{"id":"a11b5866-c860-4627-a11c-e59cc2dbb0ac","arxiv_id":"2506.20781","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DaT-CTLESS, using scatter-window deep-learning segmentation, achieved ICC 0.96 with CT-based attenuation correction on regional DaT uptake in an in silico trial, significantly outperforming uniform attenuation correction.","lead":"This paper presents DaT-CTLESS, a deep-learning method that corrects for tissue attenuation in dopamine transporter brain SPECT scans without needing a separate CT scan. In a simulated trial with 197 virtual patients, it matched CT-based correction closely on regional uptake and beat the standard uniform-attenuation method.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Primary ICC is computed across pooled VOIs; between-VOI variance may inflate the reported 0.96, and the ICC-difference significance is not reported with clustered inference.","rationale":"The reader's concern about simulation fidelity is a legitimate external-validity limitation and is explicitly acknowledged by the authors; but the paper's central claim is explicitly scoped to the in silico trial, and the RELAINCE conclusion quotes the ICC as the evidence. The statistical construction of that ICC is the condition that must hold for the central claim to be true even in silico. If the pooled ICC is inflated, the reported number is not a reliable measure of agreement. The concern is concrete and testable from the existing data. I recommend conditional acceptance: require per-VOI ICCs and clustered inference for the ICC difference. The simulation-fidelity limitation does not, by itself, invalidate the scoped claim; it affects the strength of the motivation for clinical translation, which the paper already phrases as future work.","tokens_in":18079,"tokens_out":11706,"duration_ms":142573,"concrete_test":"Request or recompute the per-VOI ICC(3,1) for each of the six regions (LC, RC, LP, RP, LGP, RGP) for DaT-CTLESS vs CTAC and UAC vs CTAC using the 47 test patients, plus a patient-level bootstrap 95% CI for the pooled-minus-within-region ICC difference and for the DaT-CTLESS-vs-UAC ICC difference. If the pooled 0.96 is driven by between-VOI variance (per-VOI ICCs below ~0.90, especially for GP), or the bootstrap CI for the ICC difference includes zero, the primary claim should be revised to a per-VOI or mixed-effects analysis.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Table 3 reports a single ICC of 0.96 (95% CI 0.94–0.97) for DaT-CTLESS vs CTAC, and 0.44 for UAC vs CTAC, as the primary endpoint. The Methods state the ICC is a two-way fixed single-score model for absolute agreement, but do not specify whether the unit of analysis is the individual patient–VOI combination. If, as appears, the 47 test patients × 6 VOIs are pooled into one ICC, the between-target variance includes large systematic differences among caudate, putamen, and GP uptake magnitudes. A two-way ICC is highly sensitive to this between-target spread; the same per-region measurement error yields a much higher pooled ICC than a within-region ICC. The manuscript does not report per-VOI ICCs, nor a confidence interval for the ICC difference between DaT-CTLESS and UAC, nor inference that accounts for the six repeated VOIs per patient. Thus the headline 'excellent agreement' (ICC 0.96) and the 'significantly higher correlation' claim are not yet supported by the reported statistical analysis. This is an internal, checkable issue, independent of how realistic the SIMIND simulations are.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DaT-CTLESS, a deep-learning-based transmission-less attenuation compensation method for dopamine transporter (DaT) SPECT. The method reconstructs scatter-window and photopeak-window projections, uses a U-net with attention gates to segment an initial scatter-window reconstruction into attenuation regions (white/gray matter, skull, scalp, head gear), assigns predefined literature-based attenuation coefficients to those regions, and uses the resulting attenuation map for OSEM reconstruction. The method is evaluated in an in silico imaging trial (ISIT-DaT) with 150 virtual patients for training and 47 for testing, generated from real MR/CT anatomy and clinical SBR distributions, simulated with SIMIND on GE and Siemens scanners. The primary endpoint is the ICC between regional DaT uptake estimates from DaT-CTLESS and CTAC versus the ICC between UAC and CTAC; secondary endpoints include AUC for normal versus reduced putamen SBR, test-retest repeatability, cross-scanner generalizability, fidelity-based FoMs, sensitivity to intra-regional heterogeneity, and comparison with two other DL-based AC methods. The authors report ICC 0.96 (95% CI 0.94–0.97) for DaT-CTLESS versus CTAC and 0.44 for UAC, and conclude that DaT-CTLESS is a reliable transmission-less AC method.","tokens_in":18283,"tokens_out":6821,"duration_ms":69526,"significance":"If the central claim is fully supported, DaT-CTLESS addresses a real clinical need: many DaT SPECT systems lack CT, and CT-based AC adds dose, cost, and misregistration risk. The study is significant for its adoption of the in silico imaging trial paradigm with a pre-specified primary endpoint, a power analysis, bootstrap confidence intervals, and multi-vendor scanner simulation. The authors credit the RELAINCE guidelines and evaluate on clinically relevant tasks. The method's use of scatter-window information is physically motivated, and the segmentation-based approach with literature-assigned attenuation coefficients avoids circularity from fitting attenuation values to the outcome. The secondary analyses, including test-retest repeatability and training-set-size sensitivity, are thoughtful. However, the primary endpoint analysis has a statistical weakness that must be addressed before the headline claim can be considered established.","major_comments":[{"comment":"The unit of analysis for the primary ICC is not specified. The Methods state that the ICC was computed using a two-way fixed single-score model for absolute agreement, but the reader cannot determine whether the 47 test patients × 6 VOIs were pooled into a single ICC. If pooled, the between-VOI variance—systematically different uptake magnitudes in caudate, putamen, and GP—will inflate the ICC for both methods, and the reported superiority of DaT-CTLESS (0.96) over UAC (0.44) may reflect the methods' ability to reproduce the overall VOI pattern rather than accurate per-region quantification. The paper does not report per-VOI ICCs, nor does it report the confidence interval or p-value for the ICC difference, despite the Methods stating that a bootstrapping strategy would be used for this difference. Thus the primary claim of a significantly higher correlation with CTAC is not yet supported by the reported statistical analysis.","section":"Statistical Considerations; Results, Table 3"},{"comment":"The power analysis is designed to test a single ICC against a null value (0.90 vs 0.75) using Walter et al.'s approach, not to detect a difference between two correlated ICCs obtained from the same set of subjects. It also does not account for the clustered, repeated-measures structure of six VOIs per patient. Consequently, the sample size of N=47 has no demonstrated power to support the primary comparison, and the F-test cited there cannot be directly applied to the DaT-CTLESS versus UAC ICC difference. The authors should either re-frame the power analysis for the actual comparison or justify why the simpler analysis is sufficient.","section":"Statistical Considerations"},{"comment":"The reported ICC for DaT-CTLESS (0.96) and the claim of 'excellent agreement' with CTAC rest on a single pooled value. To make the primary endpoint interpretable, the authors should report ICCs per VOI (or per VOI-group) and an ICC computed with a model that treats patients as random effects and accounts for repeated VOIs. Fig. 9 reports ICCs across heterogeneity levels, but again without specifying whether the unit is patient–VOI or patient; the same concern applies there. Providing per-VOI ICCs and a cluster-bootstrap confidence interval for the ICC difference would be a straightforward correction.","section":"Results, Table 3 and Fig. 9"}],"minor_comments":[{"comment":"The first page contains a note that the manuscript has been accepted for publication in Medical Physics on June 15, 2025; this note is inappropriate for a submitted manuscript and should be removed.","section":"Page 1 (header)"},{"comment":"In the Abstract, 'DaT-CLTESS' appears to be a typo for 'DaT-CTLESS' in the third sentence of the Background paragraph and at the start of the Results paragraph.","section":"Abstract"},{"comment":"The text says the segmentation network was trained to estimate 'white/gray matters, skull, scalp, and the head gear' with a weighted cross-entropy loss, but the choice of weights is not described.","section":"Methods, DaT-CTLESS method"},{"comment":"The caption states 'Normalized RMSE and SSIM between (a) attenuation maps as well as (b) activity maps obtained by the CTAC method and those obtained by DaT-CTLESS and UAC.' It would help to state explicitly that DaT-CTLESS and UAC are compared against CTAC as reference, and to clarify whether the FoMs in (a) and (b) are computed over the same VOIs or over the whole image.","section":"Results, Fig. 7 caption"},{"comment":"The power analysis says 'We found that N = 33 samples were needed to detect an ICC under the proposed method of 0.90 against a null ICC of 0.75' but the test dataset includes 47 patients; the text should explain why 47 was chosen over 33 (e.g., to accommodate the 23 group B patients).","section":"Statistical Considerations"},{"comment":"The statement that DaT-CTLESS is expected to be insensitive to patient motion because motion equally affects photopeak and scatter-window data is plausible but not directly tested in the trial; it should be phrased as a hypothesis rather than a demonstrated property.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The primary statistical issue is fixable and does not undermine the overall contribution, but the authors must re-analyze the data with appropriate clustering (per-VOI ICCs, cluster bootstrap for the ICC difference, and a proper power analysis) before the central claim is supported. The header stating acceptance is also a submission-process concern that the editor may wish to address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you care about quantitative SPECT and DL-based attenuation compensation. The real contribution is not the network—it's the evaluation. They ran a pre-specified in silico trial with a power analysis, a primary endpoint, bootstrap CIs, and a broad set of secondary analyses: AUC for normal-vs-reduced SBR, test-retest repeatability, cross-scanner generalizability, sensitivity to intra-regional heterogeneity, and training-set size. That level of rigor is rare in the DL-imaging literature, and it shows.\n\nThe method itself is an incremental extension of their earlier cardiac CTLESS work to brain DaT SPECT: segment scatter-window reconstructions into tissue classes with a U-net, assign literature attenuation coefficients per region, then use the resulting map for AC. The physics rationale is reasonable—scatter-window data does carry attenuation information, and the head gear is a genuine problem that uniform AC ignores. The head-gear example in Fig. 8 is convincing.\n\nThe main soft spot is statistical. The primary ICC pools all six VOIs across patients. Between-VOI differences in uptake magnitude can inflate a two-way fixed single-score ICC, so the headline 0.96 may be optimistic as an absolute number. The methods say they computed a bootstrap CI for the ICC difference, but it never appears in the results—only the two individual CIs. A referee should ask for per-VOI ICCs and a clustered analysis that treats the six VOIs as repeated measures within each patient. This affects the strength of the claim more than its direction: UAC is clearly worse on every secondary metric, so the practical conclusion is unlikely to flip.\n\nThe other obvious limitation is simulation realism. SIMIND is a well-validated simulator, but the virtual patients come from a single-center MR/CT cohort and the tracer kinetics are static. The authors acknowledge this, but readers should treat \"excellent agreement with CTAC\" as a statement about the simulated world, not yet about patients. No code or data is released, which weakens reproducibility, though the trial design is described in enough detail to be reconstructed.\n\nWho is this for? Imaging scientists working on SPECT AC, especially those planning virtual trials for AI methods. It deserves a serious referee. I would engage with it, and I would ask for the clustered ICC analysis before taking the 0.96 at face value.","headline":"A carefully designed in silico imaging trial for a transmission-less DaT SPECT AC method; the pooled-ICC analysis needs a closer look, but the core result is credible.","tokens_in":18855,"tokens_out":2156,"would_cite":true,"duration_ms":25757,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DaT SPECT attenuation correction can be done without a CT scan by segmenting scatter-window reconstructions with a U-net, matching CT-based correction in an in silico trial.","keywords":["DaT SPECT","transmission-less attenuation compensation","deep learning","in silico imaging trial","scatter-energy window","regional uptake quantification","specific binding ratio","attenuation map segmentation"],"falsifier":"Scan real patients on a SPECT/CT system, derive DaT-CTLESS attenuation maps from SPECT data alone, and compare regional uptake with same-day CTAC; if the ICC falls to the level of UAC (around 0.44) or head-gear attenuation is missed in patients who moved during acquisition, the central claim would be refuted.","tokens_in":17884,"feed_emoji":"🧠","tokens_out":9109,"duration_ms":93174,"temperature":0.7,"pith_summary":"The paper proposes and evaluates DaT-CTLESS, a deep-learning method that performs attenuation compensation for dopamine-transporter (DaT) brain SPECT without any transmission or CT scan. Its central claim is that regional DaT uptake measured after DaT-CTLESS correction tracks CT-based attenuation correction (CTAC) far more closely than the standard CT-free fallback, a uniform attenuation map (UAC). In the ISIT-DaT in silico trial, the intraclass correlation with CTAC was 0.96 (95% CI [0.94, 0.97]) for DaT-CTLESS versus 0.44 for UAC. The method also beat UAC at distinguishing normal from reduced putamen binding, matched CTAC in test-retest repeatability, generalized across two simulated scanners, and held its accuracy with as few as 50 training patients. These results motivate moving the method into clinical evaluation.","feed_headline":"CT-free DaT SPECT correction matches CT-based accuracy in trial","feed_subtitle":"Regional uptake agreed with CTAC at ICC 0.96 versus 0.44 for uniform correction in a 197-patient virtual trial.","key_machinery":"The load-bearing mechanism is the scatter-energy-window reconstruction as the physical source of attenuation information, combined with a U-net (a convolutional encoder-decoder network) that converts that reconstruction into a small set of attenuation regions rather than per-voxel coefficients. Because Compton scatter probability at a location tracks the attenuation coefficient there, scatter-window projections carry attenuation information even from structures with no tracer uptake, such as the head gear that UAC cannot see. The segmentation step reframes a high-dimensional, ill-posed voxel-estimation problem as a low-dimensional region-labeling problem, which is why the method needs little training data and outperforms voxel-wise deep-learning estimators.","core_discovery":"On the paper's own terms, the discovery is that the attenuation map needed for quantitative DaT SPECT can be recovered from the emission data itself by treating attenuation-map estimation as a segmentation problem. DaT-CTLESS takes photopeak- and scatter-energy-window reconstructions as input to a U-net that labels attenuation regions, including white/gray matter, skull, scalp, and head gear, and then assigns each region a fixed attenuation coefficient. In the ISIT-DaT virtual trial this pipeline produced regional uptake estimates with an ICC of 0.96 against CTAC, significantly higher than the 0.44 ICC between UAC and CTAC, at similar fidelity to ground truth and similar test-retest repeatability to CTAC. It also outperformed direct and per-voxel indirect deep-learning AC methods across training set sizes, which the paper attributes to the dimensionality reduction of segmentation.","pith_inferences":["The central comparison is only as strong as the simulation; a real-world clinical study is the natural next test, and the paper itself lists this as future work.","If real scatter data include motion or head-gear variations not captured in the Monte Carlo model, scanner- or protocol-specific retraining may be needed before deployment.","The same scatter-window-plus-segmentation idea could be tested for attenuation compensation in other SPECT tracers and possibly outside the brain, since the scatter-attenuation correlation is not specific to DaT imaging.","Because training labels come from CT-derived segmentations, the method's ceiling is the accuracy of those segmentations; clinical transfer would depend on CT-derived regions matching real anatomy."],"forward_implications":["SPECT systems without a CT component could obtain attenuation compensation approaching CT-based accuracy, removing a barrier for community, mobile, and solid-state-detector scanners.","Eliminating the CT scan removes SPECT-CT misalignment as a source of quantification error, a practical concern for Parkinson patients who move during scanning.","Large existing SPECT databases that lack transmission scans could become usable for quantitative DaT uptake analysis.","Because the method is formulated as segmentation, it may be retrainable for other tracers or brain SPECT protocols with modest training data."],"supporting_citations":[{"why":"Fisher information analysis showing scatter-window emission data carries information about the attenuation distribution, the physical premise for DaT-CTLESS.","marker":"[32]"},{"why":"Earlier segmentation-based method that assigns attenuation coefficients to regions segmented from scatter and photopeak data; the family DaT-CTLESS extends.","marker":"[21]"},{"why":"Prior scatter-window-plus-deep-learning transmission-less AC method for myocardial perfusion SPECT, whose positive results motivated the DaT extension.","marker":"[38]"},{"why":"Practice guideline that defines the clinical imaging protocol and the uniform attenuation map (UAC) used as the CT-free comparator.","marker":"[43]"},{"why":"Clinical SBR distributions from Parkinson patients and normal controls used to assign realistic uptake values to virtual patients.","marker":"[48]"},{"why":"Monte Carlo SPECT simulation used to generate the virtual scans for both simulated scanners in the trial.","marker":"[53]"},{"why":"Best-practices guideline that motivates task-based in silico evaluation of AI methods in nuclear medicine.","marker":"[39]"},{"why":"ICC guideline used to interpret the agreement between DaT-CTLESS and CTAC as high.","marker":"[55]"}],"fun_headline_variants":["DaT SPECT attenuation map from emission data matches CTAC","Deep learning eliminates CT for DaT SPECT quantification","CT-less DaT SPECT hits ICC 0.96 in virtual imaging trial","Uniform AC falls short: DL AC matches CT in DaT SPECT","DaT SPECT without CT achieves ICC 0.96 in trial"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trial assumes that the Monte Carlo simulations and the virtual patient anatomy faithfully reproduce clinical DaT SPECT, especially the scatter-window projections; if that simulated physics differs from real scans, the learned segmentation may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["DaT SPECT attenuation map from emission data matches CTAC","Deep learning eliminates CT for DaT SPECT quantification","CT-less DaT SPECT hits ICC 0.96 in virtual imaging trial","Uniform AC falls short: DL AC matches CT in DaT SPECT","DaT SPECT without CT achieves ICC 0.96 in trial"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000936,"raw_usage":{"total_tokens":4073,"prompt_tokens":1085,"completion_tokens":2988,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2896}},"tokens_in":701,"tokens_out":2988,"duration_ms":22261,"temperature":1.0,"reasoning_tokens":2896,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:41:29.428232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan real patients on a SPECT/CT system, derive DaT-CTLESS attenuation maps from SPECT data alone, and compare regional uptake with same-day CTAC; if the ICC falls to the level of UAC (around 0.44) or head-gear attenuation is missed in patients who moved during acquisition, the central claim would be refuted.","supporting_citations":[{"cited_title":"Fisher information analysis of list-mode SPECT emission data for joint estimation of activity and attenuation distribution","cited_arxiv_id":null,"evidence_quote":"Fisher information analysis showing scatter-window emission data carries information about the attenuation distribution, the physical premise for DaT-CTLESS."},{"cited_title":"Segmentation of the body and lungs from Compton scatter and photopeak window data in SPECT: a Monte-Carlo investigation [published online ahead of print 1996/01/01]","cited_arxiv_id":null,"evidence_quote":"Earlier segmentation-based method that assigns attenuation coefficients to regions segmented from scatter and photopeak data; the family DaT-CTLESS extends."},{"cited_title":"CTLESS: A scatter-window projection and deep learning-based transmission-less attenuation compensation method for myocardial perfusion SPECT","cited_arxiv_id":null,"evidence_quote":"Prior scatter-window-plus-deep-learning transmission-less AC method for myocardial perfusion SPECT, whose positive results motivated the DaT extension."},{"cited_title":"SNM practice guideline for dopamine transporter imaging with 123I-ioflupane SPECT 1.0","cited_arxiv_id":null,"evidence_quote":"Practice guideline that defines the clinical imaging protocol and the uniform attenuation map (UAC) used as the CT-free comparator."},{"cited_title":"Imaging analysis of Parkinson’s disease patients using SPECT and tractography","cited_arxiv_id":null,"evidence_quote":"Clinical SBR distributions from Parkinson patients and normal controls used to assign realistic uptake values to virtual patients."},{"cited_title":"SIMIND Monte Carlo simulation of a single photon emission CT","cited_arxiv_id":null,"evidence_quote":"Monte Carlo SPECT simulation used to generate the virtual scans for both simulated scanners in the trial."},{"cited_title":"A guideline of selecting and reporting intraclass correlation coefficients for reliability research","cited_arxiv_id":null,"evidence_quote":"ICC guideline used to interpret the agreement between DaT-CTLESS and CTAC as high."}],"review_version":1}