{"id":"f2f2ad9d-fefe-40bd-8828-45e5f84880ef","arxiv_id":"2501.19094","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ADDGAN learns clean image distributions from noisy CT and DBT measurements by feeding generated objects through the known imaging operator, and it beats AmbientGAN baselines on FID and observer-task metrics.","lead":"The authors introduce ADDGAN, a generative model that learns to produce realistic medical images from noisy scans by combining denoising diffusion GANs with a model of the imaging system. It produces sharper CT and breast tomosynthesis images than earlier AmbientGAN approaches, which could improve how medical imaging systems are tested and compared.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generator in Eq. (5) computes ŷ0 = H(x̂0) without adding the measurement noise n that defines the real y0 in Eq. (6), so the discriminator may exploit noise differences and bias the learned object distribution.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the generator's simulated measurement ŷ0 = H(x̂0) omits the additive measurement noise n that defines the real measurement y0 in Eq. (6). This is the single most important threat to the central claim that ADDGAN learns a realistic SOM from noisy measurements, because it directly undermines the statistical equivalence between the training-time measurement model and the real imaging system. My independent reading of the paper confirms that no noise is sampled in the generator path for either CT or DBT: the CT text explicitly defines y0 = FBP(R f + n) and then says 'the Radon transform R and the FBP operator were applied to the generated objects x̂0 to produce ŷ0', with no n; the DBT text says 'directly added Gaussian noise ... to generate noisy image data y0', but the generator's ŷ0 is described as the result of applying the imaging operator to x̂0, without re-adding that noise. The concern is therefore not a disagreement with consensus but an internal inconsistency between the measurement model used to create training data and the one used in the generator. The concrete test I propose would settle whether this mismatch is observable in practice, by comparing the noise-inclusive variant with the reported results. I do not see other concerns that are equally load-bearing: the lack of error bars and code is a reproducibility weakness, but it does not by itself invalidate the central claim; the reported FID and ROC improvements are consistent across metrics, though based on single runs. The verdict remains CONDITIONAL, as the reader suggested, because the measurement-noise omission is plausible and fixable, but it must be addressed before the claim of establishing realistic SOMs is accepted fully. I agree with the reader's assessment and see no reason to change the verdict.","tokens_in":6345,"tokens_out":3372,"duration_ms":36955,"concrete_test":"Retrain ADDGAN on the same datasets with the generator's simulated measurement modified to include the real measurement noise: for CT, sample n per sinogram element with n~N(0,1) and set ŷ0 = FBP(R x̂0 + n); for DBT, sample n~N(0,0.06²) and set ŷ0 = x̂0 + n. Keep all other architecture, hyperparameters, and data identical, and run at least 3 seeds. Then compare (a) FID between generated x̂0 and ground-truth x0, (b) FID to the noisy y0, and (c) Hotelling observer AUC, against the values reported in Tables 1–2 and Fig. 5. If the noise-inclusive version produces FID and AUC statistically closer to ground truth (e.g., improved FID to x0), the reported results were biased by unmodeled measurement noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ADDGAN learns a realistic stochastic object model from noisy measurements. For the CT study, the real measurement process is y0 = FBP(R f + n), where n is i.i.d. Gaussian with std 1 in sinogram space (Eq. 6). In the ADDGAN generator path, however, the simulated measurement is ŷ0 = FBP(R x̂0), with no noise sampled or added. The same omission occurs in the DBT study, where the real y0 = x0 + n (Gaussian, std 0.06), but the generator's ŷ0 is just x̂0. Because the discriminator is trained on pairs (yt, yt−1) derived from real noisy y0, it can learn to distinguish real measurements from the noiseless simulated measurements solely by the presence or absence of measurement noise. The generator then faces an incentive to encode the missing noise into the object x̂0 itself, making the learned SOM contaminated by measurement noise rather than being the desired object statistics. This is not a minor detail: the whole point of an ambient model is to invert the full measurement operator including its stochastic components. The paper does not sample measurement noise anywhere in the generator's forward simulation, so the training signal is statistically mismatched. If the discriminator can reliably detect the absence of noise, the learned x̂0 distribution will be biased. The DBT study's 'toy' noise model actually makes the omission clearer, since the noise is simple additive Gaussian and should be trivial to include. The paper does not flag this omission or argue that the diffusion noise in the forward process absorbs it; the two noise sources play different roles. This is the weakest load-bearing assumption in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ADDGAN, an ambient variant of the denoising diffusion GAN (DDGAN), intended to learn stochastic object models (SOMs) from noisy medical image measurements. The generator synthesizes an object, maps it through a known imaging operator (with reconstruction for CT) to produce simulated measurements, and uses these in the posterior sampling of the denoising diffusion process. Experiments on a stylized CT dataset (DeepLesion) and a DBT dataset (BCS-DBT) compare ADDGAN against DDPM, DDGAN, Ambient StyleGAN3, and ProAmGAN, using FID and Hotelling observer AUC. The paper reports that ADDGAN achieves the best FID and yields AUC closest to ground truth.","tokens_in":6656,"tokens_out":5595,"duration_ms":57645,"significance":"If correct, the paper would provide a fast, high-quality method for estimating object statistics from noisy clinical data, an important goal for task-based image quality assessment. The use of a diffusion GAN as the backbone is well-motivated, and the empirical results are promising. However, the central modeling assumption—that the generator's simulated measurement properly represents the real noisy measurement—is not established, so the significance is conditional on resolving that issue.","major_comments":[{"comment":"The generator's simulated measurement ŷ0 = H(x̂0) omits the additive measurement noise n that defines the real y0. In the CT study, real y0 = FBP(Rf + n), while simulated ŷ0 = FBP(Rx̂0); in the DBT study, real y0 = x0 + n, while simulated ŷ0 = x̂0. Because the discriminator is trained on real noisy y0 and noiseless simulated ŷ0, it can distinguish them based solely on the presence of measurement noise. The generator may then encode the missing noise into the object x̂0, biasing the learned SOM. This is a load-bearing flaw for the paper's claim of learning SOMs from noisy data. Please either modify the forward simulation to sample measurement noise from the known noise model, or provide a theoretical argument or ablation study showing the omission is harmless.","section":"Section 2.3 (Eq. 5) and Section 3.1 (Eq. 6)"},{"comment":"The evaluation measures generated objects against ground-truth clean objects, which is appropriate for SOM assessment. However, because the generator's simulated measurements are noiseless, the reported FID improvement may stem from the generator producing noise-free objects rather than correctly inverting the stochastic measurement process. To substantiate the ambient claim, please report a distributional comparison between real y0 and model-simulated y0 (e.g., Wasserstein distance in sinogram/image space), not only x0-based metrics.","section":"Section 3.2 (Tables 1, 2 and Fig. 5)"}],"minor_comments":[{"comment":"In the last expression of Eq. (5), the argument of G is written as Gθ(xt, z, t), but in the ambient setting it should be yt to match the preceding definitions; please correct this typo.","section":"Eq. (5)"},{"comment":"For the CT study, the sinogram noise standard deviation is set to 1 while the object intensities are normalized to [0,1]; please clarify how this noise level was chosen and report the resulting noise level in the reconstructed image domain.","section":"Section 3.1"},{"comment":"The captions contain a typo, 'Denoisng' instead of 'Denoising'; please also make the caption labels consistent with the subfigures.","section":"Figures 2 and 3"},{"comment":"The paper assumes the imaging operator H (and reconstruction operator O) is exactly known and differentiable; a brief discussion of robustness to operator mismatch would be valuable given the method's reliance on this assumption.","section":"Conclusions and broader context"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This manuscript builds on the authors' prior AmbientGAN work and proposes a plausible extension to DDGAN. The main concern is the missing measurement noise in the generator's forward model, which we believe warrants major revision. The paper is within the journal's scope as a methodological contribution to task-based image quality. No concerns about citation practices beyond typical self-citation within a research group. I recommend requiring the authors to address the noise-modeling issue before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the ADDGAN paper. The core idea is genuinely new: take DDGAN's fast few-step reverse process and give it the AmbientGAN treatment, so the generator produces an object, passes it through the known measurement operator (plus reconstruction, for CT), and the discriminator works on the resulting 'measurement' pairs. The CT and DBT experiments are well designed—two clinical datasets, realistic forward model for CT, and comparison against DDPM, DDGAN, and two AmbientGAN variants. The reported FID and Hotelling observer results favor ADDGAN, and the qualitative images look cleaner than the baselines. That part is solid and worth building on.\n\nThe soft spot is not small. In Eq. (6) the real CT measurement is y0 = FBP(Rf + n), with i.i.d. Gaussian noise added to the sinogram. But in the generator path (Eq. (5) and Section 3.1) the simulated measurement is ŷ0 = H(x̂0) — no noise term is sampled. Same story for DBT: real y0 = x0 + Gaussian noise, but ŷ0 = x̂0. The discriminator is trained on pairs derived from real y0 and fake y0, so it can learn to tell them apart simply by the presence or absence of measurement noise. The generator then has an incentive to fold that missing noise into x̂0, which would bias the learned object distribution. The paper never flags this or argues that diffusion noise absorbs it; those are separate sources. This goes to the heart of the ambient-model claim. The authors need to either add noise to ŷ0 in the generator path, or show formally that the mismatch doesn't bias the object statistics.\n\nThe other issue is that the numbers come from single runs—no error bars, no significance tests. The FID gap to AmbientGAN is large, but without variance we can't tell if it's robust. And no code is released, which makes it hard to check the noise handling directly.\n\nWho gets value from this? People working on SOMs and task-based IQ assessment, and anyone applying GAN/diffusion hybrids to inverse problems. For them, this is a useful new architecture and a clear cautionary example about measurement noise. I'd send it to peer review with a request for major revision—the flaw is fixable, the experiments are relevant, and the construction is worth publishing once the noise modeling is fixed. If that's addressed, it would be a credible contribution.","headline":"ADDGAN is a sensible combination of DDGAN and AmbientGAN with good empirical results, but the generator's omission of measurement noise in its simulated measurements is a load-bearing flaw that biases the learned object distribution.","tokens_in":7248,"tokens_out":4136,"would_cite":true,"duration_ms":39780,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes ADDGAN, which trains a denoising diffusion GAN through the imaging operator, so stochastic object models can be learned from noisy CT and DBT data without clean ground truth.","keywords":["stochastic object model","denoising diffusion GAN","AmbientGAN","computed tomography","digital breast tomosynthesis","task-based image quality","generative adversarial network","noisy image data"],"falsifier":"Train ADDGAN on CT data whose sinogram noise is Poisson-distributed rather than the i.i.d. Gaussian used in the paper, then measure the Hotelling-observer AUC on generated versus ground-truth patches; if the AUC gap grows or the discriminator can reliably tell real from simulated measurements, the ambient noise model is insufficient and the 'noise-independent SOM' claim fails for realistic noise.","tokens_in":6064,"feed_emoji":"🩻","tokens_out":6421,"duration_ms":59888,"temperature":0.7,"pith_summary":"The paper proposes ADDGAN, an augmented denoising diffusion GAN that learns stochastic object models (SOMs) directly from noisy medical image data, without access to clean ground truth. The central claim is that by routing reverse-diffusion steps through a known imaging operator, the model can disentangle object variability from measurement noise and generate images whose textures match the true object distribution. On clinical CT and digital breast tomosynthesis datasets, ADDGAN reports lower Fréchet Inception Distance than state-of-the-art AmbientGAN variants and yields Hotelling-observer detection performance closer to ground truth. The authors argue this makes ADDGAN a practical tool for task-based image quality assessment with experimental data.","feed_headline":"ADDGAN learns object models from noisy CT and DBT scans","feed_subtitle":"An ambient denoising diffusion GAN beats AmbientGAN on texture fidelity and observer tasks.","key_machinery":"The load-bearing mechanism is the ambient posterior-sampling identity in Eq. (5): pθ(y_{t-1} | y_t) = ∫ p(z) q(y_{t-1} | y_t, H(Gθ(x_t, z, t))) dz. It replaces the clean image used in standard DDGAN denoising with the simulated measurement of the generated object, enforcing consistency between the synthesized object and the observed noisy data. For CT, H includes the Radon transform and FBP reconstruction; for DBT in this study, H is additive Gaussian noise. This identity is what allows the generator to be trained without clean target images.","core_discovery":"ADDGAN extends the DDGAN framework by inserting the imaging system into the denoising loop. At each reverse step the generator produces a candidate object x̂0, and the known measurement operator H (for CT, Radon transform followed by filtered back-projection) converts it into a simulated noisy image ŷ0 that conditions the posterior sampling. Training therefore uses only the measured noisy images y0, yet the final output is an object-space sample. In the reported experiments, ADDGAN achieves FID of 30.34 on CT and 41.13 on DBT, outperforming ProAmGAN (41.80) and Ambient StyleGAN3 (70.04), and its generated DBT patches produce SSIM and Hotelling-observer (a standard linear signal-detection template) statistics much closer to ground truth than the comparison models.","pith_inferences":["The ambient loop should transfer to other computed imaging systems (PET, MRI, ultrasound) whenever the forward operator is differentiable and the noise process can be simulated; the paper only demonstrates CT and DBT.","If real measurement noise is signal-dependent or correlated, the simple additive or Gaussian noise models used here may need to be replaced by a learned noise model, otherwise ADDGAN could memorize noise statistics rather than object statistics.","A natural next test is to use ADDGAN-generated SOMs as backgrounds for virtual clinical trials and compare the resulting task-based image quality rankings with physical phantom measurements."],"forward_implications":["ADDGAN can establish SOMs from raw measured image data, so task-based image quality studies no longer require clean or denoised training images.","The same architecture inherits DDGAN's fast sampling (four to eight diffusion steps in these experiments), avoiding the slow generation of standard diffusion models.","Generated objects from ADDGAN are, by the paper's metrics, statistically closer to ground-truth anatomy than those from AmbientGAN models, particularly in high-resolution textures.","The Hotelling-observer comparison indicates that ADDGAN-produced backgrounds support signal-detection tasks with performance similar to real backgrounds, a prerequisite for using the SOM in virtual clinical trials."],"supporting_citations":[{"why":"Introduces the AmbientGAN idea of training a generator through a known measurement operator on noisy data, which ADDGAN adapts to diffusion.","marker":"[8]"},{"why":"Provides the advanced AmbientGAN baselines (Ambient StyleGAN3 and ProAmGAN) and the SOM-from-measurements framework that ADDGAN compares against.","marker":"[9]"},{"why":"Defines the denoising diffusion GAN architecture that ADDGAN augments with the ambient measurement loop.","marker":"[13]"},{"why":"Supplies the DDPM baseline and the forward/reverse diffusion formalism used for training and comparison.","marker":"[14]"},{"why":"Supplies the clinical CT images used as ground-truth objects in the stylized CT study.","marker":"[16]"},{"why":"Supplies the differentiable Radon transform and filtered back-projection operators used to simulate CT measurements in the training loop.","marker":"[17]"},{"why":"Supplies the DBT images used as ground-truth objects for the tomosynthesis study.","marker":"[18]"},{"why":"Defines the StyleGAN3 architecture underlying the Ambient StyleGAN3 baseline.","marker":"[20]"}],"fun_headline_variants":["ADDGAN turns noisy scans into clean object models","Diffusion GAN learns object models from noisy CT and DBT","ADDGAN beats AmbientGAN on noisy-image object models","Ambient DDGAN establishes object models from noisy images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on assuming that the simulated measurements H(x̂0) generated inside the network are statistically indistinguishable, including their noise, from the real measured images y0; for CT the real noise enters the sinogram and is never sampled in the generator path, so if a discriminator can detect that difference, the learned object distribution is biased.","fun_headline_variants_meta":{"raw":{"variants":["ADDGAN turns noisy scans into clean object models","Diffusion GAN learns object models from noisy CT and DBT","ADDGAN beats AmbientGAN on noisy-image object models","Ambient DDGAN establishes object models from noisy images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00056,"raw_usage":{"total_tokens":2699,"prompt_tokens":1019,"completion_tokens":1680,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":1615}},"tokens_in":635,"tokens_out":1680,"duration_ms":11730,"temperature":1.0,"reasoning_tokens":1615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:21:10.927534+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ADDGAN on CT data whose sinogram noise is Poisson-distributed rather than the i.i.d. Gaussian used in the paper, then measure the Hotelling-observer AUC on generated versus ground-truth patches; if the AUC gap grows or the discriminator can reliably tell real from simulated measurements, the ambient noise model is insufficient and the 'noise-independent SOM' claim fails for realistic noise.","supporting_citations":[{"cited_title":"Detection of masses and architectural distortions in digital breast tomosynthesis: a publicly available dataset of 5,060 patients and a deep learning model","cited_arxiv_id":"2011.07995","evidence_quote":"Supplies the DBT images used as ground-truth objects for the tomosynthesis study."},{"cited_title":"Alias-freegenerative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Defines the StyleGAN3 architecture underlying the Ambient StyleGAN3 baseline."},{"cited_title":"Deeplesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the clinical CT images used as ground-truth objects in the stylized CT study."},{"cited_title":"Ambientgan: Generative models from lossy measurements,","cited_arxiv_id":null,"evidence_quote":"Introduces the AmbientGAN idea of training a generator through a known measurement operator on noisy data, which ADDGAN adapts to diffusion."},{"cited_title":"Learning stochastic object models from medical imaging measurements by use of advanced ambient generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the advanced AmbientGAN baselines (Ambient StyleGAN3 and ProAmGAN) and the SOM-from-measurements framework that ADDGAN compares against."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the DDPM baseline and the forward/reverse diffusion formalism used for training and comparison."}],"review_version":1}