{"id":"dcafa6ec-154f-40bb-8de7-69d4171f0eaf","arxiv_id":"2412.09962","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A wavelet diffusion model generates patient-specific pseudo-healthy knee MR images for trochlear dysplasia, improving sulcus angle, groove depth, and Dejour classification.","lead":"Researchers trained a wavelet diffusion model on healthy knee MRIs to fill in the deformed trochlear groove of patients with trochlear dysplasia, creating 'pseudo-healthy' images that a surgeon judged less dysplastic. The goal is to give surgeons a patient-specific target for trochleoplasty planning without extra CT radiation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline quantitative result (SA/TGD improvement, p=0.001/0.0002) is computed from 16 non-blinded, single-rater measurements; without blinded multi-rater reliability the \"significant improvement\" claim is not yet established.","rationale":"The reader's weakest assumption was the domain gap between adult healthy fastMRI training data and adolescent TD patients. That is a real concern, but the paper's own conclusion acknowledges it and flags growth plates and swelling as limitations. The more load-bearing weakness is the validity of the outcome measurements themselves: the single quantitative result is based on 16 non-blinded single-rater measurements with no reliability analysis. If those measurements are biased, the headline claim fails regardless of how well the model handles domain shift. A blinded multi-rater reliability study is a concrete, feasible check that would settle whether the reported improvements are real. The reader already issued a CONDITIONAL verdict; this concern reinforces that assessment rather than changing it. I therefore recommend UNCHANGED, with the understanding that the conditional status is justified and should be resolved by the proposed reliability experiment.","tokens_in":7432,"tokens_out":2456,"duration_ms":30676,"concrete_test":"Recruit two additional orthopedic surgeons blinded to the study condition and to the location of the inpainted region. Present the 16 paired pre/post scans (and, if feasible, all 49) in randomized order, with the inpainted region masked or indicated only after measurement. Have each rater measure SA, TGD, and Dejour twice with a washout period of at least two weeks. Compute intra-rater and inter-rater intraclass correlation coefficients (ICC) and the smallest detectable difference (SDD). If the mean pathological-to-pseudo-healthy change is smaller than the SDD, or if the improvement is not replicated across raters, the significant-improvement claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim rests on measurements described in Sections 3.4 and 4.1: 49 TD scans, but only 16 have measurable SA and TGD before inpainting; the Wilcoxon tests are on these 16 pairs. The measurements were made by a single deputy attending surgeon with no reported blinding, no intra-rater repeatability check, and no inter-rater reliability analysis. Because the inpainted region is visually apparent and the before/after images are compared side by side, expectation bias can systematically shift SA and TGD toward the healthy range. Additionally, the 16 measurable cases are a selected subsample: 33 of 49 scans were too severe to measure before inpainting, so regression to the mean and selection effects could explain part of the improvement. The paper is honest about the domain shift between adult fastMRI training data and adolescent TD patients (Section 5), but it does not address measurement validity, which is the more immediate threat to the claim that inpainting \"significantly improves\" SA, TGD, and Dejour classification. Without reliability evidence, the reported p-values cannot be distinguished from rater bias or measurement noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D wavelet diffusion model trained on healthy fastMRI knee volumes to inpaint the trochlear region in patients with trochlear dysplasia, generating patient-specific pseudo-healthy MR images intended to support trochleoplasty planning. The method masks the patellar region with a 30 mm offset, conditions the model on the masked image and mask in the wavelet domain, and evaluates the result by having one orthopedic surgeon measure sulcus angle, trochlear groove depth, and Dejour classification on 49 pathological scans before and after inpainting. The authors report significant SA/TGD changes on 16 paired cases, Dejour improvement in 41/49 cases, and MSE/PSNR/SSIM scores on a fastMRI test set. The code is publicly released.","tokens_in":7657,"tokens_out":5102,"duration_ms":59776,"significance":"If the quantitative clinical claims hold, this is a useful proof-of-concept: it offers a patient-specific target shape for trochleoplasty from standard MR images, without additional radiation, and it is conditioned on the patient's own patella. Concrete strengths are the use of a public training dataset, released code, a clear clinical motivation, and an honest statement of applicability limitations. The central quantitative claim, however, is currently carried by a small, non-blinded, single-rater measurement study; the reported p-values should not be read as established evidence until measurement reliability, blinding, and selection effects are addressed.","major_comments":[{"comment":"The headline SA/TGD improvement rests on only 16 paired scans measured by a single deputy attending surgeon, with no blinding, no repeatability check, and no inter-rater reliability analysis. Because the before/after images are visually distinguishable and are compared side by side, expectation bias can systematically shift the measurements toward the healthy range. Please provide blinded repeated measurements with intra-class correlation coefficients and Bland-Altman limits, preferably with a second rater, and report a sensitivity analysis that includes all 49 scans rather than only the 16 in which SA/TGD were measurable before inpainting. Without this, the reported p=0.001 and p=0.0002 values cannot be separated from rater bias or measurement noise.","section":"§4.1, Fig. 7"},{"comment":"The Dejour improvement in 41 of 49 scans is also based on the same unblinded single-rater evaluation. Please provide a blinded or independently adjudicated Dejour classification, report the full contingency table including the cases in which SA/TGD could not be measured before inpainting, and clarify how the pre-inpainting Dejour stage was assigned in those 33 cases. If the pre-inpainting classification was necessarily subjective for severe cases, the 41/49 improvement rate should be interpreted with that caveat stated explicitly.","section":"§3.4, §4.1"},{"comment":"The training procedure does not specify how the masked image m1 and mask m2 are generated during training. This matters because inference uses a 30 mm bowl-shaped mask around the patella, and a systematic mismatch between training and inference masks changes the conditional distribution the model has learned. Please state the training mask sampling protocol (random, anatomical, or derived from SegmentAnyBone) and, ideally, include an ablation on mask placement and mask size.","section":"§2.4, Fig. 3"},{"comment":"The model is trained on adult fastMRI knees and applied to adolescent TD patients, a domain shift the authors acknowledge in Section 5. Because the intended use is surgical planning, the fastMRI reconstruction metrics (MSE/PSNR/SSIM) and the observed reduction in dysplasia measurements are not sufficient to establish anatomical correctness for a given patient. Please add a concrete validation that the generated trochlea is congruent with the patient's own patella, for example by measuring patellotrochlear contact or alignment, or by comparing against an adolescent healthy control group. Absent such evidence, the claim that the output 'should theoretically align with the respective patient's patella' remains an unverified assumption.","section":"§5, §3.1"}],"minor_comments":[{"comment":"The phrase 'quantitativ analysis' should be corrected to 'quantitative analysis'.","section":"§4.1"},{"comment":"The abstract states 'significantly improves the SA, TGD and D\\'ejour classification in a study with 49 knee MR scans,' but the SA/TGD statistics are computed on only 16 scans; the abstract should qualify this to avoid overstating the sample size.","section":"Abstract"},{"comment":"Please specify whether the L1 regularization on high-frequency wavelet coefficients is applied to the predicted coefficients or to the ground-truth coefficients; the current notation leaves this ambiguous.","section":"§2.4, Eq. (3)"},{"comment":"The 30 mm mask offset is a free parameter; please state how this value was chosen and whether the results are sensitive to it.","section":"§2.1"},{"comment":"The description of computing SSIM from a 'full SSIM image' from torchmetrics is unclear; standard SSIM is a scalar, so please clarify the exact procedure used.","section":"§3.5"},{"comment":"Reference [3] appears incomplete in the reference list; please provide the full bibliographic information.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is an application-oriented contribution; its main novelty is the clinical target and the patient-specific conditioning, while the generative architecture follows the authors' prior WDM work. The measurement-validity issue is the key gate for publication: the central clinical claim currently rests on a non-blinded single-rater evaluation of 16 paired cases. I would invite a revision that adds blinded multi-rater reliability analysis and a more complete handling of the unmeasurable cases, rather than reject, because the core idea is sound, the code is available, and the limitations are stated honestly. The missing specification of training mask generation should also be fixed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Best read as an early clinical feasibility report, not as evidence that the method normalizes anatomy. The new thing here is not the diffusion model—that is the Basel group's WDM—but the transfer to trochlear dysplasia: a patella-conditioned bowl mask that targets the pathological trochlear region, and a clinical evaluation with a surgeon measuring sulcus angle, trochlear groove depth, and Dejour stage. Using standard pre-op MR instead of CT is a genuine practical advantage for this mostly adolescent population, and the authors make code available and state limitations. That part is solid.\n\nThe weak joint is the quantitative claim. The SA/TGD \"significant improvement\" uses 16 paired scans where both pre and post measurements were possible; 33 of 49 were too severe to measure before inpainting. That is a selected subsample, and regression to the mean is not addressed. The measurements were taken by a single deputy attending surgeon, unblinded, with no intra- or inter-rater reliability. The inpainted region is visually obvious and the comparison is side by side, so expectation bias is a real threat. The Dejour improvement in 41/49 has the same problem. I cannot distinguish these p-values from rater bias or measurement noise.\n\nThe domain shift from adult fastMRI knees to adolescent TD patients is acknowledged but not tested, and the claim that the inpainted anatomy aligns with the patient's patella is asserted rather than verified geometrically. There is also no baseline comparison—no simple anatomical prior, no statistical shape model—so I cannot tell how much of the apparent normalization is due to the diffusion model specifically. The missing L1 weight for the regularization term is minor.\n\nAll that said, this is a coherent, honest application paper with a useful clinical target. It deserves referee time, but the authors should be pushed to add blinded multi-rater measurements, report all cases including the severe ones where feasible, and compare against a non-generative baseline. I would not cite the quantitative result as it stands, but I would mention the approach in related work.","headline":"The application is promising and the code is out, but the headline clinical claim rests on 16 unblinded single-rater measurements—send it to review, not to press.","tokens_in":8236,"tokens_out":3008,"would_cite":false,"duration_ms":36851,"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":"Pseudo-healthy knee MRI generation normalizes trochlear dysplasia measures in 49 scans.","keywords":["Pseudo-Healthy","Trochlear Dysplasia","Diffusion Models","Inpainting","Wavelet Diffusion","Knee MRI","Surgical Planning","Image Generation"],"falsifier":"Measure the congruence between the generated trochlear groove and the patient's patella in a cohort that later undergoes trochleoplasty, and compare the generated target with the groove shape that actually stabilizes the patella; if the two disagree, the pseudo-healthy image is not a reliable surgical target.","tokens_in":7227,"feed_emoji":"🦵","tokens_out":5894,"duration_ms":62913,"temperature":0.7,"pith_summary":"This paper tries to establish that a wavelet diffusion model trained on healthy knee MRI scans can generate a patient-specific pseudo-healthy image of the trochlear groove for adolescents with trochlear dysplasia, simply by masking the pathological region and inpainting. An orthopedic surgeon's measurements on 49 dysplastic knee scans show that after inpainting the sulcus angle and trochlear groove depth move toward normal values and the Dejour severity stage drops, with many severe cases becoming measurable for the first time. If the claim holds, surgeons would gain a concrete, per-patient visual target for reshaping the femur during trochleoplasty, a procedure that currently has no standardized preoperative plan.","feed_headline":"Inpainted knee MRIs show surgeons a healthy trochlear target","feed_subtitle":"Sulcus angle drops from 154° to 145° and groove depth deepens to 5.2 mm across 49 dysplasia scans.","key_machinery":"The machinery is the Wavelet Diffusion Model (WDM), a denoising diffusion probabilistic model whose forward and reverse processes operate on the Haar wavelet decomposition of the volume instead of the raw image. At every reverse step, the network receives the noisy wavelet coefficients concatenated with the wavelet transform of the masked pathological scan and the mask itself, and it is trained to predict the denoised volume with an MSE loss plus an $\\ell^1$ penalty on high-frequency components to suppress edge noise. This conditioning on the masked scan is what ties the generated healthy anatomy to the individual patient's patella.","core_discovery":"The central claim is that a diffusion model operating in the wavelet domain can restore a plausible healthy trochlear shape while remaining conditioned on the patient's own patella. The evidence: in 41 of 49 pathological scans the Dejour classification decreased by one or more stages, and in the 16 scans measurable both before and after, the sulcus angle fell from a mean of $154^\\circ$ to $145^\\circ$ ($p = 0.001$) and groove depth increased from 3.6 mm to 5.2 mm ($p = 0.0002$). Additionally, 33 of 49 severe cases had unmeasurable sulcus angle and groove depth before inpainting, but 46 of 49 were measurable afterward. The paper concludes that pseudo-healthy MR generation is feasible and can serve as a patient-specific planning aid for trochleoplasty.","pith_inferences":["A direct test of the planning value would be to register the generated groove to the patient's patella and measure patellotrochlear congruence, separating 'looks healthy' from 'fits this patella'.","The pseudo-healthy image could be converted into a quantitative resection guide, e.g., the volume to remove to reach the generated groove surface.","If combined with virtual patella repositioning, the same model might extend to dislocated-patella cases the authors had to exclude."],"forward_implications":["Surgeons can preview a patient-specific healthy trochlear shape from standard preoperative MRI, without needing an additional CT scan and its radiation.","Severe dysplasia cases where sulcus angle and groove depth cannot be measured become quantifiable after inpainting, providing numerical targets for the surgical correction.","The same masked-inpainting recipe could transfer to other joints or bone deformities that have a recognizable healthy shape but no standardized operative plan.","With a broader pediatric training set, the method might handle the currently excluded cases of dislocated patellae, swelling, and growth-plate edema."],"supporting_citations":[{"why":"supplies the wavelet diffusion architecture and training hyperparameters this work adapts","marker":"[20]"},{"why":"provides the masked-inpainting conditioning scheme that the method transfers from brain to knee","marker":"[19]"},{"why":"supplies the bone segmentation used to locate the patella and build the inpainting mask","marker":"[21]"},{"why":"provides the healthy knee MRI dataset used to train the model","marker":"[23]"},{"why":"defines the Dejour classification used as the clinical severity outcome","marker":"[10]"}],"fun_headline_variants":["Wavelet diffusion restores healthy trochlea in knee MRIs","AI inpaints healthy knee shape to plan trochleoplasty","Diffusion model shows surgeons a healthy trochlear target","Pseudo-healthy knee MRIs from diffusion aid surgery planning","Inpainted MRIs reveal normal trochlear groove for surgeons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model is trained only on healthy adult knee MRI scans and is applied to adolescent dysplasia patients without adapting for growth plates, swelling, or scanner differences, so the generated 'healthy' groove may not actually match that patient's patella.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet diffusion restores healthy trochlea in knee MRIs","AI inpaints healthy knee shape to plan trochleoplasty","Diffusion model shows surgeons a healthy trochlear target","Pseudo-healthy knee MRIs from diffusion aid surgery planning","Inpainted MRIs reveal normal trochlear groove for surgeons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1526,"prompt_tokens":1032,"completion_tokens":494,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":648,"tokens_out":494,"duration_ms":5843,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:30:51.570737+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the congruence between the generated trochlear groove and the patient's patella in a cohort that later undergoes trochleoplasty, and compare the generated target with the groove shape that actually stabilizes the patella; if the two disagree, the pseudo-healthy image is not a reliable surgical target.","supporting_citations":[{"cited_title":"In: MICCAI Workshop on Deep Generative Models, pp","cited_arxiv_id":null,"evidence_quote":"supplies the wavelet diffusion architecture and training hyperparameters this work adapts"},{"cited_title":"In: MICCAI Workshop on Deep Generative Models, pp","cited_arxiv_id":null,"evidence_quote":"provides the masked-inpainting conditioning scheme that the method transfers from brain to knee"},{"cited_title":"Radiology: Artificial Intelligence 2(1), 190007 (2020)","cited_arxiv_id":null,"evidence_quote":"provides the healthy knee MRI dataset used to train the model"},{"cited_title":"Knee Surgery, Sports Traumatology, Arthroscopy 2, 19–26 (1994)","cited_arxiv_id":null,"evidence_quote":"defines the Dejour classification used as the clinical severity outcome"}],"review_version":1}