Paper deep dive
Uncertainty as a Proxy for Semantic Correctness in Diffusion-Based Medical Image Synthesis
Yuxuan Ou, Konstantinos Kamnitsas, OxAAA Study, AICT Consortium, Regent Lee, Vicente Grau
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Diffusion models can synthesise contrast-enhanced CT (CECT) from non-contrast CT (NCCT), avoiding contrast administration and its environmental and patient-access costs. However, visually realistic images are not necessarily anatomically correct, and the pixel-intensity and feature-space similarity metrics used to assess generation quality do not directly measure anatomical correctness. In this work, we investigate whether uncertainty can serve as a proxy for semantic correctness in diffusion-based medical image synthesis. We study NCCT-to-CECT synthesis using AortaDiff, a multitask diffusion framework that jointly generates CECT images and lumen segmentations. The segmentation output provides an explicit representation of the generated vascular anatomy, enabling segmentation-derived errors to be used as a quantitative measure of generation correctness. Six methods spanning weight (Ensemble, HyperDiff, BayesDiff), architecture-perturbation (MCDropout), generative-stochasticity (RDS) and input-perturbation (TTA) uncertainty are compared at the pixel, region and image levels, and for detection of clinically relevant out-of-distribution (OOD) cases. Uncertainty proves informative at all three spatial scales, remains informative on an external multi-centre dataset under distribution shift, and supports OOD detection. MCDropout stands out among the six: it ranks among the leading methods at every scale, generalizes well on the external dataset, and can be enabled at inference on any model already trained with dropout, so reliable uncertainty comes at no extra training cost. Uncertainty reliably flags severe failures but discriminates poorly among already high-quality images. These findings support uncertainty as a practical and computationally economical signal for quality filtering, reliability assessment and OOD detection in NCCT-to CECT synthesis.
Tags
Links
- Source: https://arxiv.org/abs/2610.03224v1
- Canonical: https://arxiv.org/abs/2610.03224v1
Trouble viewing inline? Open PDF directly â
Full Text
49,121 characters extracted from source content.
Expand or collapse full text
IEEE TRANSACTIONS AND JOURNALS TEMPLATE1 Uncertainty as a Proxy for Semantic Correctness in Diffusion-Based Medical Image Synthesis Yuxuan Ou, Konstantinos Kamnitsas, OxAAA Study, AICT Consortium, Regent Lee, Vicente Grau Abstract â Diffusion models can synthesise contrast- enhanced CT (CECT) from non-contrast CT (NCCT), avoid- ing contrast administration and its environmental and patient-access costs. However, visually realistic images are not necessarily anatomically correct, and the pixel-intensity and feature-space similarity metrics used to assess gener- ation quality do not directly measure anatomical correct- ness. In this work, we investigate whether uncertainty can serve as a proxy for semantic correctness in diffusion- based medical image synthesis. We study NCCT-to-CECT synthesis using AortaDiff, a multitask diffusion framework that jointly generates CECT images and lumen segmentations. The segmentation out- put provides an explicit representation of the gener- ated vascular anatomy, enabling segmentation-derived er- rors to be used as a quantitative measure of gener- ation correctness. Six methods spanning weight (En- semble, HyperDiff, BayesDiff), architecture-perturbation (MCDropout), generative-stochasticity (RDS) and input- perturbation (TTA) uncertainty are compared at the pixel, region and image levels, and for detection of clinically relevant out-of-distribution (OOD) cases. Uncertainty proves informative at all three spatial scales, remains informative on an external multi-centre dataset under distribution shift, and supports OOD detection. MC- Dropout stands out among the six: it ranks among the leading methods at every scale, generalizes well on the external dataset, and can be enabled at inference on any model already trained with dropout, so reliable uncertainty comes at no extra training cost. Uncertainty reliably flags severe failures but discriminates poorly among already high-quality images. These findings support uncertainty as a practical and computationally economical signal for quality filtering, reliability assessment and OOD detection in NCCT-to-CECT synthesis. Index Termsâ Contrast-Enhanced CT Generation, Diffu- sion Models, Uncertainty Quantification, Trustworthy AI I. INTRODUCTION Diffusion models have recently achieved strong perfor- mance in medical image synthesis [1]â[3]. A clinically im- portant application is the synthesis of contrast-enhanced im- ages from non-contrast scans, including contrast-enhanced CT (CECT) from non-contrast CT (NCCT) [4]â[7] and contrast- enhanced from non-contrast MRI [8]â[10]. Contrast adminis- tration adds costs beyond the scan itself. Iodinated contrast me- dia are poorly removed by conventional wastewater treatment Y. Ou, K. Kamnitsas, and V. Grau are with the Department of Engi- neering Science, University of Oxford, Oxford, United Kingdom. R. Lee is with the Nuffield Department of Surgical Sciences, University of Oxford, Oxford, United Kingdom. and are now widely detected as environmental contaminants [11]. Moreover, concerns about kidney risk may lead to contrast-enhanced studies being withheld from patients with impaired renal function [12]. Synthesising CECT from NCCT therefore has implications for the environmental footprint, re- source resilience, and accessibility of routine vascular imaging. These benefits depend on the generated vascular anatomy being trustworthy. However, a generated image that appears visually realistic is not necessarily anatomically correct [13]. Small structural errors in the aorta may lead to incorrect downstream analysis and clinical decisions. Existing image quality metrics such as PSNR, SSIM, LPIPS and MSE mea- sure similarity either between pixel intensities or between feature representations, and do not directly assess whether the generated anatomy is semantically correct [14]. Therefore, reliable assessment of generation correctness remains an im- portant challenge [15]. Uncertainty in a diffusion model can arise from different sources, and uncertainty estimation methods differ primarily in where stochasticity or variability is introduced [16]. We therefore organize diffusion-based UQ methods according to four sources: weight uncertainty, which perturbs or samples the learned parameters [17]â[20]; architecture-perturbation uncertainty, which keeps weights fixed but stochastically modifies the computation graph, with Monte Carlo Dropout (MCDropout) as the predominant approach [21]; generative- stochasticity uncertainty, which captures the variability of diffusion generation across repeated sampling from different initial noise realizations [22], [23]; and input-perturbation uncertainty, which probes model stability under test-time trans- formations of the input image [24]. These sources correspond to the main stages at which variability can enter a sampling- based generation pipeline: the model parameters, the network computation, the diffusion sampling process, and the input condition. Existing evaluations of uncertainty in medical image gener- ation do not test it against semantic correctness, i.e. whether the generated anatomy matches the ground truth. Some mea- sure error as pixel-intensity or feature-space dissimilarity between generated and reference images, using metrics such as MAE, SSIM and LPIPS [25]â[28], which capture anatom- ical error only indirectly. Others validate uncertainty against distribution-level scores such as FID under synthetically de- graded inputs [29]. To our knowledge, no study establishes whether uncertainty corresponds to semantic errors in the generated anatomy. arXiv:2610.03224v1 [cs.CV] 2 Oct 2026 2IEEE TRANSACTIONS AND JOURNALS TEMPLATE Ground Truth Predictions NCCT Multitask Diffusion Multitask Diffusion Framework CECTCECT Lumen Seg Lumen Seg Uncertainty and Semantic Correctness Uncertainty Estimation Weight Uncertainty Input Perturbation Generative Stochasticity Architecture Perturbation BayesDiff, HyperDiff, Ensemble MCDropout RDS TTA Pixel-Level Semantic Error Detection Localized-Region Error Localization Sample-Level Quality Filtering OOD Detection Localized Uncertainty Localized Error Gen CECT GT CECT Quality Filtering: Removing highest-uncertainty samples Gen CECT UQ Map Multi-Scale Reliability Analysis Uncertainty Map Semantic Error Map |GT Lumen â Predicted Lumen| = Semantic Error Map CECT Uncertainty Map Real NCCT Generation uncertainty is evaluated against segmentation-derived semantic errors. UQ Rank 0% 20% 40% 80% 100% Study Overview Aorta Fig. 1.Overview of the proposed evaluation framework. AortaDiff jointly generates CECT and lumen segmentation from NCCT. Uncertainty is evaluated against segmentation-derived semantic error at pixel, region, and image levels, with additional external OOD evaluation. We study these uncertainty quantification methods using a multitask diffusion framework for NCCT to CECT synthesis [30]. The model jointly generates a CECT image and a lumen segmentation mask, as shown in Figure 1. This pro- vides a natural way to evaluate semantic correctness because the segmentation output directly reflects the generated aorta structure. We show that the multitask modelâs segmentation output agrees closely with three independently trained external segmentation networks (nnU-Net, UNet++ and Attention U- Net [31]â[33]) applied to the generated image, confirming that the segmentation branch faithfully reflects the generated vascular anatomy. Its agreement with the ground-truth lumen therefore provides a measure of generation correctness. Using this framework, we systematically evaluate represen- tative uncertainty estimation methods for diffusion models. We perform a multi-scale evaluation that studies uncertainty at the pixel, region, and image level, and assess whether uncer- tainty can detect clinically relevant out-of-distribution (OOD) cases, specifically vascular stents, on an external multi-center dataset. Together, these analyses reveal whether uncertainty corresponds to local semantic errors, whether it can flag failed generations, and whether it remains consistent on an external dataset and identifies OOD cases under distribution shift. Our experiments lead to three main findings. First, uncer- tainty provides meaningful information about semantic cor- rectness at the pixel, region and image levels, and is most informative for severe generation failures. This information is preserved on an external multi-centre dataset, where image- level failure detection remains effective. Second, MCDropout ranks among the leading methods at every scale. When the base model is trained with dropout, reliable uncertainty esti- mates can be obtained without training additional networks. Third, uncertainty maps can be aggregated into image-level scores for quality filtering and for detecting clinically relevant OOD structures such as vascular stents. The main contributions of this work are: âą We propose an evaluation framework that tests diffu- sion model generation uncertainty against semantic cor- rectness, measured by agreement with the ground-truth anatomy rather than image similarity, using the lumen segmentation jointly predicted by a multitask diffusion model. âą We introduce a multi-scale evaluation protocol spanning the pixel, region and image levels, extended to external multi-centre data under distribution shift and to detection of clinically relevant OOD cases (vascular stents). âą We compare six uncertainty quantification methods for diffusion models, spanning four sources of uncertainty. âą We show that MCDropout remains among the leading methods at every scale and under distribution shift, re- quiring no additional training when the base model is trained with dropout. AUTHOR et al.: TITLE3 I. RELATED WORK: UNCERTAINTY QUANTIFICATION METHODS Uncertainty in diffusion models can arise from model parameters, input perturbations, architectural perturbations, and stochasticity in the diffusion sampling process [16]. We therefore evaluate representative methods from these four sources: weight uncertainty (BayesDiff, HyperDiff, and deep ensembles), input perturbation (TTA), architecture perturba- tion (MC Dropout), and generative stochasticity (repeated DDIM sampling). A. Weight Uncertainty Weight uncertainty models variability in the learned param- eters and assumes that parameter disagreement propagates to uncertainty in the generated output [18]â[20]. We consider three approaches with complementary implementations. a) BayesDiff: BayesDiff [19] applies a last-layer Laplace approximation (LLLA) [34] to the diffusion noise-prediction network, fitting a Gaussian posterior around the pre-trained final layer. The resulting predictive variance is propagated through the reverse diffusion process, with the covariance term estimated by Monte Carlo sampling, yielding a pixel-wise uncertainty map. b) HyperDiff: HyperDiff [20] uses a Bayesian hypernet- work [35] to generate plausible weights for the diffusion denoising network. Epistemic uncertainty is obtained from variation across sampled weights, while aleatoric uncertainty captures variation within each weight configuration. c) Deep Ensembles: Deep ensembles [18] approximate weight uncertainty by independently training multiple models and measuring disagreement among their predictions. B. Input Perturbation Uncertainty Input perturbation uncertainty measures model sensitivity to plausible changes in the input. Test-time augmentation (TTA) [24] generates multiple predictions from randomly transformed inputs and estimates uncertainty from their vari- ability. Unlike weight-based approaches, TTA keeps the model parameters fixed and probes sensitivity to input perturbations. C. Architecture Perturbation Uncertainty Architecture perturbation uncertainty estimates reliability by introducing stochasticity into the network computation during inference. Monte Carlo Dropout (MCDropout) [21] keeps dropout active during inference, producing different effective subnetworks across repeated forward passes. Uncertainty is estimated from the variability of the resulting predictions. D. Generative Stochasticity Generative stochasticity captures variability arising from the diffusion sampling process. We use repeated deterministic DDIM sampling (RDS) [23] with different initial noise real- izations. Because the reverse process is deterministic given the initial noise, variability across samples reflects differences in the initial noise and provides a direct measure of generation diversity. I. METHODOLOGY A. Multitask Diffusion Framework We study uncertainty quantification in the setting of NCCT to CECT synthesis of the abdominal aorta. Given an NCCT slice, a multitask diffusion model jointly predicts a synthetic CECT image and a lumen segmentation mask, as shown in Fig. 1 [30]. The two tasks are optimised simultaneously and share a common image representation, so the segmentation output provides an explicit representation of the lumen struc- ture present in the generated image. This allows segmentation- derived errors to be used as a semantic proxy for generation correctness, rather than the appearance similarity measured by PSNR, SSIM and LPIPS. B. Generation Uncertainty For all methods, uncertainty is estimated from the CECT generation branch, yielding a pixel-wise uncertainty map over the synthetic CECT image; the segmentation branch is used only to define the semantic correctness target. C. Semantic Correctness Definition and Evaluation Semantic correctness is defined with respect to the aorta region. All semantic error measures are computed from the lumen segmentation, at three spatial scales: pixel, region and image. 1) Pixel Level: Semantic error is the disagreement between the predicted and ground-truth lumen masks, s pred and s gt : pixels where the masks disagree are semantically incorrect, giving a binary error map. Pixel-wise AUROC, computed within the aortic region, measures whether uncertainty sep- arates correct from incorrect pixels. For sparsification, let A be the aortic pixels and u i the uncertainty at pixel iâA. At fraction tâ [0, 1] we remove the fraction t of pixels with the highest uncertainty, leaving K(t), and recompute Dice over the retained pixels only: Dice(t) = 2|(s pred â© s gt )â© K(t)| |s pred â© K(t)| +|s gt â© K(t)| .(1) Removing an incorrect pixel drops a term from the denomina- tor only and raises retained Dice, whereas removing a correctly predicted lumen pixel also drops the numerator and lowers it. An ideal uncertainty map therefore follows the oracle curve, which removes pixels in order of their true semantic error. We summarise each curve by the area under the sparsification error curve (AUSE, the gap to the oracle, lower is better) and the area under the sparsification curve (AUSC, the retained Dice itself, higher is better). Both are computed from the mean curve aggregated across test slices and therefore carry no standard deviation. 2) Region Level: Pixel-wise analysis does not reveal whether uncertainty can spatially localize errors. We partition the aortic region into superpixels using SLIC [36] with com- pactness 20. Regional semantic error is the proportion of in- correctly predicted pixels in each superpixel R k , and regional uncertainty its mean pixel uncertainty Ìu k = 1 |R k | P iâR k u i . Spearman correlation between the two measures rank agree- ment across regions. For sparsification, whole superpixels are 4IEEE TRANSACTIONS AND JOURNALS TEMPLATE discarded in order of decreasing Ìu k until a fraction t of aortic pixels has been removed, and the retained Dice is recomputed as in (1); curves are again summarised by AUSE and AUSC. 3) Image Level: Semantic correctness is quantified by the lumen Dice of the whole image, which distinguishes successful from failed generations. Since the preceding analyses produce pixel-wise maps, a scalar uncertainty score U is required per image; we aggregate each map by its mean, maximum, 90th percentile, and maximum local average (window size 7). For rejection, each image receives a failure label y = 1[ Dice < Ï ] and detection performance is the AUROC of U against y; because Ï is application-dependent, we sweep it across a range and report AUROC at each. For sparsification, images are progressively removed in order of decreasing U and the average Dice of the retained images is recomputed at each step, compared against the oracle obtained by removing images in order of their true Dice. D. Out-of-Distribution Detection In addition to semantic correctness, we investigate whether uncertainty can identify clinically relevant out-of-distribution (OOD) cases. It evaluates whether uncertainty can detect previously unseen structures that were not present during training. We use vascular stents as a clinically relevant OOD sce- nario. Stents are absent from the training dataset and therefore represent previously unseen structures. Pixel-wise uncertainty maps are aggregated into image-level uncertainty scores and used to classify stent-containing and non-stent slices. Detec- tion performance is quantified using AUROC. Together, the semantic correctness and OOD evaluations provide a comprehensive assessment of whether uncertainty can be used as a reliability signal for diffusion-based medical image synthesis, both for identifying semantic errors and for detecting clinically relevant distribution shifts. IV. DATASETS In this study, we use two datasets containing paired NCCT and CECT scans. Both datasets were acquired with full regulatory approval, which enabled secondary use of imaging data acquired as part of the routine healthcare service. The Oxford dataset (UK Health Research Authority Approval Ref: 13/SC/0250) is used for model training and in-distribution evaluation, while the AICT dataset (UK Health Research Authority Approval Ref: 22/HRA/2303) serves as an exter- nal validation dataset for evaluating the relationship between uncertainty and semantic error under distribution shift and OOD detection. The two datasets were collected from different clinical sites and imaging scanners providing a realistic setting for assessing the robustness. A. Oxford Dataset The Oxford dataset is underpinned by the Oxford Abdomi- nal Aortic Aneurysm (OxAAA) Study [37]â[39]. It consists of 264 patient cases, each containing spatially registered NCCT and CECT scans. Aortic segmentation labels are available for all cases, while 65 cases additionally contain expert-annotated lumen segmentations. Following our previous work [30], the 199 cases without lumen annotations, together with 45 of the 65 lumen-annotated cases, were used for training and validation; the remaining 20 lumen-annotated cases were held out for testing. For model development and evaluation, volumetric scans were processed as independent 2D axial slices. The 20 test cases yielded a total of 1,547 paired NCCTâCECT slices. Slices were retained only if the Dice coefficient between the registered NCCT and CECT aorta masks exceeded 0.90, leaving 1,316 slices in the final test set. B. AICT Dataset The AICT repository is sourced from 10 hospitals across 3 continents (w.aict.ai). It is used as an external dataset to evaluate the generalizability of uncertainty estimation under data distribution shift. The dataset contains paired NCCT and CECT scans acquired from multiple clinical centers using different scanners and acquisition protocols. As expert lumen annotations were not available, automated segmentation methods were used during preprocessing. First, NCCT and CECT volumes were aligned using rigid regis- tration. Aortic segmentations were first generated using an nnU-Net model [31] trained on the Oxford dataset [39], and each result was manually inspected. When the nnU-Net segmentation was judged inadequate, TotalSegmentator [40] was applied as a fallback and used if it produced a satisfactory mask; cases for which neither method yielded an acceptable aortic segmentation were discarded. Lumen segmentations were generated using another nnU-Net model trained on the Oxford dataset, and each mask was manually reviewed. Quality control was subsequently performed to ensure ade- quate image quality and successful preprocessing. Cases were retained only if: (i) the CECT scan exhibited sufficient vascular enhancement to allow clear visualization of the aortic lumen; (i) the scan contained the abdominal aorta; (i) NCCT and CECT volumes could be successfully aligned using rigid registration; (iv) at least one automated segmentation method successfully extracted the aortic mask; and (v) the Dice similarity coefficient between the registered NCCT and CECT aortic masks exceeded 0.90. In addition to the above quality- control criteria, all retained slices and the corresponding automatically generated aortic and lumen segmentation masks were manually reviewed. After preprocessing and quality control, 61 cases were retained for analysis. The final cohort consisted of 21 cases from Opole, Poland; 18 from Leicester, UK; 17 from Herak- lion, Greece; 3 from S Ì ao Paulo, Brazil, and 2 from Brugge, Belgium. This yielded 5,950 slices, of which 1,575 (18 cases) contained vascular stents. Throughout this work, AICT was used exclusively for external evaluation, allowing us to assess the robustness of uncertainty estimation methods on previously unseen data distributions. AUTHOR et al.: TITLE5 V. UNCERTAINTY ESTIMATION IMPLEMENTATION DETAILS Table I summarises the four uncertainty sources, the six representative methods, and their configuration and cost. Two details are not captured there. For BayesDiff, the Laplace approximation is fitted on the 2,244 paired NCCTâCECT slices with lumen masks from the Oxford training set, the co- variance Cov[x,Δ] between the current state and the predicted noise is estimated with 100 Monte Carlo samples per step, and generation uncertainty is the final propagated variance. HyperDiff yields both aleatoric and epistemic estimates; their mean pixel-wise Spearman correlation within the aortic region is 0.962±0.031 and they perform comparably at all evaluation levels, so we report epistemic uncertainty throughout. VI. RESULTS We first validate the use of multitask segmentation as a se- mantic measure for generation correctness. We then investigate the relationship between generation uncertainty and semantic correctness at the pixel, region, and image levels. Finally, we evaluate uncertainty on external data and assess its utility for detecting clinically relevant out-of-distribution structures. A. Using Segmentation as a Semantic Proxy for Generation Correctness We adopt the training data and model configuration of our previous work [30], evaluated on the refined test set of Section IV-A. Under DDIM inference with 10 steps, the multitask framework improves on a two-stage pipeline (a generation-only diffusion model followed by an external nnU- Net) in PSNR (25.37 vs. 25.12), SSIM (0.84 vs. 0.80) and LPIPS (0.06 vs. 0.08) at equal segmentation Dice (0.90). The proxy additionally requires that the segmentation branch reflect the lumen structure present in the modelâs jointly generated image. We therefore applied three independently trained 2D lumen segmenters â nnU-Net [31], UNet++ [32] and Attention U-Net [33], all trained on real CECT annotations from the Oxford dataset â to the generated CECT images. The multitask segmentation agreed with each at a mean Dice of 0.95, confirming that it faithfully represents the synthesized vascular anatomy. Its disagreement with the reference lumen annotation therefore provides a quantitative measure of seman- tic error. 1) Robustness to Registration Quality: Disagreement be- tween predicted and reference lumen masks could also re- flect residual NCCT-to-CECT registration error. We therefore repeated the pixel-, region- and image-level analyses after stratifying the OxAAA test slices by aortic-mask registration Dice (above 0.95, n = 791; 0.90â0.95, n = 525). Across all six methods, pixel-level AUROC differed by at most 0.03 between strata and the AUROC ranking was identical in both, as was the region-level Spearman ranking. Residual misalignment is therefore unlikely to be the dominant driver of the observed uncertaintyâerror relationship. Fig. 2.Pixel-wise sparsification curves for different uncertainty esti- mation methods. Pixels within the aortic region are gradually removed according to their uncertainty ranking, and the retained Dice is recom- puted at each step. The uncertainty-guided curve (blue) is compared against the theoretical best (green) and random removal (red). B. Relationship Between Generation Uncertainty and Semantic Correctness In this section, we investigate whether generation uncer- tainty corresponds to semantic correctness. Specifically, we examine whether high uncertainty is associated with prediction errors, while low uncertainty corresponds to correct predic- tions. We systematically evaluate this relationship at three different scales: pixel-level, region-level, and image-level. 1) Pixel-Level Evaluation: MCDropoutâs uncertainty corre- sponds most closely with pixel-level semantic errors on AUSE, AUSC and AUROC (Table I), though HyperDiff and RDS achieve similar AUROC and the three form a close top cluster; BayesDiff performs no better than random guessing (AUROC 0.493) and diverges rapidly from the oracle in Fig. 2. Notably, MCDropout and RDS reach this level as lightweight, retraining-free methods, whereas HyperDiff requires training a hypernetwork â strong pixel-level correspondence is therefore achievable without algorithmic changes or retraining. The sparsification curves show additional information that the summary metrics do not. MCDropout and HyperDiff both show the largest early gains in retained Dice, confirming that their most-uncertain pixels are frequently erroneous, but HyperDiff declines substantially at high sparsification fractions while MCDropout tracks the oracle throughout. HyperDiff assigns very low uncertainty to some individual error pixels (middle panel, Fig. 4): it identifies highly uncertain error pixels well, but assigns uncertainty less consistently across the full ranking. This inconsistency is what its region-level performance later recovers. 2) Region-Level Evaluation: HyperDiff and MCDropout both outperform the other methods and are effectively tied across all three metrics (Fig. 3, Table I), indicating that both localize error-containing regions effectively. BayesDiff is again weakest, with a negative correlation (â0.24± 0.28) meaning its most uncertain regions tend to be less erro- neous. The ranking differs from the pixel-level ranking, so the two scales capture distinct properties. HyperDiffâs rise is the clearest case. It reaches the top cluster on localized sparsification despite its inconsistent pixel-level uncertainty 6IEEE TRANSACTIONS AND JOURNALS TEMPLATE TABLE I EVALUATED UNCERTAINTY ESTIMATION METHODS. ALL USE DDIM SAMPLING WITH 10 STEPS; GENERATION IS THE MEAN OVER SAMPLES AND UNCERTAINTY THE PIXEL-WISE SAMPLE VARIANCE. INFERENCE TIME IS PER TEST CASE. SourceMethodMechanism and configurationPassesTime (s)Training overhead Input perturbationTTA400 augmented views (bounding-box expansion, ro- tation, isotropic scaling), inverse-warped to original coordinates 40042None Generative stochasticityRDSDeterministic DDIM (η = 0) from 400 initial noise realizations 40042None Architecture perturbationMCDropoutDropout p = 0.5 in U-Net residual blocks; 20 seeds Ă 20 trajectories 40044None Weight Ensemble4 independently trained models Ă 100 samples400444 models HyperDiffBayesian hypernetwork generating denoiser weights; 20 weight sets Ă 20 samples 40055Hypernetwork BayesDiffLast-layer Laplace approximation (diagonal Hessian) on the final convolutional layer 10.22Laplace fitting TABLE I PIXEL-LEVEL COMPARISON OF UNCERTAINTY ESTIMATION METHODS. AUSE (â) AND AUSC (â ) SUMMARISE THE PIXEL-WISE SPARSIFICATION CURVE; AUROC (â ) MEASURES CLASSIFICATION OF CORRECT VERSUS INCORRECT PIXELS. BEST IN BOLD. MethodAUSEâAUSCâAUROCâ TTA0.0370.9040.787 ± 0.094 Ensemble0.0630.8780.689 ± 0.122 RDS0.0350.9060.809 ± 0.094 MCDropout 0.0280.911 0.830 ± 0.073 HyperDiff0.0490.8900.815± 0.084 BayesDiff0.1700.7710.493 ± 0.049 Fig. 3.Region-level sparsification curves. SLIC superpixels are pro- gressively removed in order of decreasing average uncertainty and the retained Dice recomputed at each step, comparing each methodâs uncertainty-guided curve (blue) against the oracle (green) and random removal (red). because region-level scoring averages over each superpixel: the high uncertainty of neighbouring error pixels outweighs the under-flagged ones, so the region still scores high. Fig. 4 shows this â error pixels that survive pixel-level filtering (red box, middle panel) fall inside superpixels flagged as highly uncertain in the region-level map (right panel). Averaging smooths over exactly the inconsistency that limits HyperDiff at the pixel level. 3) Image-Level Evaluation: Fig. 5 reports AUROC as a function of the Dice threshold that defines failure. MCDropout, TTA, RDS and Ensemble exceed 0.9 for severe errors (Dice < 0.65) and stay above 0.8 up to a threshold of roughly Fig. 4. HyperDiff pixel-level under-flagging versus region-level uncer- tainty on a representative slice. TABLE I REGION-LEVEL COMPARISON OF UNCERTAINTY ESTIMATION METHODS. AUSE (â) AND AUSC (â ) SUMMARISE THE LOCALIZED SPARSIFICATION CURVE; SPEARMAN CORRELATION (â ) MEASURES RANK AGREEMENT BETWEEN REGION UNCERTAINTY AND REGION SEMANTIC ERROR. MethodAUSE âAUSC âSpearman â MCDropout0.0230.9010.65 ± 0.19 TTA0.0280.8970.56 ± 0.23 RDS0.0420.8810.60 ± 0.20 Ensemble0.0410.8820.44 ± 0.29 HyperDiff0.0210.8990.64 ± 0.16 BayesDiff0.1600.764-0.24 ± 0.28 0.85. AUROC falls consistently as the threshold rises from 0.6 to 0.9, indicating that uncertainty is most informative for identifying severe failures. Image-level sparsification assesses the ranking rather than the detection (Fig. 6). HyperDiff follows the oracle most closely and retains the highest Dice, reaching about 0.93 once 95% of the most uncertain images are removed. MCDropout, TTA and Ensemble are almost identical; RDS matches them early but declines at high removal fractions, so its ranking is less reliable among the less uncertain images; BayesDiff is worst throughout. All methods leave a clear gap from the oracle, so none ranks images optimally and uncertainty-guided filtering should be applied with care. The four aggregation strategies give highly similar trends and leave the relative ranking of methods unchanged, so these conclusions do not depend on the aggregation choice. 4) External Validation Under Distribution Shift: We repeat the image-level failure detection analysis on the external AICT AUTHOR et al.: TITLE7 Fig. 5. Image-level failure detection. A sample is labelled failed when its lumen Dice falls below a threshold; the threshold is swept from 0.6 to 1.0 and AUROC reported at each. Fig. 6.Image-level sparsification curves. Images are progressively removed in order of decreasing aggregated uncertainty and the average Dice of the retained images recomputed at each step, compared against the oracle (green) and random removal (red). dataset (Fig. 8). MCDropout is the most consistent, exceeding 0.8 across most aggregation strategies and nearly the entire threshold range; TTA, RDS and HyperDiff are intermediate and largely preserve the internal ranking; BayesDiff remains substantially weaker across most thresholds. The AUROC curves are noticeably flatter than on the internal dataset, so failure detection varies less across Dice thresholds under distribution shift. Fig. 7 shows representative external cases. In both, uncer- tainty concentrates where the generated anatomy deviates from the ground-truth CECT, corresponding to the generation errors induced by the unseen structure, though the highest values are not necessarily on the stent itself. 5) Out-of-Distribution Detection: We evaluate whether un- certainty can identify clinically relevant OOD cases involv- ing previously unseen vascular stents (Table IV). All meth- ods except BayesDiff perform similarly, spanning 0.75â0.85 across aggregation strategies, with the highest value shared between TTA, RDS and MCDropout depending on the strategy used. BayesDiff behaves differently from the other methods: mean aggregation is weak (0.48) while maximum aggregation reaches 0.82, the same pattern seen in the image-level analysis. As a reference, A KNN feature-space detector [41] achieves only 0.43â0.44, substantially below every uncertainty-based method. Generated CECT Uncertainty Map Ground Truth CECT Case 1 Case 2 Fig. 7. Representative OOD cases containing vascular stents that were absent from the training dataset. Fig. 8. Image-level failure detection plots for different thresholds in AICT dataset. VII. DISCUSSION Uncertainty provides meaningful information about seman- tic correctness in diffusion-based CECT synthesis, and can ad- ditionally identify clinically relevant out-of-distribution cases. Among the methods evaluated, MCDropout emerges as the method of choice: it performs strongly at every scale and under distribution shift, and requires no additional training when the base model already contains dropout, making it the practical default. In practice this makes MCDropout potentially deploy- able: a single uncertainty map, produced without retraining, lets clinicians flag suspect pixels, localize regions likely to contain errors, and reject unreliable generations. These differences follow from the source each method esti- mates. BayesDiff places a posterior on the final convolutional layer alone, treating all preceding layers as a fixed feature extractor, so its uncertainty reflects one layer rather than the full denoiser â a plausible explanation for its consistently weak performance and its negative region-level correlation. Ensemble and HyperDiff instead capture variability across the entire network, and HyperDiff concentrates it within spatially coherent regions, consistent with its region-level strength. MC- Dropout produces many closely related subnetworks whose disagreement varies smoothly across neighbouring pixels, con- 8IEEE TRANSACTIONS AND JOURNALS TEMPLATE TABLE IV OOD STENT DETECTION AUROC (â ) UNDER FOUR AGGREGATION STRATEGIES: MEAN, MAXIMUM, 90TH PERCENTILE, AND MAXIMUM LOCAL AVERAGE WITH WINDOW SIZE 7. MethodMean âMax âP90 âMWA 7 â MCDropout0.770.830.810.82 RDS0.790.840.790.85 HyperDiff0.750.780.750.81 TTA0.800.850.810.83 BayesDiff0.480.820.530.72 Ensemble0.780.810.790.80 sistent with its stable pixel-level correspondence. TTA reflects sensitivity to input changes, aligning with its OOD perfor- mance, since unseen structures introduce input characteristics the model is less robust to. RDS reflects output diversity across initial noise realizations, and greater variability tends to indicate lower-quality generations. The most appropriate method therefore depends on the intended clinical application. Computational requirements differ substantially (Table I). TTA and RDS apply directly to any pretrained model, and MCDropout is likewise training-free when the base model was trained with dropout, whereas Ensemble and HyperDiff require additional training. Since the training-free methods match or exceed the retraining-based ones at every scale, deployment- time reliability assessment need not carry a retraining cost â relevant where computational budget or energy footprint constrains clinical deployment. Several limitations should be acknowledged. The experi- ments were conducted on a single application, NCCT to CECT synthesis of the abdominal aorta, and generalizability to other organs, modalities and generation problems remains to be validated empirically. We expect the analysis to translate, since many medical image generation tasks involve anatomically meaningful structures whose correctness can be quantified through task-specific segmentations or structural measure- ments, and the framework is independent of the particular segmentation formulation used. In deployment, the estimates studied here may help identify unreliable generations, lo- calize likely semantic errors, and alert clinicians to out-of- distribution structures or artifacts. More broadly, these results suggest that uncertainty should be treated not as a single quantity but as a multi-scale reliability signal, with different mechanisms providing complementary information about a generated image. The evaluation framework proposed here offers a systematic way to study that signal in diffusion-based medical image synthesis, and may support the development of more reliable and deployable generative models for clinical use. ACKNOWLEDGMENT We acknowledge the contribution from the Oxford Ab- dominal Aortic Aneurysm Study and the AICT Consortium (w.aict.ai) for the individual siteâs equivalent approval of the AICT study protocol. Specifically to the following: Oxford Abdominal Aor- tic Aneurysm Study (Ashok Handa, Pierfrancesco Lapolla- Lapolla, Anirudh Chandrashekar); AICT Consortium (Ewa Strauss, Grzegorz Oszkinis, Matt Bown, Michael Klontzas, Edwaldo Edner Joviliano, Saulo Da Silva Cordeiro, Lotte Pyfferoen, Kris Van De Moortele, Steven Van Wortswinkel, Filip Decker, John Kellas, Alex Pitman, David Chang, Juliette Raffor-Lareyre Fabien Lareyre, Stan Snijders, Daniel Feller, Ittai Dayan, Carlos Santin, Ernest Puyol, Anders Nordell, Erik Ranschaert) Yuxuan Ou acknowledges the funding provided by the China Scholarship Council and University of Oxford through the CSCâPAG Oxford Scholarship. Vicente Grau and Regent Lee are partially funded by NetZero AICT (EU New Horizon Grant), which has also funded data collection and curation for the AICT dataset. [1] A. Kazerouni, E. K. Aghdam, M. Heidari, R. Azad, M. Fayyaz, I. Hacihaliloglu, and D. Merhof, âDiffusion models in medical imaging: A comprehensive survey,â Medical image analysis, vol. 88, p. 102846, 2023. [2] J. Luo, L. Yang, Y. Liu, C. Hu, G. Wang, Y. Yang, T.- L. Yang, and X. Zhou, âReview of diffusion models and its applications in biomedical informatics,â BMC Medical Informatics and Decision Making, vol. 25, no. 1, p. 390, 2025. [Online]. Available: https://doi.org/10.1186/s12911-025-03210-5 [3] P. Mirowski and A. Fabija Ì nska, âDiffusion model-based synthesis of brain images for data augmentation,â Biomedical Signal Processing and Control, vol. 113, p. 108940, 2026. [Online]. Available: https://w.sciencedirect.com/science/article/pii/S174680942501451X [4] A. Handa, A. Chandrashekar, R. Uberoi, V. Grau, R. Lee, P. Lapolla, and N. Shivakumar, âA deep learning approach to visualize aortic aneurysm morphology without the use of intravenous contrast agents,â Annals of Surgery, vol. 277, no. 2, p. 449â459, Jan. 2023. [5] P. Shiri, X. Yi, N. P. Mistry, S. Javadinia, M. Chegini, S.- B. Ko, A. Baniasadi, and S. J. Adams, âGenerating synthetic contrast-enhanced chest ct images from non-contrast scans using slice-consistent brownian bridge diffusion network,â 2025. [Online]. Available: https://arxiv.org/abs/2508.16897 [6] L. Zhong, R. Xiao, H. Shu, K. Zheng, X. Li, Y. Wu, J. Ma, Q. Feng, and W. Yang, âNcct-to-cect synthesis with contrast-enhanced knowledge and anatomical perception for multi-organ segmentation in non-contrast ct images,â Medical Image Analysis, vol. 100, p. 103397, 2025. [Online]. Available: https://w.sciencedirect.com/science/article/pii/ S1361841524003220 [7] K. Zheng, M. Huang, X. Li, J. Ma, Q. Feng, W. Yang, and L. Zhong, âContrast flow pattern and cross-phase specificity-aware diffusion model for ncct-to-multiphase cect synthesis,â in Medical Image Computing and Computer Assisted Intervention â MICCAI 2025, J. C. Gee, D. C. Alexander, J. Hong, J. E. Iglesias, C. H. Sudre, A. Venkataraman, P. Golland, J. H. Kim, and J. Park, Eds.Cham: Springer Nature Switzerland, 2026, p. 89â99. [8] Q. Zhang, M. K. Burrage, E. Lukaschuk, M. Shanmuganathan, I. A. Popescu, C. Nikolaidou, R. Mills, K. Werys, E. Hann, A. Barutcu, S. D. Polat, H. C. R. H. Investigators, M. Salerno, M. Jerosch-Herold, R. Y. Kwong, H. C. Watkins, C. M. Kramer, S. Neubauer, V. M. Ferreira, and S. K. Piechnik, âToward replacing late gadolinium enhancement with artificial intelligence virtual native enhancement for gadolinium-free cardiovascular magnetic resonance tissue characterization in hypertrophic cardiomyopathy,â Circulation, vol. 144, no. 8, p. 589â599, 2021. [Online]. Available: https://w. ahajournals.org/doi/abs/10.1161/CIRCULATIONAHA.121.054432 [9] Q. Zhang, M. K. Burrage, M. Shanmuganathan, R. A. Gonzales, E. Lukaschuk, K. E. Thomas, R. Mills, J. L. Pelado, C. Nikolaidou, I. A. Popescu, Y. P. Lee, X. Zhang, R. Dharmakumar, S. G. Myerson, O. Rider, O. A. M. I. O. Study, K. M. Channon, S. Neubauer, S. K. Piechnik, and V. M. Ferreira, âArtificial intelligence for contrast-free mri: Scar assessment in myocardial infarction using deep learningâbased virtual native enhancement,â Circulation, vol. 146, no. 20, p. 1492â1503, 2022. [Online]. Available: https://w. ahajournals.org/doi/abs/10.1161/CIRCULATIONAHA.122.060137 [10] A. F. I. Osman and N. M. Tamam, âContrast-enhanced MRI synthesis using dense-dilated residual convolutions based 3D network toward elimination of gadolinium in neuro-oncology,â J Appl Clin Med Phys, vol. 24, no. 12, p. e14120, Aug. 2023. AUTHOR et al.: TITLE9 [11] A. Sengar and A. Vijayanandan, âComprehensive review on iodinated x-ray contrast media: Complete fate, occurrence, and formation of disinfection byproducts,â Sci Total Environ, vol. 769, p. 144846, Jan. 2021. [12] M. S. Davenport, M. A. Perazella, J. Yee, J. R. Dillman, D. Fine, R. J. McDonald, R. A. Rodby, C. L. Wang, and J. C. Weinreb, âUse of intravenous iodinated contrast media in patients with kidney disease: Consensus statements from the american college of radiology and the national kidney foundation,â Radiology, vol. 294, no. 3, p. 660â668, Jan. 2020. [13] S. Sun, G. Goldgof, A. Butte, and A. M. Alaa, âAligning synthetic medical images with clinical knowledge using human feedback,â in Ad- vances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, p. 13 408â13 428. [14] S. Wang, S. Klein, J. A. Hernandez-Tamames, and D. H. J. Poot, âEvaluating and calibrating diffusion model-derived uncertainty for quantitative mri mapping,â 2026. [Online]. Available: https: //arxiv.org/abs/2608.11942 [15] B. Kocak, M. E. Klontzas, A. Stanzione, A. Meddeb, A. Demircio Ì glu, C. Bluethgen, K. K. Bressem, L. Ugga, N. Mercaldo, O. D Ì Ä±az, and R. Cuocolo, âEvaluation metrics in medical imaging ai: fundamentals, pitfalls, misapplications, and recommendations,â European Journal of Radiology Artificial Intelligence, vol. 3, p. 100030, 2025. [Online]. Available: https://w.sciencedirect.com/science/article/pii/ S3050577125000283 [16] M. Abdar, F. Pourpanah, S. Hussain, D. Rezazadegan, L. Liu, M. Ghavamzadeh, P. Fieguth, X. Cao, A. Khosravi, U. R. Acharya,V.Makarenkov,andS.Nahavandi,âAreviewof uncertainty quantification in deep learning: Techniques, applications and challenges,â Information Fusion, vol. 76, p. 243â297, 2021. [Online]. Available: https://w.sciencedirect.com/science/article/pii/ S1566253521001081 [17] C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra, âWeight uncertainty in neural network,â in International conference on machine learning. PMLR, 2015, p. 1613â1622. [18] B. Lakshminarayanan, A. Pritzel, and C. Blundell, âSimple and scalable predictive uncertainty estimation using deep ensembles,â Advances in neural information processing systems, vol. 30, p. 6402â6413, 2017. [19] S. Kou, L. Gan, D. Wang, C. Li, and Z. Deng, âBayesdiff: Estimating pixel-wise uncertainty in diffusion via bayesian inference,â in Interna- tional Conference on Learning Representations, vol. 2024, 2024, p. 17 046â17 063. [20] M. A. Chan, M. J. Molina, and C. A. Metzler, âEstimating epistemic and aleatoric uncertainty with a single model,â Advances in Neural Information Processing Systems, vol. 37, p. 109 845â109 870, 2024. [21] Y. Gal and Z. Ghahramani, âDropout as a bayesian approximation: Representing model uncertainty in deep learning,â in international conference on machine learning. PMLR, 2016, p. 1050â1059. [22] J. Ho, A. Jain, and P. Abbeel, âDenoising diffusion probabilistic models,â Advances in neural information processing systems, vol. 33, p. 6840â 6851, 2020. [23] J. Song, C. Meng, and S. Ermon, âDenoising diffusion implicit models,â in International Conference on Learning Representations, 2021. [24] M. S. Ayhan and P. Berens, âTest-time data augmentation for estimation of heteroscedastic aleatoric uncertainty in deep neural networks,â in Medical imaging with deep learning, 2018. [25] T. K Ì ustner, K. Hammernik, D. Rueckert, T. Hepp, and S. Gatidis, âPredictive uncertainty in deep learning-based MR image reconstruction using deep ensembles: Evaluation on the fastMRI data set,â Magn Reson Med, vol. 92, no. 1, p. 289â302, Jan. 2024. [26] G. Kutiel, R. Cohen, M. Elad, D. Freedman, and E. Rivlin, âConformal prediction masks: Visualizing uncertainty in medical imaging,â in Trust- worthy Machine Learning for Healthcare. Springer Nature Switzerland, 2023, p. 163â176. [27] C. Ekmekci and M. Cetin, âConformalized generative bayesian imaging: An uncertainty quantification framework for computational imaging,â 2025. [Online]. Available: https://arxiv.org/abs/2504.07696 [28] J. Teneggi, J. W. Stayman, and J. Sulam, âConformal risk control for semantic uncertainty quantification in computed tomography,â in Med- ical Image Computing and Computer Assisted Intervention â MICCAI 2025. Springer Nature Switzerland, 2026, p. 45â55. [29] C. Bench, E. Ahmed, and S. Thomas, âTrustworthy image-to-image translation: evaluating uncertainty calibration in unpaired training sce- narios,â in 2025 International Joint Conference on Neural Networks (IJCNN), 2025, p. 1â10. [30] Y. Ou, N. Bi, J. Pan, J. Yang, B. Yu, U. Zidan, R. Lee, and V. Grau, âAortadiff: A unified multitask diffusion framework for contrast-free a imaging,â in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2026, p. 8242â8251. [31] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, ânnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,â Nature Methods, vol. 18, no. 2, p. 203â211, Feb. 2021. [32] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, âUnet++: A nested u-net architecture for medical image segmentation,â in International workshop on deep learning in medical image analysis. Springer, 2018, p. 3â11. [33] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz et al., âAtten- tion u-net: Learning where to look for the pancreas,â arXiv preprint arXiv:1804.03999, 2018. [34] A. Kristiadi, M. Hein, and P. Hennig, âBeing bayesian, even just a bit, fixes overconfidence in relu networks,â in International conference on machine learning. PMLR, 2020, p. 5436â5446. [35] D. Krueger, C.-W. Huang, R. Islam, R. Turner, A. Lacoste, andA.Courville,âBayesianhypernetworks,âarXivpreprint arXiv:1710.04759, 2017. [36] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S Ì usstrunk, âSlic superpixels compared to state-of-the-art superpixel methods,â Los Alamitos, p. 2274â2282, 2012. [Online]. Available: https: //infoscience.epfl.ch/handle/20.500.14299/80789 [37] R. Lee, A. Jones, F. Woodgate, K. Bellamkonda, N. Killough, L. Fulford- Smith, K. Hurst, I. Cassimjee, and A. Handa, âThe experience of patients during the clinical management pathway of abdominal aortic aneurysms at a NHS trust,â J Patient Exp, vol. 4, no. 4, p. 202â209, Jun. 2017. [38] R. Lee, P. D. Charles, P. Lapolla, I. Cassimjee, H. Huang, B. Kessler, R. Fischer, A. Handa, and Oxford Abdominal Aortic Aneurysm Study, the Oxford Regional Vascular Services, âIntegrated physiological and biochemical assessments for the prediction of growth of abdominal aortic aneurysms in humans,â Ann Surg, vol. 270, no. 1, p. e1âe3, Jul. 2019. [39] A. Chandrashekar, A. Handa, N. Shivakumar, P. Lapolla, R. Uberoi, V. Grau, and R. Lee, âA deep learning pipeline to automate High- Resolution arterial segmentation with or without intravenous contrast,â Annals of Surgery, vol. 276, no. 6, p. 1017â1027, Nov. 2020. [40] J. Wasserthal, H.-C. Breit, M. T. Meyer, M. Pradella, D. Hinck, A. W. Sauter, T. Heye, D. T. Boll, J. Cyriac, S. Yang, M. Bach, and M. Segeroth, âTotalsegmentator: Robust segmentation of 104 anatomic structures in ct images,â Radiology: Artificial Intelligence, vol. 5, no. 5, p. e230024, 2023. [Online]. Available: https: //doi.org/10.1148/ryai.230024 [41] Y. Sun, Y. Ming, X. Zhu, and Y. Li, âOut-of-distribution detection with deep nearest neighbors,â in International conference on machine learning. PMLR, 2022, p. 20 827â20 840.