{"id":"03c0e2a4-0fb8-4d3e-b4fe-0a2b70236beb","arxiv_id":"2412.16979","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CDEIT, a diffusion model that denoises conductivity images conditioned on boundary voltages, outperforms prior EIT reconstruction methods on clean simulated data and transfers to real data via a voltage-current normalization.","lead":"The paper applies a conditional diffusion model to electrical impedance tomography, reconstructing internal conductivity images directly from boundary voltage measurements, and reports higher scores than prior methods on clean simulated data. It also introduces a normalization scheme intended to let models trained on simulations work on real tank data with different currents and background conductivities.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. IV normalization derivation is not a valid transfer law: it asserts current scaling equality without justification, mixes dimensional quantities, and calibrates SU on target-domain reference voltages, so the real-data generalization claim is unsupported.","rationale":"Good-faith reading: the core CDEIT construction, a conditional DDPM with voltage-conditioned denoising, Eq. (21), and DDIM sampling, is standard, and the simulated benchmark in Table I supports a real, though modest, improvement over DHU-Net. The open-source implementation is a strong reproducibility asset. The reader's conditional verdict is appropriate. My stress-test converges on the same weakest point: the generalization contribution. The normalization is not an innocent approximation; it is internally inconsistent as written. The current scaling factor is asserted, not derived, and the voltage scaling factor is calibrated using the target domain's own reference data. A controlled simulation experiment with known transfer parameters can settle whether the formula is even approximately correct. If it fails, the paper should be revised to drop or substantially weaken the generalization claim, while the simulated-data claim can stand. I therefore recommend keeping the verdict at CONDITIONAL: the methodological core is defensible, but the transfer claim needs either a corrected derivation with defined quantities and a validation experiment, or explicit removal from the abstract.","tokens_in":19230,"tokens_out":5812,"duration_ms":57815,"concrete_test":"Run a controlled sim-to-sim transfer experiment: train CDEIT on the paper's simulated dataset (background 1 S/m, 1 mA, radius R0), then generate a matched EIDORS dataset with the same inclusion masks but background 2 S/m, excitation 2 mA, and radius 2R0. Apply Eq. (32) using SU from reference voltages and compare the reconstructed δσ against ground truth. If normalized PSNR/SSIM are not close to the within-distribution test values (within about 2 dB PSNR), the transfer law is invalid. Repeat with nonzero contact impedance z_l in Eq. (2) to test whether the simplified resistor model is the cause of the failure.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"CDEIT's third contribution, sim-to-real generalization, is carried entirely by Sec. IV, and that derivation does not hold together. Eq. (26) is a resistance formula for two contact points in an extended conductor; it is not the response of a finite 16-electrode domain with the complete electrode model and contact impedances z_l in Eq. (2). The subsequent equations I00=U00σ00 and I01=U01σ01 treat σ as a scalar conductance, but σ in the paper is a conductivity distribution, so the product Uσ is not a well-defined current between measurement electrodes, which in EIT draw essentially no current. The key assertion SI=I10/I00=I11/I01, stated 'due to linearity,' is exactly the property that needs proof: it assumes the unknown domain mismatch scales identically in the background and inclusion frames even though the forward map f(σ) is nonlinear. Furthermore, SU is computed as Ur/Us from the reference voltage of the target real dataset, so the transfer is calibrated on the deployment domain rather than derived from training-domain quantities; calling this 'directly applied... without retraining' overstates what was tested. The real-data experiments evaluate only four hand-picked samples against manually labeled reference images, so if the normalization is wrong, the claimed real-world generalization is unsupported. The simulated-data SOTA claim rests on Table I and is not affected by this concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDEIT, a conditional denoising diffusion model for electrical impedance tomography (EIT) image reconstruction. The model adds Gaussian noise to clean conductivity images in a forward diffusion process and trains a Transformer-based U-Net to reverse this process conditioned on boundary voltage measurements, using a standard DDPM-style loss with an L1 simplification. Inference is performed with a DDIM sampler using only five steps. The authors report state-of-the-art performance on a simulated dataset (PSNR 39.57, SSIM 0.998, CC 0.999) and propose a normalization procedure in Section IV intended to allow models trained on simulated data to be applied to real datasets with different sizes, excitation currents, and background conductivities. Experiments on two real datasets (UEF2017 and KTC2023) are presented, but with only four hand-picked samples and manually labeled references.","tokens_in":19527,"tokens_out":2627,"duration_ms":24110,"significance":"If the simulated-dataset results hold, the paper provides a useful and well-documented application of conditional diffusion models to EIT, with open-source code and a standard derivation of the DDPM objective. The reported 0.81 dB PSNR improvement over DHU-Net on the test set is a modest but consistent gain, and the ablation and complexity analyses are reasonable. However, the third contribution, sim-to-real generalization, is not established: the normalization derivation in Section IV is not physically valid as presented, and the real-data evaluation is too limited to support the claimed generalization. The abstract's blanket claim of outperforming state-of-the-art methods is also contradicted by the paper's own noise-robustness results in Table IV.","major_comments":[{"comment":"The abstract states that the proposed model 'outperforms state-of-the-art methods' without qualification, but Table IV shows that at 30 dB Gaussian noise the CDEIT PSNR is 26.70, below DHU-Net (28.94) and Ec-Net (28.78). The text in Sec. V-E3 explicitly acknowledges this degradation. The SOTA claim must be restricted to the noise-free or low-noise setting, or the claim should be removed from the abstract.","section":"Abstract, Table IV, Sec. V-E3"},{"comment":"The normalization derivation is not a valid transfer law. Eq. (26) is a two-point resistance formula for an extended conductor and does not describe the response of a 16-electrode setup governed by the complete electrode model with contact impedances z_l in Eq. (2). Eqs. (27)-(28) write I = Uσ, but σ is a conductivity distribution, not a scalar conductance, and in EIT the measurement electrodes draw essentially no current; the product Uσ is not the current flowing between measurement electrodes. The key assertion SI = I10/I00 = I11/I01, attributed to 'linearity', is exactly the property that needs proof, since the forward map f(σ) is nonlinear and the domain mismatch may scale differently in the background and inclusion frames. Finally, SU is computed from reference voltages of the target real dataset (text following Eq. (32)), so the claimed 'prediction' on real data is calibrated on the deployment domain rather than derived from training-domain quantities; calling this 'directly applied... without retraining' overstates what was actually tested.","section":"Sec. IV, Eqs. (26)-(32)"},{"comment":"The real-data validation uses only four hand-picked samples with manually labeled reference images, and the text states that the quantitative indicators are 'shown for reference purposes only'. This does not provide a quantitative test of the normalization procedure; the sim-to-real generalization claim rests entirely on the derivation in Sec. IV, which has the problems noted above. A more systematic evaluation on the full real datasets, or at least a clearly stated limitation, is needed.","section":"Sec. V-D, Fig. 10, Fig. 11"}],"minor_comments":[{"comment":"Eq. (20) retains a negative coefficient from the KL divergence in Eq. (18), but the simplified loss in Eq. (21) is a positive L1 norm. Since the L1 loss is what is actually optimized, this is a presentation inconsistency rather than a fatal error, but the sign should be clarified.","section":"Sec. III-D, Eq. (20)"},{"comment":"There is a typo, 'bilion' instead of 'billion', in the complexity discussion.","section":"Sec. V-E1, Table II"},{"comment":"The conclusion contains a typo, 'generalizion' instead of 'generalization'.","section":"Sec. VI"},{"comment":"The reference list has inconsistent author name spellings, e.g., 'Husain' in [14] versus 'Hussain' in [6] and [7], which should be unified.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's central simulated benchmark is sound and the code is provided, but the third contribution, sim-to-real transfer, is not supported by the current derivation and evaluation. The overclaim in the abstract relative to Table IV also needs to be addressed. I would recommend major revision rather than rejection because the simulated results and the diffusion framework are defensible; however, the authors should either substantially strengthen the real-data validation or explicitly limit the generalization claim to the specific datasets and conditions tested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper does something genuinely new in the EIT diffusion literature — it conditions the reverse diffusion process directly on the boundary voltage vector, whereas the two cited diffusion-based EIT works (CSD [43] and [44]) use unconditional priors or classical initial guesses. The DDPM machinery in Sec. III is standard and correctly derived, and the simulated-data result is a real, if modest, improvement (PSNR 39.57 vs 38.76 for DHU-Net). Credit where due: code is open-source, the ablation and complexity tables are informative, and the main benchmark evaluation is clean.\n\nThe soft spots are in the claims, not the core denoising network. The abstract says the model 'outperforms state-of-the-art methods' without qualification, but their own Table IV shows CDEIT at 26.70 PSNR under 30 dB noise, below both DHU-Net (28.94) and Ec-Net (28.78). That is a direct contradiction of the blanket claim. The second issue is bigger. Section IV's normalization procedure is supposed to let a simulation-trained model transfer to real tanks with different geometry, current, and background conductivity. But the derivation does not hold together. Eq. (26) is the resistance between two contact points in an extended conductor; it says nothing about a 16-electrode tank with contact impedances. Then Eqs. (27)–(28) treat σ as a scalar conductance in products like Uσ, but σ is a conductivity distribution. The key assumption SI = I10/I00 = I11/I01, justified as 'due to linearity,' is exactly what needs proof, and the EIT forward map is nonlinear. Finally, SU is computed from the average reference voltage of the target real dataset, so the 'prediction' on real data is calibrated on the very test domain. Calling this 'directly applied without retraining' overstates what was done. The real-data evaluation compounds the problem: four hand-picked samples with manually labeled references.\n\nSo: the core method is defensible, the generalization claim is not, at least not as derived. The paper deserves a serious referee, but it needs major revision. I would send it to review, and in the report I would ask the authors to (1) scope the SOTA claim to clean/low-noise data, (2) reframe Sec. IV as a heuristic with clear limitations or redo it with proper CEM quantities, and (3) report real-data metrics with error bars and a larger sample. I will bring it to the reading group because it is a good example of how diffusion conditioning can help in inverse problems, and also of how a weak transfer argument can undermine an otherwise reasonable paper.","headline":"A solid, modest application of conditional diffusion to EIT with a real SOTA bump on clean simulated data, but the sim-to-real normalization is a calibration-based heuristic, not a derivation, and the abstract overclaims.","tokens_in":20099,"tokens_out":2948,"would_cite":true,"duration_ms":26522,"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":"Conditional diffusion model reconstructs EIT images from boundary voltages and beats all compared methods on simulated tests.","keywords":["electrical impedance tomography","image reconstruction","conditional diffusion model","denoising diffusion probabilistic model","Transformer U-net","DDIM sampling","ill-posed inverse problem","domain generalization"],"falsifier":"Withhold the reference voltage $U_{10}$ used to compute $S_U$ from a real dataset, derive $S_U$ instead from the forward model or from a separately calibrated phantom, and run the reconstructions; if accuracy collapses, the claimed parameter-free domain transfer is refuted. A second check is already in the paper: at 30 dB input noise, CDEIT's PSNR drops below DHU-Net and Ec-Net, so the superiority claim is restricted to clean or mildly noisy measurements.","tokens_in":18978,"feed_emoji":"⚡","tokens_out":10165,"duration_ms":80945,"temperature":0.7,"pith_summary":"This paper proposes CDEIT, a generative model that reconstructs cross-sectional conductivity images from electrical impedance tomography (EIT) boundary-voltage measurements. Instead of regressing an image from voltage data in one step, CDEIT adds Gaussian noise to a clean conductivity image and then learns to denoise it back to a clean image while conditioning on the measured boundary voltages at every step. The authors argue that this multi-step, voltage-conditioned denoising implicitly learns a prior over conductivity images and refines spatial details progressively, which is why it reports higher accuracy than single-step deep-learning and classical methods on a simulated test set. The paper also claims that a voltage-and-current normalization procedure lets a model trained on simulated data be applied to real water-tank datasets with different sizes, excitation currents, and background conductivities without retraining.","feed_headline":"Voltage-conditioned diffusion model outdoes EIT baselines","feed_subtitle":"CDEIT turns boundary voltages into conductivity maps via iterative denoising and generalizes from simulated to real data.","key_machinery":"The central objects are the forward diffusion chain $\\sigma_t = \\sqrt{\\bar{\\alpha}_t}\\,\\sigma_0 + \\sqrt{1-\\bar{\\alpha}_t}\\,\\epsilon$, which turns a clean conductivity image into Gaussian noise over $T=1000$ steps; the conditional reverse network $\\sigma_\\theta(\\sigma_t, U, t)$, a Transformer-based U-net with global and windowed Swin attention, time-step embedding, patch merging, and skip connections, which estimates the clean image from the noisy image and the upsampled boundary-voltage map; the DDIM sampler, which allows the reverse process to run in five steps; and the normalization rule $\\delta\\hat{\\sigma} = g\\left(\\frac{U_{11}-U_{10}}{S_U}\\right)\\cdot \\frac{S_I}{S_U}$, which rescales real-world voltage differences by factors derived from a simplified resistor model, an assumed linear current scaling, and a voltage ratio measured from the target dataset's reference voltages.","core_discovery":"The central discovery is that a denoising diffusion probabilistic model, re-purposed as a conditional generator, can solve the EIT inverse problem by modeling the conditional distribution $p(\\sigma|U)$ of conductivity images given boundary voltages. In training, Gaussian noise is added to clean conductivity images, and a Transformer-based U-net learns to predict the original image from the noisy image together with the voltage measurements and the diffusion time step. At inference, a DDIM sampler runs the learned reverse process in as few as five steps, starting from pure noise and the measured voltages, to generate the conductivity image. On the simulated benchmark the method reaches a PSNR of 39.57 dB, SSIM of 0.998, and correlation coefficient of 0.999, outperforming all eight compared methods, including another diffusion-based approach that is not conditioned on voltages. The paper further claims that the model transfers to real water-tank data through a normalization procedure that compensates for differences in excitation current and background conductivity.","pith_inferences":["The conditional-diffusion formulation naturally supports uncertainty quantification: sampling several reconstructions from the same voltage vector would yield per-pixel variance estimates, which the paper does not compute.","The normalization procedure compensates only for global scalings of current and conductivity; extending it to electrode contact impedance or non-circular domains would require a richer transfer model and would directly test the simplified resistor assumption.","The 30 dB noise result suggests that training the denoiser with voltage noise, or conditioning on an estimated noise level, could make generative refinement robust where single-step regression currently wins.","If $S_U$ could be estimated from the forward model instead of from the target dataset's own reference voltages, the simulation-to-real transfer would become truly parameter-free; as presented, the scale factor is calibrated on the test domain."],"forward_implications":["On the simulated test set, CDEIT achieves PSNR 39.57 dB, SSIM 0.998, and CC 0.999, outperforming all compared conventional and deep-learning methods.","The DDIM sampler with five reverse steps keeps most of the quality gain while bringing inference to 1.55 GFLOPs, less than several lighter-parameter baselines.","The proposed normalization lets a model trained on simulated data produce usable reconstructions on two real water-tank datasets with different sizes, currents, and background conductivities, without retraining.","At low noise (40 dB), CDEIT keeps the top PSNR, but at 30 dB the single-step baselines DHU-Net and Ec-Net overtake it, bounding the reported advantage to clean or mildly noisy measurements."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model formulation that CDEIT adapts by conditioning the reverse process on voltage data.","marker":"[21]"},{"why":"Provides the DDIM sampler that lets CDEIT run the reverse process in five steps.","marker":"[52]"},{"why":"Supplies the Transformer attention mechanism used in the denoising U-net backbone.","marker":"[47]"},{"why":"Supplies the windowed Swin attention blocks used for multi-scale feature extraction in the U-net.","marker":"[48]"},{"why":"Establishes conditional diffusion for image restoration, the template for conditioning generation on boundary voltages.","marker":"[24]"},{"why":"Provides the finite-element simulation environment used to create the paired conductivity-voltage training data.","marker":"[54]"},{"why":"Provides the first real water-tank EIT dataset used to test the claimed simulation-to-real generalization.","marker":"[56]"},{"why":"Provides the second real EIT dataset, with reference conductivity masks, used to test the same generalization claim.","marker":"[57]"}],"fun_headline_variants":["Diffusion model reconstructs EIT images from voltages","Conditional diffusion outshines EIT reconstruction baselines","Voltage-guided denoising yields high-quality EIT maps","CDEIT: diffusion-based EIT outperforms eight baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the normalization formula, with the current scale factor $S_I$ assumed linear and the voltage scale factor $S_U$ measured from the target dataset's reference voltages, correctly maps real voltage differences into the simulation's units; if the simplified resistor model $R \\approx \\frac{2}{\\pi\\sigma}\\ln\\frac{2d}{\\delta}$ fails under electrode contact impedance and nonlinear current distribution, the reported real-data generalization is not parameter-free.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model reconstructs EIT images from voltages","Conditional diffusion outshines EIT reconstruction baselines","Voltage-guided denoising yields high-quality EIT maps","CDEIT: diffusion-based EIT outperforms eight baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1513,"prompt_tokens":1050,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":394}},"tokens_in":666,"tokens_out":463,"duration_ms":4361,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:54:46.474500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Withhold the reference voltage $U_{10}$ used to compute $S_U$ from a real dataset, derive $S_U$ instead from the forward model or from a separately calibrated phantom, and run the reconstructions; if accuracy collapses, the claimed parameter-free domain transfer is refuted. A second check is already in the paper: at 30 dB input noise, CDEIT's PSNR drops below DHU-Net and Ec-Net, so the superiority claim is restricted to clean or mildly noisy measurements.","supporting_citations":[{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion probabilistic model formulation that CDEIT adapts by conditioning the reverse process on voltage data."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Supplies the Transformer attention mechanism used in the denoising U-net backbone."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows,","cited_arxiv_id":null,"evidence_quote":"Supplies the windowed Swin attention blocks used for multi-scale feature extraction in the U-net."},{"cited_title":"Image super-resolution via iterative refinement,","cited_arxiv_id":null,"evidence_quote":"Establishes conditional diffusion for image restoration, the template for conditioning generation on boundary voltages."},{"cited_title":"Eidors: Towards a community-based extensible software base for eit,","cited_arxiv_id":null,"evidence_quote":"Provides the finite-element simulation environment used to create the paired conductivity-voltage training data."},{"cited_title":"Kuopio tomography challenge 2023 – electrical impedance tomography competition and open dataset,","cited_arxiv_id":null,"evidence_quote":"Provides the second real EIT dataset, with reference conductivity masks, used to test the same generalization claim."}],"review_version":1}