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

cs.CVSep 25, 2026

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

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

Organizations: Department of Clinical Neurosciences, Cumming School of Medicine, University of Calgary, Calgary, AB, Canada. · Ron and Rene Ward Centre for Healthy Brain Aging Research, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada. · Department of Biomedical Engineering, School of Life Science and Technology, Huazhong University of Science and Technology, Wuhan, Hubei, China. · Department of Biomedical Engineering, Schulich School of Engineering, University of Calgary, Calgary, AB, Canada.

Abstract

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.

Figures & tables

Fig. 1: Overview of PredRA . The deterministic model first captures the part of the target that can be predicted directly from the source. The residual generator then models what remains, including useful stable residual structure that the first model missed. PredRA reuses one residual proposal and varies only its strength $\lambda$ . Validation selects $\lambda^{*}$ to maximize SSIM while keeping LPIPS within the fixed constraint in ( 19 ). In this way, the final prediction recovers additional residual detail without accepting unnecessary generative variation. The schematic shows one CT input for clarity; the experiments also include CT+CTA inputs.

Method	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
SelfRDB [ 36 ]	0.2636	0.7173	15.9121	0.2202
I2I-Mamba [ 4 ]	0.2492	0.7229	15.8951	0.2379
ResViT [ 3 ]	0.2538	0.7139	15.8931	0.2700
MADM [ 35 ]	0.3171	0.7228	16.9847	0.2567
YODA [ 39 ]	0.1620	0.8048	20.7667	0.1395
DET-15.25M	0.1657	0.8033	21.2576	0.1389

TABLE I: CT+CTA

\rightarrow

DWI+ADC frozen-test benchmark under the shared evaluator. Bold and underlining mark the best and second-best among the ranked comparison methods; DET and Full-* rows are reference settings and are excluded from that ranking. External baselines use the compute-bounded schedules described in the text.

Method	$\lambda$	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
SelfRDB [ 36 ]	–	0.1929	0.8112	12.6603	0.1289
I2I-Mamba [ 4 ]	–	0.1888	0.8080	12.8245	0.1254
ResViT [ 3 ]	–	0.1874	0.8125	12.8727	0.1187
MADM [ 35 ]	–	0.1744	0.8324	13.8996	0.1426
YODA [ 39 ]	–	0.1542	0.8521	14.9571	0.1203
DET-15.25M	0	0.1744	0.8347	13.9691	0.1627

TABLE II: CT+CTA

\rightarrow

CTP frozen-test macro benchmark. Bold and underlining mark the best and second-best among the ranked comparison methods; DET and Full-* rows are reference settings and are excluded from that ranking. “val.” denotes target-specific validation-selected residual strengths.

Method	$\lambda$	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
SelfRDB [ 36 ]	–	0.1364	0.7038	15.7203	0.1900
I2I-Mamba [ 4 ]	–	0.1240	0.7140	16.4184	0.1563
ResViT [ 3 ]	–	0.1184	0.7195	16.8226	0.1564
MADM [ 35 ]	–	0.1079	0.7621	17.6970	0.1692
YODA [ 39 ]	–	0.1539	0.6245	15.0081	0.3291
DET-15.25M	0	0.1139	0.7510	17.3619	0.1823

TABLE III: SynthRAD CT

\rightarrow

MRI frozen-test benchmark under the common canonical all-slice evaluator. Bold and underlining mark the best and second-best among the ranked comparison methods; DET and Full-* rows are reference settings and are excluded from that ranking.

Fig. 2: Qualitative comparison across the three translation settings. Rows show SynthRAD CT $\rightarrow$ MRI, CT+CTA $\rightarrow$ CTP targets (Tmax, CBF, and CBV), and CT+CTA $\rightarrow$ DWI+ADC. Columns compare the available input image(s), paired ground truth, five external baselines, the deterministic anchor, and the selected PredRA result. All examples use fixed frozen-test cases and the same display protocol within each target. Quantitative conclusions are based on the complete frozen test sets.

Fig. 3: Focused Tmax example comparing the deterministic anchor, the unallocated Full-Flow32-12.63M proposal, and PredRA-Flow32-12.63M. The first row shows the full CT and CTA inputs and the corresponding target and predictions. The second row uses the same square region for every panel. The deterministic prediction captures the broad perfusion pattern, Full-Flow32-12.63M adds finer local structure, and PredRA-Flow32-12.63M uses the same generated residual with limited amplitude.

Model	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
DET-2.44M	0.1029	0.7385	17.7259	0.1933
DET-9.76M	0.0999	0.7446	17.9875	0.1847
DET-15.25M	0.0967	0.7596	18.2548	0.1750
DET-61.00M	0.0975	0.7537	18.1865	0.1750

TABLE IV: Representative SynthRAD deterministic-capacity points on the full validation set. DET-15.25M is the validation-selected anchor. The complete seven-capacity sweep, including architecture width

b

, is reported in Appendix Table VIII .

Residual model	$\lambda$	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
PredRA-DDPM-1.55M	0.4000	0.1155	0.7436	17.2249	0.1618
PredRA-DDPM-12.63M	0.4500	0.1147	0.7490	17.3000	0.1537
PredRA-Flow32-1.55M	0.4000	0.1151	0.7442	17.2541	0.1649
PredRA-Flow32-8.85M	0.4500	0.1152	0.7451	17.2450	0.1599
PredRA-Flow32-12.63M	0.5000	0.1144	0.7494	17.3293	0.1537

TABLE V: Representative post hoc SynthRAD frozen-test residual-backend capacity sensitivity with DET-15.25M fixed. Each

\lambda

is selected independently on validation; residual capacity itself is not selected from these test curves. The complete DDPM/Flow capacity sweep is reported in Appendix Table IX ; full (

\lambda=1

) results are reported in Appendix Table X .

Fig. 4: SynthRAD residual-backend capacity sensitivity with DET-15.25M fixed. Points show the SSIM–LPIPS settings; labels give residual-network parameter counts. Markers are not connected because the response is non-monotonic. The 12.63M PredRA-Flow32 and PredRA-DDPM refinement points are nearly coincident.

Steps	$\lambda$	Full-Flow				PredRA-Flow
		MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
1	0.9500	0.1128	0.7571	17.5260	0.1858	0.1127	0.7572	17.5257	0.1858
2	0.5500	0.1146	0.7493	17.3139	0.1786	0.1134	0.7541	17.4293	0.1824
4	0.6500	0.1183	0.7338	16.9401	0.1693	0.1152	0.7461	17.2474	0.1740
8	0.5500	0.1217	0.7207	16.6486	0.1628	0.1157	0.7435	17.2100	0.1686
16	0.4500	0.1241	0.7109	16.4274	0.1595	0.1154	0.7440	17.2361	0.1664

TABLE VI: SynthRAD Flow sampling-step sensitivity using the 1.55M residual backend and DET-15.25M. Full-Flow uses

\lambda=1

; PredRA-Flow uses the independently validation-selected

\lambda

shown for each step count.

Method	Params (M)	Trainable (M)	Sampler steps	GFLOPs/native eval	Net eval (ms)	Peak infer GPU (GB)
PredRA-DDPM-12.63M	27.8820	12.6280	1000	$\geq$ 116.5470	17.9030	0.2710
PredRA-Flow32-12.63M	27.8820	12.6280	32	$\geq$ 116.5470	17.7570	0.2710
SelfRDB	39.7270	39.7270	–	$\geq$ 71.2600	21.1320	1.8710
YODA	53.5070	53.5070	–	$\geq$ 182.3400	13.8380	3.1360
I2I-Mamba	109.3010	109.3010	–	$\geq$ 169.8630	9.3290	0.9620
ResViT	123.4420	123.4420	–	$\geq$ 119.8000	12.6290	0.5170

TABLE VII: Standardized representative-forward efficiency analysis for the reported 12.63M PredRA configurations and external SynthRAD baselines. These PredRA configurations combine DET-15.25M with a 12.63M residual backend and contain 27.882M unique parameters; only the 12.628M residual network is optimized during residual-stage training because the deterministic anchor is frozen. The Flow32 realization is the fast PredRA operating variant, requiring 32 residual-network evaluations rather than the 1000 ancestral DDPM steps. Net evaluation time denotes one deterministic-anchor forward pass plus one native residual-network evaluation at

192\times 192

, batch size 1; it is an per-network measure and does not represent complete sampling wall time. GFLOPs marked with

\geq

are lower bounds when the profiler does not account for all operators.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Width $b$	Params (M)	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
DET-2.44M	32	2.4430	0.1029	0.7385	17.7259	0.1933
DET-3.82M	40	3.8160	0.1028	0.7324	17.7003	0.1951
DET-5.49M	48	5.4940	0.1016	0.7340	17.7611	0.1938
DET-9.76M	64	9.7640	0.0999	0.7446	17.9875	0.1847
DET-15.25M	80	15.2540	0.0967	0.7596	18.2548	0.1750
DET-29.89M	112	29.8930	0.0998	0.7428	18.0064	0.1840

Appendix

TABLE VIII: Complete deterministic-capacity validation sweep. The selected DET-15.25M anchor corresponds to width

b=80

Backend	Width $b$	Params (M)	$\lambda$	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
DDPM	32	1.5470	0.4000	0.1155	0.7436	17.2249	0.1618
DDPM	48	3.3100	0.4000	0.1152	0.7458	17.2515	0.1594
DDPM	64	5.7450	0.4500	0.1152	0.7462	17.2513	0.1547
DDPM	80	8.8510	0.4500	0.1150	0.7480	17.2709	0.1542
DDPM	96	12.6280	0.4500	0.1147	0.7490	17.3000	0.1537
Flow32	32	1.5470	0.4000	0.1151	0.7442	17.2541	0.1649

Appendix

TABLE IX: Complete post hoc frozen-test PredRA residual-backend capacity sensitivity with DET-15.25M fixed. Width

b

is reported only to document the architecture; model names use actual residual parameter counts. Bold marks the best value in each metric across the displayed sweep.

Backend	Width $b$	Params (M)	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	Steps
DDPM	32	1.5470	0.1273	0.7007	16.1777	0.1560	1000
DDPM	48	3.3100	0.1270	0.7042	16.1971	0.1530	1000
DDPM	64	5.7450	0.1259	0.7088	16.2848	0.1502	1000
DDPM	80	8.8510	0.1255	0.7121	16.3108	0.1491	1000
DDPM	96	12.6280	0.1248	0.7139	16.3686	0.1481	1000
Flow32	32	1.5470	0.1254	0.7043	16.3045	0.1571	32

Appendix

TABLE X: Complete post hoc frozen-test unallocated residual-proposal (

\lambda=1

) capacity sensitivity with DET-15.25M fixed. Bold marks the best value in each metric across the displayed sweep.

Method	CBF	CBV	Tmax
DET-15.25M	0.2045 / 0.8287 / 12.9148 / 0.1720	0.1792 / 0.8304 / 13.7866 / 0.1651	0.1395 / 0.8450 / 15.2059 / 0.1511
Full-DDPM-12.63M	0.2087 / 0.8037 / 12.4641 / 0.1241	0.1934 / 0.8005 / 12.7943 / 0.1174	0.1676 / 0.8116 / 13.6962 / 0.1163
PredRA-DDPM-12.63M	0.1986 / 0.8169 / 13.0624 / 0.1260	0.1808 / 0.8160 / 13.5926 / 0.1189	0.1483 / 0.8325 / 14.7830 / 0.1182
Full-Flow32-12.63M	0.1980 / 0.8083 / 12.8494 / 0.1181	0.1691 / 0.8051 / 13.6533 / 0.1121	0.1490 / 0.8217 / 14.6028 / 0.1120
PredRA-Flow32-12.63M	0.1945 / 0.8199 / 13.1944 / 0.1208	0.1662 / 0.8192 / 14.0914 / 0.1145	0.1431 / 0.8322 / 14.9864 / 0.1145
SelfRDB	0.2112 / 0.8093 / 12.0095 / 0.1333	0.2011 / 0.8065 / 12.4451 / 0.1276	0.1662 / 0.8179 / 13.5263 / 0.1257

Appendix

TABLE XI: Per-target CT+CTA

\rightarrow

CTP results from the shared-evaluator validated benchmark. Each cell reports MAE / SSIM / PSNR / LPIPS.

Method	DWI	ADC
DET-15.25M	0.0924 / 0.8414 / 26.1227 / 0.1402	0.2389 / 0.7652 / 16.3926 / 0.1376
Full-DDPM-12.63M	0.1003 / 0.8194 / 24.1156 / 0.1427	0.2900 / 0.7359 / 14.4691 / 0.1197
PredRA-DDPM-12.63M	0.0912 / 0.8433 / 26.1071 / 0.1346	0.2354 / 0.7689 / 16.4847 / 0.1220
Full-Flow32-12.63M	0.0901 / 0.8136 / 25.4219 / 0.1424	0.3306 / 0.6865 / 13.6535 / 0.1537
PredRA-Flow32-12.63M	0.0897 / 0.8434 / 26.4032 / 0.1331	0.2387 / 0.7654 / 16.4208 / 0.1371
SelfRDB	0.1697 / 0.7798 / 18.9103 / 0.2196	0.3574 / 0.6549 / 12.9138 / 0.2207

Appendix

TABLE XII: Per-target CT+CTA

\rightarrow

DWI+ADC results from the shared-evaluator benchmark. Each cell reports MAE / SSIM / PSNR / LPIPS.

Image-to-Image Translation

Image-to-image translation maps images between visual domains or modalities while preserving relevant content, structure, and semantics.

See all 113 papers

Aug 23, 2026

GET: Generative Embedding Translation for Medical Image Segmentation

Generative segmentation provides an alternative to direct pixel-wise prediction by operating on learned latent representations, but effective image-to-mask translation must preserve target structure while remaining computationally efficient. We propose Generative Embedding Translation (GET), a structured embedding-translation framework that progressively transforms image embeddings into mask embeddings within the frozen latent space of a Stable Diffusion VAE. GET uses a U-Net-style Embedding Translation Network with 1.07M trainable parameters, combining Mobile Bottleneck Convolutions, Subsampled Self-Attention, and Multi-scale Feature Enrichment for local modeling, global context, and multi-scale refinement. Across five medical segmentation datasets, GET outperforms generative, CNN, and Transformer baselines. Compared with the strongest generative baseline, GMS, GET improves average Dice and IoU by 0.93% and 1.26%, reduces HD95 by 0.81 pixels, and uses 31.41% fewer trainable parameters. Under bidirectional BUS-BUSI domain shift, GET further improves Dice and IoU by 3.51% and 3.39%, while reducing HD95 by 27.37 pixels. Our code is available at: https://github.com/maklachur/GET.

Md Maklachur Rahman, Md Hasan Al Banna, Saraf Anjum +2

Texas A&M University, College Station, TX, USA · Independent Researcher · Alfa Laval

Image Segmentation Medical Image Segmentation Diffusion Models for Medical Image Segmentation

Ask PDF

Oct 25, 2026

StaX-Net: Stackelberg Interleaved Asymmetric Network for Cross-Modality Medical Image Translation

Cross-modality medical image translation is challenged by anatomically inconsistent synthesis that can compromise clinical diagnosis. We propose StaX-Net, a Stackelberg interleaved asymmetric network that incorporates a partially coupled encoder-discriminator structure, and a 2D feature aggregation and spatial transformation (FAST) layer that distills target-supervised priors under the learning using privileged information paradigm. The shared blocks between the generator and discriminator enhance architectural feature alignment. To stabilize the leader-follower training, a Stackelberg-inspired asymmetric optimization scheme employs explicit shared-parameter freezing and asymmetric learning rates. The FAST layer distills target-domain priors during training and applies a source-only transform at inference. This constrains the synthesis process through target-informed feature representations relative to standard generative architectures. The code is available on GitHub for public use.

Amin Golnari, Saeid Sanei

College of Engineering & Computer Science, VinUniversity, Hanoi, Vietnam · College of Health Science, VinUniversity, Hanoi, Vietnam · Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK

Image-to-Image Translation Medical Image Synthesis Adversarial Training

Ask PDF

May 13, 2026

Cross Modality Image Translation In Medical Imaging Using Generative Frameworks

Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and Positron Emission Tomography (PET) provide complementary information about tissues. Medical image-to-image (I2I) translation enables virtual scanning by synthesizing a target modality from a source one without requiring an additional acquisition. Despite growing interest, many methods operate on 2D slices, are evaluated on isolated tasks under different experimental settings, and lack clinically oriented assessment. This work presents a reproducible benchmark for 3D I2I translation in oncological imaging that compares seven generative models: three Generative Adversarial Networks (Pix2Pix, CycleGAN, and SRGAN) and four latent models (Latent Diffusion Model, Latent Diffusion Model+ControlNet, Brownian Bridge, and Flow Matching). The benchmark comprises 77 experiments across eleven configurations drawn from five datasets, covering three anatomical regions (head/neck, lung, and pelvis) and four translation directions (cone-beam CT to CT, MRI to CT, CT to PET, and T2-weighted MRI to T2-FLAIR). Under the evaluated configurations, SRGAN achieves the highest quantitative image fidelity across all tasks, while latent models perform less well, due to information loss introduced by the variational autoencoder. A tumor-level analysis reveals that all models struggle with small lesions and that, in CT to PET synthesis, models reproduce tumor shape more reliably than tracer uptake values. A Visual Turing test involving 17 physicians, including 15 radiologists, shows near-chance classification accuracy (56.7%), suggesting that experts struggle to distinguish real from synthetic volumes under the viewing conditions of the study. Expert preferences do not follow quantitative rankings, exposing a dissociation between quantitative metrics and clinical preference.

Giulia Romoli, Filippo Ruffini, Francesco Di Feola +20

Department of Diagnostics and Intervention, Radiation Physics, Biomedical Engineering, Umeå University, Universitetstorget, 4, Umeå, 90187, Sweden · Unit of Artificial Intelligence and Computer Systems, Department of Engineering, Università Campus Bio-Medico di Roma, Via Alvaro del Portillo, 21, Roma, 00128, Italy · Università Cattolica del Sacro Cuore, Largo Francesco Vito, 1, Roma, 00168, Italy +12

Image-to-Image Translation Medical Image Synthesis Medical Image Benchmarks Generative Adversarial Networks

Ask PDF

Abstract

Image-to-Image Translation

Explore similar work

GET: Generative Embedding Translation for Medical Image Segmentation

StaX-Net: Stackelberg Interleaved Asymmetric Network for Cross-Modality Medical Image Translation

Cross Modality Image Translation In Medical Imaging Using Generative Frameworks