{"id":"74135629-e787-4f0b-9cc8-5e04ee210f9c","arxiv_id":"2412.03617","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TriPLET couples sinogram denoising, wavelet-domain reconstruction, and adversarial image evaluation to reconstruct standard-dose PET from low-dose PET sinograms, reporting small but statistically significant improvements over prior state-of-the-art methods.","lead":"This paper introduces TriPLET, a deep learning framework that reconstructs standard-dose PET images from low-dose PET sinograms by combining denoising in the projection domain, wavelet-domain reconstruction, and image-domain adversarial evaluation. The method is tested on 70 paired chest-abdomen PET exams and shows modest but consistent metric gains over prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on synthetic sinograms: inverse-Radon transforms of reconstructed images, not real raw LPET sinograms; projection-domain gains may not transfer.","rationale":"I read the paper in good faith and find the architecture coherent, the ablation study useful, and the external validation a positive step toward generalization. The decisive weakness is not internal inconsistency but the fidelity of the projection-domain training data. The method is advertised as reconstructing SPET from LPET sinograms, yet no real sinogram is ever used: the sinograms are inverse-Radon transforms of reconstructed image patches, and the authors explicitly acknowledge this in Section 4.1. Because the projection-domain noise statistics are precisely where the LPET/SPET difference originates, the central quantitative claim in Table 3 is tested on a proxy. The margin over the strongest baseline is small (0.04 dB PSNR, 0.001 SSIM, 0.029 rRMSE), which makes the proxy issue more consequential: if the synthetic sinogram domain is even slightly easier or differently structured than real raw data, the reported superiority may not persist. The concrete test I propose is direct and would settle the matter: retrain on real raw sinograms and compare. I agree with the reader that the appropriate verdict is CONDITIONAL, pending raw-sinogram validation and code release; my stress-test does not change that verdict.","tokens_in":24161,"tokens_out":5550,"duration_ms":54768,"concrete_test":"Retrain TriPLET and the strongest baseline TriDoRNet on real raw sinograms (or faithful list-mode simulations with matched Poisson noise) from the same uEXPLORER scanner, using the 70-subject 120s-vs-1200s paired data and physically consistent whole-sinogram or context-aware patch extraction. Then compare image-domain PSNR/SSIM/rRMSE. If TriPLET no longer reproducibly outperforms TriDoRNet by a margin larger than the reported 0.04 dB PSNR, or if its absolute metrics drop substantially relative to Table 3, the pseudo-sinogram training is the reason the central claim does not extend to clinical raw data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that the 'LPET sinograms' fed to DenNet are representative of actual low-dose PET raw data. Section 4.1 states plainly: 'we did not collect PET raw data (i.e., sinograms). The sinograms indirectly obtained from reconstructed PET images may indeed differ from the real PET raw data.' All projection-domain training and evaluation uses sinogram patches obtained by inverse Radon transform of reconstructed image patches. This is not a faithful proxy: reconstructed OSEM images already incorporate attenuation/scatter/random corrections, TOF modeling, and iterative regularization, so their inverse Radon transforms do not share the Poisson counting statistics, detector-response correlation, or truncation artifacts of raw list-mode sinograms. Moreover, the Radon transform is only lossless for full-field, full-angle data; cropping an image and applying inverse Radon yields a synthetic global sinogram of the patch with zero outside, not a local patch of the true sinogram. Therefore the improvement in Table 3 over TriDoRNet (PSNR +0.04 dB, SSIM +0.001, rRMSE -0.029) may reflect denoising of a synthetic projection domain that does not transfer to the real sinogram domain. The external validation does not close this gap because it also relies on reconstructed-image data converted to sinograms by the same inverse-Radon pipeline. The reported p<0.05 is also insufficiently documented (no per-fold values or test details), so the significance of the small margin is unverifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TriPLET, an end-to-end framework for reconstructing standard-dose PET (SPET) images from low-dose PET (LPET) sinograms. The method couples three subnetworks: a CNN-Transformer denoising network operating on sinograms, a wavelet-transform-based reconstruction network, and a paired adversarial network in the image domain, with losses in the projection, wavelet, and image domains. The authors evaluate TriPLET on 70 paired chest-abdomen uEXPLORER scans (converted to sinogram patches via inverse Radon transforms of reconstructed images) and on an external Ultra Low Dose PET challenge dataset, reporting the highest PSNR/SSIM and lowest rRMSE among six baselines, including their own prior TriDoRNet. The paper includes an ablation study of each component and acknowledges that the sinograms are not real raw data.","tokens_in":24429,"tokens_out":5298,"duration_ms":45147,"significance":"If the projection-domain results transfer to real raw sinograms, the hybrid denoising-and-reconstruction design is a sensible way to reduce noise propagation in low-dose PET, and the triple-domain supervision is a useful methodological contribution. The paper is transparent about its main limitation, ships a clear ablation chain, uses five-fold cross-validation, and provides an external dataset evaluation. However, the central quantitative claim is only as strong as the fidelity of the synthetic sinograms; the reported gain over TriDoRNet is small (PSNR +0.04 dB, SSIM +0.001, rRMSE -0.029), and the external validation reuses the same synthetic-sinogram pipeline, so it does not independently confirm that the projection-domain denoising generalizes to real raw data.","major_comments":[{"comment":"The LPET sinograms used for training and testing are obtained by applying the inverse Radon transform to reconstructed image patches, not by acquiring raw list-mode or sinogram data. As the authors explicitly note, \"sinograms indirectly obtained from reconstructed PET images may indeed differ from the real PET raw data.\" Reconstructed OSEM images already incorporate attenuation/scatter/random corrections, TOF modeling, and iterative regularization, so their inverse Radon transforms do not reproduce the Poisson counting statistics, detector-response correlation, or truncation characteristics of true low-dose raw sinograms. Because the projection-domain denoising loss L_P is the mechanism that distinguishes TriPLET from image-domain-only methods, the reported improvements in Table 3 over TriDoRNet may not transfer. The external validation in Section 5.1 also uses the same inverse-Radon conversion, so it does not close this gap. The central claim therefore needs either validation on real raw sinogram data or a more realistic forward simulation (e.g., re-projection with Poisson noise) to be convincing.","section":"Section 4.1, Fig. 4"},{"comment":"The claim that TriPLET is \"significantly better\" than all baselines with p < 0.05 is not adequately documented. It is unclear whether the paired t-test was computed on the five fold-level means, per-subject metrics, or per-patch metrics; no p-values, test statistics, or degrees of freedom are reported. Since the gain over TriDoRNet is small relative to the reported standard deviations (Table 3: 25.932 ± 0.671 vs 25.892 ± 0.784 for PSNR), the significance claim is load-bearing and must be supported with exact details of the test procedure.","section":"Section 4.4, \"Quantitative Comparison\""},{"comment":"The statement that \"the sinogram and image can be losslessly converted ... hence this method of obtaining sinogram patches does not introduce additional errors\" is misleading for patch-based data. Applying the inverse Radon transform to a zero-padded 96×96×96 image patch produces the global sinogram of that patch (with zero attenuation outside the patch), not a local patch of the true scanner sinogram; the two are not equivalent because the integration paths and detector geometry differ. This affects both the denoising network's input distribution and the projection-domain loss. The authors should clarify the relationship to the full-field sinogram or extract patches directly in the sinogram domain for a true correspondence.","section":"Section 4.1, Fig. 4"}],"minor_comments":[{"comment":"The phrase \"normalized to [0, 1] using Z-score normalization\" is contradictory because Z-score normalization does not bound values to [0, 1]; please describe the actual preprocessing (e.g., Z-score followed by min-max scaling).","section":"Section 4.1"},{"comment":"The description \"filter each PET image's values (i.e., only retaining values within the 5% to 95% range)\" is actually a clipping or truncation operation, not a filter; the terminology should be corrected for clarity.","section":"Section 4.1"},{"comment":"The number of CNN and Transformer blocks is stated inconsistently: the text first describes seven blocks with {1st, 3rd, 5th, 7th} as CNN and {2nd, 4th, 6th} as Transformer (four CNN, three Transformer), then says \"an alternating combination of 3 CNN blocks and 3 Transformer blocks.\" Please reconcile these counts.","section":"Section 3.1"},{"comment":"The caption refers to \"Table 1\" when describing the methods shown, but the ablation results are presented in Table 2; this cross-reference should be corrected.","section":"Figure 5 caption"},{"comment":"The notation for rRMSE uses a symbol in the denominator that is described only in words as \"the all-voxel-averaged value of the ground-truth image.\" Define the symbol explicitly and confirm whether the same ground-truth maximum is used for PSNR in Eq. (6).","section":"Section 4.2, Eq. (5)"},{"comment":"The external validation describes 40 pairs but gives no details about whether the same value clipping, normalization, and patch-extraction protocol were applied to the external data; a brief description would help assess the fairness of the cross-dataset comparison.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The contribution is incremental over the authors' own TriDoRNet, and the reported improvement is small. The main risk is that the sinogram-based claim is validated only on synthetic sinograms; I would encourage the editor to require either raw-data validation or a realistic simulation before acceptance. The paper's honesty about the limitation is commendable, but the external validation does not remedy the gap because it reuses the same inverse-Radon pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper you'll want to know: Jiang et al. build a three-stage PET enhancement pipeline (sinogram denoising with a CNN-transformer hybrid, wavelet-domain reconstruction, adversarial image-pair evaluation) trained end-to-end. It is a direct extension of their TriDoRNet, and the differences are clearly itemized: Swin-style transformer blocks in the denoiser, a PatchGAN-like critic that takes LPET/SPET pairs, GradNorm loss weighting, and new external validation. The ablations show each piece buys something, and the reported numbers beat the prior art on their internal 70-subject uEXPLORER dataset, though the margin over TriDoRNet is small (PSNR +0.04 dB, SSIM +0.001).\n\nThe strongest thing I can say for the paper: it is honest. Section 4.1 states plainly that no real raw sinograms were collected, and that the sinograms are obtained by inverse Radon transform of reconstructed image patches. That sentence puts the central claim in perspective. The denoising network is trained and evaluated on a proxy domain. Reconstructed OSEM images already contain attenuation/scatter correction, TOF modeling, and iterative regularization; their inverse Radon transforms do not carry the Poisson counting statistics or detector-response correlations of true list-mode sinograms. Worse, as the stress-test note observes, the patch-based inverse Radon is not equivalent to extracting a sinogram patch -- it is a global transform of a cropped image with zero padding. So the projection-domain gains may not transfer to real raw data. The external validation does not close this gap, because the same inverse-Radon pipeline generates those sinograms too.\n\nThat said, the reader's concern about circularity is not the issue: the loss functions are supervised by real SPET images, and the external dataset provides an independent benchmark on different scanners. The paper is not circular; the limitation is one of domain mismatch.\n\nThe main soft spots beyond the raw-data issue: the statistical significance claims are underdocumented for the small margins (no per-fold numbers, no description of the paired test), and the improvement over TriDoRNet is at the edge of what could be noise. These are fixable in revision.\n\nBottom line: this is a modest but competent incremental contribution. The framework is sensible and the ablation is internally consistent. For a reader working on sinogram-domain PET reconstruction, the paper is worth a look, but mainly as a demonstration that synthetic sinograms are an unreliable proxy. I'd recommend sending it to peer review -- a serious referee can push for real raw-data validation and better statistical reporting. It's not a paper I'd cite in my own work yet.","headline":"Solid incremental extension of the authors' TriDoRNet with a clear, honest statement of its main limitation: all sinograms are synthesized from reconstructed images, so the raw-data claim is untested.","tokens_in":24977,"tokens_out":2164,"would_cite":false,"duration_ms":20266,"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 claims that TriPLET, a hybrid denoising-and-reconstruction framework operating on sinograms, wavelet frequency maps, and images, reconstructs standard-dose PET images from low-dose PET sinograms with higher similarity and…","keywords":["low-dose PET","standard-dose PET reconstruction","sinogram denoising","triple-domain framework","discrete wavelet transform","Transformer","generative adversarial network","Radon transform"],"falsifier":"Train TriPLET on true list-mode low-dose raw sinograms paired with standard-dose reconstructions from a clinical scanner, then test on held-out raw data and compare PSNR, SSIM, and rRMSE with the reported 25.932 dB, 0.992, and 0.296; a substantial drop would show that the simulated-sinogram assumption, not the triple-domain design, drove the reported gains.","tokens_in":1815,"feed_emoji":"🩻","tokens_out":2046,"duration_ms":67262,"temperature":0.7,"pith_summary":"TriPLET is an end-to-end framework for converting low-dose PET sinograms into standard-dose PET images. The paper's thesis is that this conversion is best done as a hybrid denoising-and-reconstruction process that works in three coupled domains: a denoising network cleans the LPET sinogram in the projection domain before any reconstruction happens; a wavelet-domain reconstruction network then builds the image with explicit frequency sub-band supervision; and an adversarial network judges the result in the image domain. The authors argue that a significant part of the LPET-to-SPET gap lives in projection-domain noise, so removing noise before reconstruction matches the physics of PET and stops noise from being amplified later. On their chest-abdomen dataset they report the highest similarity and signal-to-noise ratio against six comparison methods, with PSNR of 25.932 ± 0.671 dB, SSIM of 0.992 ± 0.004, and rRMSE of 0.296 ± 0.213 on the internal test set, and top results on an external dataset. If the claim holds, the practical consequence is that patients could receive one-tenth of the usual radiotracer dose while reconstructed images stay close to standard-dose quality.","feed_headline":"Triple-domain net reconstructs near-standard PET from low-dose sinograms","feed_subtitle":"Cascading sinogram denoising, wavelet reconstruction, and adversarial checks hits 25.9 dB PSNR on chest-abdomen PET.","key_machinery":"The machinery is the triple-domain pipeline itself, coupled end to end: DenNet turns an LPET sinogram into a denoised sinogram via residual CNN and windowed-transformer blocks; RecNet receives the denoised sinogram, passes it through an inverse Radon transform, and reconstructs the image in the wavelet domain using discrete wavelet and inverse wavelet transforms in place of pooling and upsampling; AdvNet evaluates the predicted SPET image against the real one in the image domain. The Radon transform and its inverse provide a lossless bridge between sinogram and image, and the discrete wavelet transform provides a lossless decomposition into eight sub-band frequency maps, so high-frequency details can be supervised separately. Three losses are aligned to the domains: projection-domain MSE ($\\mathcal{L}_P$), focal frequency loss over wavelet sub-bands ($\\mathcal{L}_F$), and image-domain MSE plus paired adversarial loss ($\\mathcal{L}_I$), combined with GradNorm. The design choice the whole argument rests on is that noise in low-dose PET is most visible in the sinogram and best removed there, before reconstruction.","core_discovery":"The central discovery this paper advances is that a three-network cascade—DenNet, RecNet, and AdvNet—can reconstruct SPET images from LPET sinograms at the state of the art when each network is matched to the domain where its effect is strongest. DenNet uses alternating CNN and windowed-transformer blocks with residual learning to suppress sinogram noise; RecNet replaces the usual pooling and upsampling of a 3D U-Net with discrete wavelet and inverse wavelet transforms, decomposing the image into eight sub-band frequency maps and reconstructing it from them; AdvNet discriminates real from predicted image pairs, conditioning on the LPET image. The authors show by ablations that each added component—wavelet reconstruction, sinogram denoising, frequency loss, and adversarial evaluation—improves PSNR, SSIM, or rRMSE, and that the complete framework beats all six comparison methods with p < 0.05.","pith_inferences":["An implication the paper leaves implicit is that the simulated-sinogram-to-real-raw-data gap could be small enough that the same architecture would train directly on list-mode raw data, potentially extending to dynamic or total-body PET where noise statistics change across frames.","Because the sinogram patches are generated through the inverse Radon transform, the same three-domain recipe could transfer to other tomographic modalities whose raw data are sinogram-like, such as SPECT.","A direct testable extension is to train TriPLET on real raw sinograms paired with standard-dose reconstructions and compare against the reported 25.932 dB PSNR; the authors acknowledge performance may vary without quantifying how much.","The reported gains over the earlier TriDoRNet are small, so a reader should treat the transformer-based denoiser and adversarial network as refinements whose marginal value is measured mainly by the ablation study rather than by large leaps in absolute metrics."],"forward_implications":["If TriPLET holds, clinical low-dose PET protocols could reduce injected radiotracer activity to one-tenth of the standard dose while images remain quantitatively close to standard-dose reconstructions.","Denoising the sinogram before reconstruction should suppress projection-noise amplification, so the hybrid ordering should matter most at very low doses where raw-data noise dominates.","Wavelet-domain supervision should preserve high-frequency boundaries and small structures better than pure image-domain GANs, reducing over-smoothing and edge blur.","The end-to-end design means a trained network produces a reconstructed SPET image in a single forward pass, without iterative OSEM-style reconstruction at inference time.","External validation on a different scanner suggests the benefit transfers across PET scanners, though with a measurable performance drop."],"supporting_citations":[{"why":"Supplies the prior TriDoRNet triple-domain architecture that TriPLET extends with a transformer-based denoiser, an adversarial network, and GradNorm.","marker":"Jiang, Pan and Shen, 2023"},{"why":"Establishes the lossless Radon/inverse Radon conversion used to derive sinogram patches from image patches.","marker":"Toft, 1996"},{"why":"Provides the multi-level wavelet-CNN idea that motivates replacing pooling and upsampling with wavelet transforms in RecNet.","marker":"Liu, Zhang, Zhang, Lin and Zuo, 2018"},{"why":"Supplies the shifted-window multi-head self-attention design used in DenNet's transformer blocks.","marker":"Liu et al., 2021"},{"why":"Defines the focal frequency loss used as the wavelet-domain constraint $\\mathcal{L}_F$.","marker":"Jiang, Dai, Wu and Loy, 2021"},{"why":"Provides GradNorm for automatically balancing the multi-loss training across the three stages.","marker":"Chen, Badrinarayanan, Lee and Rabinovich, 2018"},{"why":"Supplies the conditional adversarial discriminator design on which AdvNet is based.","marker":"Isola, Zhu, Zhou and Efros, 2017"},{"why":"Provides the Sinogram-Net baseline that, like TriPLET, reconstructs SPET images directly from LPET sinograms.","marker":"Feng and Liu, 2020"},{"why":"Serves as the cGAN baseline that TriPLET must beat in the quantitative comparison.","marker":"Wang et al., 2018"}],"fun_headline_variants":["Triple-domain PET boost: low-dose sinograms yield near-standard images","Hybrid denoise-reconstruct net turns low-dose PET into standard quality","Cascaded triple-domain network reconstructs standard PET from low-dose","Triple-domain denoise-reconstruct beats six methods on PET quality","End-to-end triple-domain PET: sinogram to standard image in one pass"],"cache_read_input_tokens":27136,"weakest_assumption_plain":"The load-bearing premise is that sinogram patches obtained by applying the inverse Radon transform to reconstructed PET image patches faithfully represent real low-dose PET raw sinograms; the authors explicitly acknowledge in Section 4.1 that their indirectly obtained sinograms may differ from real raw data, so the projection-domain gain could shrink on clinical raw data.","fun_headline_variants_meta":{"raw":{"variants":["Triple-domain PET boost: low-dose sinograms yield near-standard images","Hybrid denoise-reconstruct net turns low-dose PET into standard quality","Cascaded triple-domain network reconstructs standard PET from low-dose","Triple-domain denoise-reconstruct beats six methods on PET quality","End-to-end triple-domain PET: sinogram to standard image in one pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000719,"raw_usage":{"total_tokens":3287,"prompt_tokens":1064,"completion_tokens":2223,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":2128}},"tokens_in":680,"tokens_out":2223,"duration_ms":14165,"temperature":1.0,"reasoning_tokens":2128,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:27:44.240747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train TriPLET on true list-mode low-dose raw sinograms paired with standard-dose reconstructions from a clinical scanner, then test on held-out raw data and compare PSNR, SSIM, and rRMSE with the reported 25.932 dB, 0.992, and 0.296; a substantial drop would show that the simulated-sinogram assumption, not the triple-domain design, drove the reported gains.","supporting_citations":[{"cited_title":", year 1996","cited_arxiv_id":null,"evidence_quote":"Establishes the lossless Radon/inverse Radon conversion used to derive sinogram patches from image patches."},{"cited_title":", author Zhang, H","cited_arxiv_id":null,"evidence_quote":"Provides the multi-level wavelet-CNN idea that motivates replacing pooling and upsampling with wavelet transforms in RecNet."},{"cited_title":", author Lin, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the shifted-window multi-head self-attention design used in DenNet's transformer blocks."},{"cited_title":", author Yu, B","cited_arxiv_id":null,"evidence_quote":"Serves as the cGAN baseline that TriPLET must beat in the quantitative comparison."}],"review_version":1}