{"id":"cee9b74c-72f6-419e-81b8-dbc2c26117e7","arxiv_id":"1908.04809","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Both a conditional GAN and a cycleGAN generated abdominal synthetic CT from 0.35T MRI with dose-calculation accuracy within clinical tolerances for liver radiotherapy.","lead":"This paper tests two AI models that turn low-field MRI scans into synthetic CT scans for liver cancer radiotherapy planning, avoiding the need for a separate CT scan. Both models produced dose calculations within clinical tolerances for an 8-patient cohort, with the conditional GAN slightly outperforming the cycleGAN on image accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated deformable CT-to-MR registration makes dCT a questionable ground truth; sCT metrics may reflect agreement with a distorted reference.","rationale":"The reader's weakest assumption identifies the deformable registration of CT to MR as the key unvalidated step, and I agree this is the most load-bearing concern. The paper's evaluation is internally consistent: the dCT is the clinical planning reference, and sCT is compared against it for both HU accuracy and dose calculation. However, because the dCT is also the training label for the sCT generators, any systematic error in the registration is learned and then reinforced in the evaluation. Without registration accuracy metrics, the reported MAE (~90 HU), gamma passing rates, and DVH deviations cannot be interpreted as accuracy relative to true patient anatomy. The central feasibility claim is therefore conditional on the dCT being a faithful surrogate for the true CT in MR coordinates. The paper's own limitations statement acknowledges the small cohort and the lack of statistical testing for the cGAN-versus-cycleGAN comparison; while those are real limitations, they do not undermine the primary claim that both models achieved clinically acceptable dose agreement with the reference. The missing registration validation is more fundamental because it challenges the validity of the reference itself. A concrete registration-quality assessment, as proposed, would directly settle whether the concern lands. If the registration is accurate, the conditional verdict can be upgraded; if not, the results may overstate readiness for MR-only liver radiotherapy. The reader's conditional verdict remains appropriate pending this check.","tokens_in":8986,"tokens_out":4477,"duration_ms":51381,"concrete_test":"Quantify per-patient deformable registration accuracy between the original breath-hold CT and 0.35T MR used to create dCT: report liver and vertebral-body Dice similarity, mean surface distance, or target registration error on anatomical landmarks (e.g., vessel bifurcations). Then recompute MAE/PSNR and gamma/DVH deviations with sCT compared against a rigidly registered (or non-deformed) CT in MR coordinates. If registration errors are small (e.g., liver Dice >= 0.95) and the dosimetric results persist, the concern is resolved; if errors are large or results change materially, the feasibility claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that sCTs enable accurate MR-only liver dose calculation rests on comparing sCT-based doses to doses computed on the deformed CT (dCT) described in Section 2.1. The accuracy of the deformable CT-to-MR registration is never quantified. If the registration misaligns anatomy (e.g., liver boundary, ribs, bowel gas), the dCT is not a valid ground truth: the sCT generator is trained to reproduce those same registration artifacts, so low MAE and high gamma between sCT and dCT may reflect agreement with a systematically distorted reference rather than with true patient electron density. The reported dose agreement would then overstate clinical accuracy. This is the single load-bearing assumption because both the HU metrics (Eqs. 1-2) and the dosimetric endpoints (Tables 3-4) are defined relative to dCT; no independent CT-based comparison is performed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript investigates the generation of synthetic CT (sCT) images from 0.35T MR images using two generative adversarial networks, cGAN and cycleGAN, for MR-only liver radiotherapy. A retrospective cohort of 12 abdominal cancer patients (8 liver, 4 non-liver) was used. CT images were deformably registered to the MR images to create deformed CT (dCT) images, which served as the reference for training and evaluation. The models were trained on transverse slices with four-fold cross-validation, and sCT quality was assessed via MAE and PSNR inside the body contour. Dose calculation accuracy was evaluated on 8 liver patients by recalculating clinical plans on sCT and comparing dose distributions to dCT-based plans using gamma analysis and DVH metrics. The results show average MAE of 89.8 HU (cGAN) and 94.1 HU (cycleGAN), gamma passing rates above 95% (2%, 2 mm) and 99% (3%, 3 mm), and average dose metric deviations within ±0.6% for the PTV and ±0.15% for OARs. The authors conclude that both models achieve accurate dose calculation and are promising for MR-only liver radiotherapy, with cGAN appearing slightly more accurate than cycleGAN.","tokens_in":9113,"tokens_out":2957,"duration_ms":32140,"significance":"If validated, this work would be the first demonstration of deep learning-based sCT generation from low-field (0.35T) MR images for abdominal sites, specifically liver, with potential direct application to MR-guided online adaptive radiotherapy. The study has notable strengths: it uses a held-out cross-validation scheme, performs dosimetric evaluation with gamma analysis and clinically relevant DVH metrics on actual treatment plans, and directly compares two popular GAN architectures under identical data conditions. The reported sCT generation time of about 6 s per patient is practically attractive for online adaptive workflows. However, the evaluation relies entirely on a dCT reference created by deformable registration whose accuracy is not reported, and the sample size for dosimetric conclusions is small (n=8) with no statistical testing. These issues currently limit the strength of the feasibility claim, though the methodology is appropriate and the results are promising as a preliminary feasibility study.","major_comments":[{"comment":"The claim that cGAN outperforms cycleGAN (abstract, Section 4) is not supported by statistical evidence. The differences in MAE (89.8 HU vs 94.1 HU) and gamma passing rates (e.g., 97.4% vs 95.6% for D ≥ 90% at 2%, 2 mm) are based on only 12 patients for HU metrics and 8 for dosimetric metrics, with large standard deviations. No confidence intervals, hypothesis tests, or non-inferiority analyses are provided. Given the small cohort, these differences could easily be within sampling variability. The authors should either provide appropriate statistical comparisons or temper the comparative conclusion to a descriptive observation that warrants future investigation with larger cohorts.","section":"Section 2.1, 2.4, Eqs. (1)-(2), Tables 3-4"}],"minor_comments":[{"comment":"There is a typo: \"sCTcyleGAN\" should be \"sCTcycleGAN\".","section":"Abstract"},{"comment":"The PSNR formula uses a constant of 4095, but the text states the networks process 16-bit single-channel images; 4095 is the maximum value for a 12-bit range. Please clarify the bit depth used for the images and ensure consistency.","section":"Eq. (2)"},{"comment":"The column header \"PNSR\" is a typo; it should be \"PSNR\".","section":"Table 2"},{"comment":"The sentence describing cycleGAN's cycle consistency is confusing: \"convert generated sMR (sCT) slices back to cycleCT (cycleMR) slices\" should be rephrased to clearly indicate that the cycle loss computes L1 distance between the original CT/MR and the cycled CT/MR slices.","section":"Section 2.2"},{"comment":"The four-fold cross-validation uses groups of 3 patients for training and 1 for testing, but the paper does not specify whether the groups were stratified by tumor site or dose fractionation. Given the heterogeneous cohort, a stratified split would be more robust; please specify or justify the random split.","section":"Section 2.3"},{"comment":"Figure 3 shows dose distribution comparisons for only one patient; while representative, providing dose difference maps for an additional patient with larger differences (e.g., the cycleGAN patient with 1.0% PTV mean dose deviation) would strengthen visual evidence of the variability.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The central feasibility claim is plausible, but the evaluation's validity rests on the unvalidated dCT reference. I recommend requesting the authors to either report deformable registration accuracy or conduct a secondary analysis against original CT (e.g., after rigid registration) to rule out that the sCT models are learning registration artifacts. The small sample size and lack of statistical tests for the cGAN-vs-cycleGAN comparison should also be addressed. If the dCT reference cannot be validated, the conclusions may need to be reframed as demonstrating technical feasibility within the limitations of the reference standard. The manuscript is otherwise well-structured and fits the scope of physics in medicine."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a competent feasibility study, not a methodological breakthrough. What's new is the application—GAN-based sCT generation from 0.35T abdominal MR for liver radiotherapy—and the head-to-head cGAN vs cycleGAN comparison in that setting. That alone justifies a read.\n\nThe paper does several things well. The evaluation is clinically meaningful: four-fold cross-validation, HU metrics (MAE/PSNR), plus dose recalculation with gamma and DVH endpoints for 8 liver patients. The results are consistent with prior high-field sCT work, and the dose errors are small (PTV mean within ±0.6%). They also state plainly that more patients are needed for statistical testing, which is honest.\n\nThe main soft spot is the ground truth. The dCT is created by deformably registering the planning CT to the MR in the treatment planning system, and the accuracy of that registration is never quantified. If the registration is imperfect—especially at the liver boundary, ribs, or bowel gas—the sCT is trained to reproduce those registration artifacts, and the low MAE and high gamma could reflect agreement with a distorted reference rather than true anatomy. Since both the HU and dosimetric endpoints are defined relative to dCT, this is the load-bearing assumption. I don't think it invalidates the central feasibility claim—dose calculation is fairly tolerant of moderate HU errors, and the results are in line with other sCT studies—but it should be fixed (e.g., report registration accuracy or use an independent validation set).\n\nMinor issues: the cGAN vs cycleGAN comparison is not statistically tested; the differences are small relative to the standard deviations. No code or data, which limits independent verification but is typical for this field. Also, the cycleGAN is trained on unpaired slices, but the patients are the same, so the practical advantage over pairwise training isn't compelling here.\n\nOverall, a solid paper for medical physicists working on MR-only workflows. It deserves serious peer review; the dCT accuracy issue is addressable and, if resolved, would make the conclusions stronger.\n\nRegards","headline":"A competent feasibility study of GAN-based sCT from low-field MR for liver radiotherapy, undermined a bit by unvalidated deformable registration as ground truth.","tokens_in":9709,"tokens_out":2157,"would_cite":true,"duration_ms":20080,"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":"GANs convert low-field MRI into synthetic CT for liver radiotherapy","keywords":["synthetic CT","generative adversarial network","cGAN","cycleGAN","MR-only radiotherapy","liver radiotherapy","low-field MRI","dose calculation"],"falsifier":"Take one liver patient with a native breath-hold CT that can be aligned to the MR without deformable registration, generate the sCT, and recalculate the clinical plan; if the 2%, 2 mm gamma pass rate against the native-CT dose drops below 95% or the PTV dose deviation exceeds 1%, the MR-only dose calculation claim would be violated for that protocol.","tokens_in":8778,"feed_emoji":"🩻","tokens_out":5788,"duration_ms":51975,"temperature":0.7,"pith_summary":"This paper asks whether synthetic CT images generated from low-field (0.35T) MR images can replace real CT for dose calculation in liver radiotherapy. Using two generative adversarial networks, cGAN and cycleGAN, on 12 abdominal cancer patients, it reports Hounsfield-unit errors of about 90 HU and dose distributions on eight liver plans that match the CT-based reference within tolerance: average gamma passing rates above 95% at 2%, 2 mm and above 99% at 3%, 3 mm. If these results hold, MR-only liver radiotherapy becomes feasible, eliminating the separate CT scan, its registration uncertainty, radiation exposure, and workflow cost. The paper also positions this as the first deep-learning sCT study using low-field MR-guided radiotherapy images.","feed_headline":"GANs convert low-field MRI into synthetic CT for liver radiotherapy","feed_subtitle":"Two GANs hit 95%+ gamma pass rates at 2%/2 mm on eight liver plans, supporting MR-only planning.","key_machinery":"The load-bearing mechanism is the generative adversarial network, specifically two variants: a conditional GAN (cGAN) trained on paired MR and deformed-CT slices, and a cycle-consistent GAN (cycleGAN) trained on unpaired slices through cycle-consistency losses that translate MR to CT and back. Each generator is a convolutional neural network mapping 0.35T TrueFISP MR slices into synthetic CT slices; each discriminator tries to distinguish real from generated images. The synthetic CT then supplies Hounsfield units to the treatment planning system's Monte Carlo dose calculation, so the quantity that carries the argument is the learned map from MR intensity patterns to electron-density-equivalent HU values.","core_discovery":"The paper claims that abdominal synthetic CT images generated by two generative adversarial networks, cGAN and cycleGAN, from 0.35T MR images are accurate enough for dose calculation in liver radiotherapy. Against the deformed CT reference, the networks produce mean absolute errors of 89.8 HU (cGAN) and 94.1 HU (cycleGAN); on eight liver plans, both deliver average gamma passing rates higher than 95% at the 2%, 2 mm criterion and higher than 99% at 3%, 3 mm, with mean dose and DVH deviations within ±0.6% of prescribed dose for the PTV and within ±0.15% for evaluated organs. The authors conclude that MR-only liver radiotherapy is achievable with these models, and that cGAN is somewhat more accurate than cycleGAN in this cohort.","pith_inferences":["If the deformable registration that creates the dCT reference has errors comparable to the reported MAE, the true HU accuracy against native CT could be worse than the quoted ~90 HU; an independent rigidly co-registered comparison would resolve this.","Both models had visible difficulty reproducing ribs; this may matter more for plans whose beams traverse the chest wall than for the liver plans tested here, and could be checked with a rib-tissue phantom or a thoracic cohort.","The cGAN's small advantages over cycleGAN may shrink or reverse with more training data, since cycleGAN's unpaired training can exploit routine scans that lack a co-registered CT."],"forward_implications":["Both cGAN and cycleGAN produce abdominal sCTs whose HU errors are comparable to previously reported abdominal sCT results from high-field MRI, suggesting the low-field source is not a barrier.","Dose recalculations on eight liver plans meet the 2%, 2 mm gamma criterion on average, so these models could replace the planning CT in the MR-only workflow for similar patients.","With sCT generation under about 10 seconds per patient, the method is fast enough to sit inside an online adaptive treatment session without stretching it.","Because cycleGAN does not require paired MR-CT training data, it can be trained on larger unpaired clinical datasets, potentially improving generalization."],"supporting_citations":[{"why":"Supplies the cGAN architecture and training protocol used for paired MR-to-sCT generation.","marker":"Isola et al 2016"},{"why":"Supplies the cycleGAN architecture and cycle-consistency training used for unpaired MR and CT slices.","marker":"Zhu et al 2017"},{"why":"Prior deep-learning abdominal sCT study; provides the high-field MAE benchmark the results are compared with.","marker":"Liu et al 2019"},{"why":"Prior cGAN-based pelvic sCT dose evaluation; establishes the dose-accuracy evaluation approach transferred to the abdomen.","marker":"Maspero et al 2018"},{"why":"Defines the gamma analysis method used to compare dCT-based and sCT-based dose distributions.","marker":"Low et al 1998"},{"why":"Review of synthetic CT methods that frames MR-only radiotherapy and defines the registration-uncertainty motivation.","marker":"Edmund and Nyholm 2017"}],"fun_headline_variants":["Low-field MRI to synthetic CT: GANs hit 95% gamma pass for liver","cGAN edges cycleGAN for MR-only liver radiotherapy","Synthetic CT from low-field MRI accurate for liver plans","Small study: Two GANs support MR-only liver radiotherapy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reference images used for training and evaluation are CTs that were deformably registered to the MR images, and the paper does not report how accurate that registration was; the reported HU and gamma errors are measured against these registered images, not against an independently verified ground truth.","fun_headline_variants_meta":{"raw":{"variants":["Low-field MRI to synthetic CT: GANs hit 95% gamma pass for liver","cGAN edges cycleGAN for MR-only liver radiotherapy","Synthetic CT from low-field MRI accurate for liver plans","Small study: Two GANs support MR-only liver radiotherapy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001212,"raw_usage":{"total_tokens":5067,"prompt_tokens":1104,"completion_tokens":3963,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":720,"completion_tokens_details":{"reasoning_tokens":3889}},"tokens_in":720,"tokens_out":3963,"duration_ms":28182,"temperature":1.0,"reasoning_tokens":3889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:31:55.976334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one liver patient with a native breath-hold CT that can be aligned to the MR without deformable registration, generate the sCT, and recalculate the clinical plan; if the 2%, 2 mm gamma pass rate against the native-CT dose drops below 95% or the PTV dose deviation exceeds 1%, the MR-only dose calculation claim would be violated for that protocol.","supporting_citations":[],"review_version":1}