement on validation to favor paired fidelity while retaining perceptual benefit.The same proposal ensures the comparison changes only how much refinement is accepted.Source: §I · p. 1
  • The study evaluates deterministic, full-residual, and validation-selected predictions against published translation baselines across multiple real-world medical imaging datasets.Matched Full-versus-PredRA comparisons test whether controlled use of the same residual proposal improves the final prediction.Source: §I · pp. 1–2

A. Deterministic, Adversarial, and State-Space Translation

C. Deterministic Priors, Residual Diffusion, and Frequency Decomposition

Residual diffusion and deterministic priors are established, as is the perception–distortion trade-off. PredRA’s framing is that residuals can retain source-supported structure after the deterministic prediction, motivating a global refinement controller rather than voxel-wise separation.

C. Deterministic Priors, Residual Diffusion, and Frequency Decomposition

III. DETERMINISTIC COMPONENT EXTRACTION AND RESIDUAL STRUCTURE

  • The residual decomposition treats bθ(X) as source-supported structure still predictable after an imperfect deterministic model, while εθ is not assumed to be pure acquisition noise or aleatoric uncertainty.The remaining deterministic risk measures paired target error before generation, not calibrated posterior variance.Source: §III · p. 3

D. Fidelity, Uncertainty, and Controlled Stochasticity

  • PredRA uses residual predictability and spatial-structure analyses to motivate a global controller, but these analyses do not separate irreducible aleatoric uncertainty from other variation.The captioned overview describes a validation-selected λ that maximizes SSIM subject to an LPIPS constraint.Source: §III · p. 3, Figure 1 · p. 3

IV. PREDRA: RECOVERING STABLE RESIDUAL DETAIL WITH CONTROLLED REFINEMENT

PredRA keeps the deterministic prediction fixed and models the remaining correction with a residual generator. Its controller chooses how much of one generated proposal to accept, balancing validation fidelity against the perceptual benefit of generation.

A. Residual Refinement as an Optimization Problem

  • The residual proposal can contain both useful stable structure and sampling-dependent variation, and scaling the same realization by λ controls how much correction—and pre-projection sampling variability—is added.This scaling is not presented as calibrated predictive uncertainty.Source: A. Residual Refinement as an… · p. 4
  • The controller maximizes validation fidelity subject to limiting lost perceptual benefit; in the experiments, these are SSIM and an LPIPS-based constraint.The residual model learns the remaining correction, including stable structure missed by the first model.Source: A. Residual Refinement as an… · p. 4

B. Conditional Residual DDPM

  • The conditional DDPM models the residual target after the deterministic reference is fixed, using conditioning C = (X, Ŷdet) for its noise predictor.At inference, ancestral reverse sampling from Gaussian noise produces one realized residual proposal.Source: B. Conditional Residual DDPM · p. 4

C. Conditional Residual Flow Matching

Conditional flow matching replaces the DDPM reverse chain as a second residual-generation backend, using an ODE sampler. The refinement strength is then selected on validation to maximize SSIM while retaining the LPIPS benefit within a prespecified slack.

C. Conditional Residual Flow Matching

  • Conditional flow matching trains a vector field along a straight conditional path and uses ODE integration to generate the residual proposal.Evaluation uses explicit Euler steps of 1, 2, 4, 8, 16, 32, or 64; matched-width comparisons retain the same residual U-Net family.Source: C. Conditional Residual Flow Matching · p. 4

D. Validation-Guided Refinement Strength

  • The frozen controller reuses the same residual proposal across λ values, with λ = 0 giving the deterministic reference and λ = 1 the full proposal.It selects validation SSIM subject to an LPIPS slack of δ = 0.0050; PSNR and MAE are fixed tie-breakers.Source: D. Validation-Guided Refinement Strength · p. 4

V. EXPERIMENTAL DESIGN

The experiments cover three paired translation settings, compare PredRA with matched residual-generation and external baselines, and test whether residual structure can be usefully controlled at global or spatial scales.

A. Translation Tasks

  • The three settings are CT+CTA→CTP, CT+CTA→DWI+ADC, and reverse-direction SynthRAD CT→MR, each evaluated on frozen patient-level test data.The CTP split is 55/12/12 patients, the DWI+ADC split is 70/10/20, and SynthRAD uses 126/27/27 of 180 paired cases.Source: A. Translation Tasks · p. 5

B. Comparators and Matched Controls

  • Matched DET, full-residual, and validation-selected PredRA outputs test controlled refinement alongside published baselines under shared evaluators.SynthRAD uses official or protocol-adapted reproductions; post-inspection capacity and Flow-step sweeps are treated as post hoc sensitivity analyses.Source: B. Comparators and Matched Controls · p. 5

C. Mechanistic Validation Protocols

  • Spatial gates test whether residual utility varies across locations, while oracle-informed and deployable gates distinguish spatial heterogeneity from its predictability.The analysis also compares target-consistent intermediate states with autonomous reverse trajectories; final decisions rely on autonomous outputs and validation-level allocation.Source: C. Mechanistic Validation Protocols · pp. 5–6

D. Metrics and Aggregation

Across the tested tasks, controlled refinement improves paired fidelity over full residual proposals, while mechanism analyses show that residual structure and local denoising gains do not always translate directly to autonomous outputs.

D. Metrics and Aggregation

  • Metrics are aggregated from slice to patient to cohort, and external-model predictions are restored to the canonical 192×192 field before evaluation.Padding can distort full-image SSIM: cropping an I2I-Mamba canvas changed SSIM by −0.1726 and LPIPS by +0.0553.Source: D. Metrics and Aggregation · p. 6

A. Stable Structure Remaining in the Residual

  • Held-out residual magnitude had moderate predictability (Spearman ρ ≈ +0.4700), supporting structured residual information but not reliable voxel-wise control.The authors therefore use one global validation-selected refinement strength.Source: A. Stable Structure Remaining in the… · p. 6

B. Local Denoising Improvements Do Not Necessarily Survive Autonomous Sampling

  • An intervention improved SSIM by +0.0144 on a target-consistent intermediate state but changed it by −0.0003 on the autonomous reverse trajectory.The sign reversal motivates evaluating residual changes on free-running trajectories rather than oracle intermediate states.Source: B. Local Denoising Improvements Do Not… · p. 6

C. Frozen-Test Results Across Tasks

  • On CTP, all three paired fidelity metrics improve for both matched backends, while LPIPS rises slightly from 0.1193 to 0.1210 for DDPM and 0.1140 to 0.1166 for Flow.These test LPIPS changes are descriptive; the 0.0050 constraint is applied only on validation.Source: C. Frozen-Test Results Across Tasks · p. 6, Table II · p. 7
  • PredRA-DDPM-12.63M lowers DWI+ADC macro MAE from 0.1951 to 0.1633, raises SSIM from 0.7776 to 0.8061, and improves PSNR and LPIPS versus Full-DDPM-12.63M.PredRA-Flow32-12.63M also improves all four macro metrics over Full-Flow32-12.63M; Table I reports the frozen-test benchmark.Source: C. Frozen-Test Results Across Tasks · pp. 6–7, Table I · p. 7
  • On SynthRAD, PredRA-Flow32-12.63M improves all four metrics over SelfRDB, I2I-Mamba, and ResViT, while MADM remains stronger in MAE, SSIM, and PSNR.PredRA-Flow32-12.63M stays close to DET-15.25M in paired fidelity while lowering LPIPS by 0.0287 (15.7000%).Source: C. Frozen-Test Results Across Tasks · p. 7, Table III · p. 7

D. SynthRAD Deterministic and Residual-Backend Capacity Analyses

The SynthRAD analyses examine deterministic-anchor and residual-backend capacity, flow sampling steps, and spatial refinement. They find a non-monotonic capacity response, a fidelity–perception trade-off across sampling steps, and weak test-time voxel-level utility prediction.

D. SynthRAD Deterministic and Residual-Backend Capacity Analyses

  • DET-15.25M (15.254M parameters) gives the best validation SSIM and MAE among tested deterministic widths; larger models do not improve consistently.The study fixes this anchor before residual-backend scaling.Source: D. SynthRAD Deterministic and… · p. 7
  • With DET-15.25M fixed, both residual backends improve overall from 1.547M to 12.628M parameters, but the trend is non-monotonic.The 8.851M Flow point is a local regression.Source: D. SynthRAD Deterministic and… · p. 7, Table V · p. 8
  • At 12.628M residual parameters, full Flow improves all four metrics over full DDPM, so allocated Flow’s small advantage is not solely due to different selected λ values.The full-refinement comparison uses λ = 1.Source: D. SynthRAD Deterministic and… · p. 8

E. Flow Sampling-Step Analysis

  • One-step flow gives the strongest paired-fidelity metrics, while more integration steps improve LPIPS as MAE, SSIM, and PSNR decline.In the post hoc 1.55M sweep, 32 steps is near a late-stage trade-off; 64 steps adds little LPIPS benefit at measurable fidelity cost.Source: E. Flow Sampling-Step Analysis · pp. 8–9

F. Spatial Refinement Analysis

  • 0.1100 test Spearman correlation was the best reported within-trajectory utility-ranking result, and neither spatial-gating approach beat global selected-strength control.The spatial utility experiments are retained as a negative mechanism result, not part of the final controller.Source: F. Spatial Refinement Analysis · p. 9

G. Qualitative Results

PredRA keeps a deterministic prediction as a reference and adds only validation-supported residual refinement. Across the reported tasks, this improves paired fidelity over full residual refinement while retaining much of its perceptual benefit.

G. Qualitative Results

  • Figures 2–3 contrast smooth, source-structured deterministic outputs with stronger local detail from full residual proposals; PredRA adds only validation-supported refinement.Figure 3 illustrates this balance in one Tmax case; quantitative comparisons use complete frozen cohorts.Source: G. Qualitative Results · p. 9

VII. DISCUSSION

  • 32 residual-network evaluations versus 1000 ancestral DDPM steps define the Flow32 fast variant, but the efficiency table reports per-network evaluation time rather than full sampling wall time.The reported configurations combine a frozen deterministic anchor with a residual network.Source: Table VII · p. 10
  • On CTP, PredRA improves MAE, SSIM, and PSNR over matched full proposals; on DWI+ADC, it improves all four macro metrics for both matched backends.On CTP, no method leads on every metric: YODA is strongest on several distortion metrics, while Full-Flow32-12.63M has the lowest LPIPS overall.Source: §VII · p. 10
  • SynthRAD results show that both DDPM and conditional Flow benefit from controlled refinement, while larger residual models still benefit from controlling their output.Deterministic performance eventually saturates as model capacity grows.Source: §VII · p. 10
  • The number of sampling steps changes the residual proposal, whereas λ controls how much of one fixed proposal is accepted; more steps may improve perceptual similarity while reducing paired fidelity.The relation Upost(λ) ∝ λ^2 explains why stronger residual scaling admits more sampling variability.Source: §VII · p. 10
  • Residual usefulness varies spatially, but tested source- and trajectory-based features do not predict it well enough to improve outputs, making global validation control more reliable in these experiments.Some interventions that appear useful on oracle intermediate states lose their benefit during free-running generation.Source: §VII · p. 10
  • The DWI+ADC task is registered paired stroke data rather than an external-site cohort.The discussion also notes that some external baseline schedules are compute-limited.Source: §VII · p. 10

VIII. CONCLUSION

  • Across CT+CTA→CTP, CT+CTA→DWI+ADC, and SynthRAD CT→MRI, controlled refinement improves paired fidelity over matched full residual proposals while retaining much of their perceptual benefit.Both DDPM and conditional Flow support the approach, and Flow provides the fast variant.Source: §VIII · p. 11

APPENDIX A MATHEMATICAL DETAILS OF DETERMINISTIC ANCHORING AND CONTROLLED REFINEMENT

The appendix uses squared-error risk and pre-projection sampling variance to explain deterministic anchoring and partial residual refinement. These derivations motivate the method but do not assert that the trained anchor is the exact conditional mean or that residuals are calibrated uncertainty.

APPENDIX A MATHEMATICAL DETAILS OF DETERMINISTIC ANCHORING AND CONTROLLED REFINEMENT

  • The appendix’s derivations use squared Euclidean risk or pre-projection sampling variance with the same residual proposal scaled across candidate λ values; validation still selects the controller.The actual controller is selected by validation SSIM under an LPIPS constraint.Source: Appendix A · p. 11

A. Squared-Error Predictable Component

  • Under squared error, the conditional-mean predictor removes the predictable first moment, but the trained finite-network anchor is not asserted to equal that mean or leave only aleatoric residuals.The Bayes residual has zero conditional mean and covariance equal to that of Y given X.Source: A. Squared-Error Predictable Component · p. 11

B. Residual Scaling Controls Admitted Stochastic Variability

  • Before output-range projection, λ scales residual sampling variability quadratically: λ = 0 removes it and λ = 1 admits the full proposal variability.The exact quadratic relation need not survive nonlinear projection; λ remains the amplitude control applied before it.Source: B. Residual Scaling Controls Admitted… · pp. 11–12

C. Why Partial Residual Refinement Can Be Optimal

The squared-error analysis explains why full residual injection is not always optimal: proposal error can favor an interior residual weight. The experiments instead select λ using the paper’s validation objective.

C. Why Partial Residual Refinement Can Be Optimal

D. Effect of the Output-Domain Projection

The output-domain projection is analyzed for Euclidean distance to an in-range target, not as a guarantee for perceptual metrics.

D. Effect of the Output-Domain Projection

APPENDIX B COMPLETE SYNTHRAD CAPACITY ANALYSES

The appendices document capacity analyses and per-target benchmark tables, alongside reverse-trajectory checks and studies of spatial residual utility.

APPENDIX B COMPLETE SYNTHRAD CAPACITY ANALYSES

  • Width b counts channels in the first U-Net feature stage, not parameters; identical widths yield different total parameter counts for deterministic and residual networks.Tables IX–X report capacity sensitivity with Det-15.25M fixed; Table X covers the unallocated residual proposal at λ = 1.Source: Appendix B · p. 12, Table IX · p. 13, Table X · p. 13

APPENDIX C DWI+ADC PER-TARGET RESULTS

  • Table XII gives DWI and ADC per-target MAE / SSIM / PSNR / LPIPS values underlying the macro benchmark, using a frozen evaluator and patient-level aggregation.The per-target table covers CT+CTA→DWI+ADC results.Source: Appendix C · p. 12, Table XII · p. 14

A. Oracle-to-Free Reverse-Trajectory Reversal

  • At start timestep 600, the t = 297 interaction was −0.0088 despite a +0.0038 intermediate-state improvement; at start timestep 200, SSIM shifted from +0.0001 oracle to −0.0001 free-running.These cases support judging interventions on the final autonomous sample rather than inferring effects from oracle-like intermediate states.Source: A. Oracle-to-Free Reverse-Trajectory… · p. 13

B. Spatial Residual Utility and the Oracle–Deployable Gap

  • Spearman ρ was −0.0092 for early learned utility; the best within-trajectory validation/test correlations were approximately +0.1850/+0.1070, and trajectory features did not improve validation correlation.Neither learned spatial controller outperformed simple global residual-strength selection.Source: B. Spatial Residual Utility and the… · p. 13

APPENDIX E CTP PER-TARGET RESULTS

  • Table XI gives per-target CTP macro-benchmark values, while its cells report MAE / SSIM / PSNR / LPIPS using frozen evaluation and patient-level aggregation.The table reports CT+CTA→CTP results.Source: Table XI · p. 13, Appendix E · p. 14

Related papers

Loading 2609.31912v1…