{"id":"44527dd8-818a-45b4-ae95-6be828c23dbf","arxiv_id":"2505.09927","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A data-dependent frequency prompt plus BN pre-adaptation and style-layer fine-tuning improves source-free cross-modality medical image segmentation.","lead":"A source-free domain adaptation method lets a pretrained medical segmentation model be adapted to a new imaging modality without the original training data or new labels. It combines per-image frequency-domain style prompts, batch-normalization recalibration, and fine-tuning of only a few style-related layers to improve segmentation accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim unverified: key baselines quoted from another paper's random split, with no seeds or code; fairness of comparison is the load-bearing uncertainty.","rationale":"Reading the paper in good faith, the method is coherent: BN pre-adaptation is a sensible initialization, the data-dependent frequency prompt is a plausible extension of prior prompt-based style translation, and the style-layer fine-tuning is a reasonable efficiency choice. The central claim, however, is empirical superiority over state-of-the-art SFDA methods, and that claim stands or falls on the comparison protocol. The starred baseline rows (FVP, FSM*) are quoted from FVP's paper, which used a different random data partition, so those numbers are not directly comparable to DDFP's numbers even under the same evaluation metrics. Without multi-seed statistics or a code link, even the non-starred comparisons cannot be independently checked or placed in context. The aggregated Wilcoxon test does not repair this because it pools small test sets across datasets and directions and reports no per-table p-values. I do not allege any misconduct; the issue is evidentiary completeness. The reader's weakest assumption about the style/content split is a legitimate mechanism-level concern, and it should remain a condition for acceptance, but it is secondary to the comparison-control problem: an imperfect layer split would weaken one contribution, while an uncontrolled comparison would void the headline claim. Therefore I retain the reader's CONDITIONAL verdict and would not move it further, but I would weigh the reproducibility conditions at least as heavily as the architectural assumptions.","tokens_in":23051,"tokens_out":10498,"duration_ms":113482,"concrete_test":"Run a controlled replication on the abdominal CT-to-MRI task: obtain or reimplement FVP and FSM, initialize all methods from the same source U-Net checkpoint, use the same 8:2 patient split as DDFP, and evaluate on the same test slices; repeat with 3 random seeds and report Dice mean +/- std (and 95% CI) for every method. If DDFP's advantage over FVP/FSM falls inside the CI overlap, the SOTA claim is unsupported; if the margin persists, the comparison concern is resolved. As a secondary check, verify that the claimed 'code available online' link exists and runs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DDFP outperforms existing SFDA methods. The strongest evidence (Tables 1-3) is not on a common test bed: Section 4.1.4 says FVP (and FSM*) results are taken from [38], which used 'random data partition', while DDFP's numbers come from a different 8:2 split of the same datasets. A reported gap such as 0.8987 vs 0.8382 (ProContra) is controlled only if all rows share the same test slices; for the starred rows this is not the case. No error bars, multi-seed results, or p-values for the individual tables are given (the aggregated Wilcoxon in Fig. 6 pools small test sets across tasks), and the 'code available online' statement has no URL. The style/content layer split (the reader's concern) is a mechanism-level risk, but even an imperfect split would not invalidate the method if the controlled comparisons are right. The baseline-comparison and reproducibility gap is more load-bearing because it directly determines whether the claimed superiority is real.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DDFP, a source-free domain adaptation (SFDA) framework for medical image segmentation. The method combines three components: a momentum-based batch-normalization preadaptation of the source model using target data; a data-dependent frequency prompt generated by a small network that modulates the amplitude spectrum of each input image; and fine-tuning of only the first four convolutional layers of the U-Net backbone while freezing deeper layers. The target model is trained with a weighted combination of pseudo-label, BN-statistic, and entropy losses. Experiments on abdominal (CHAOS/Multi-Atlas), cardiac (MMWHS), and brain tumor (BraTS) datasets report Dice and ASD improvements over several SFDA baselines, with the largest gains on cross-modality abdominal CT-to-MRI. Ablations examine loss components, prompt design, BN preadaptation, and the number of trainable layers.","tokens_in":23299,"tokens_out":9352,"duration_ms":90030,"significance":"If the controlled comparisons were solid, the paper would be a useful contribution: the per-image data-dependent frequency prompt is a genuine departure from domain-level prompts in FVP/FSM, and the BN preadaptation step is a lightweight way to improve pseudo-label quality. The ablation studies are informative, and Fig. 9 provides a partial validation of the style-layer fine-tuning strategy. However, the headline claim of outperforming state-of-the-art methods is currently supported only by uncontrolled baseline numbers for FVP and FSM, and by single-run evaluations without error bars. The empirical significance is therefore conditional on the authors redoing the comparisons under a common protocol.","major_comments":[{"comment":"The results for FVP and FSM are taken from [38] (indicated by *), which used a random data partition, whereas all DDFP numbers come from a different 8:2 split of the same datasets. Consequently, per-table comparisons such as the DeepLab CT-to-MRI row (DDFP 0.8501 vs. FVP 0.7330) are not controlled experiments: differences could be due to which slices were used for testing rather than to the method. Please re-run these baselines under the identical split, or remove the starred rows from the headline SOTA claims and qualify the comparisons accordingly.","section":"4.1.4, Tables 1-2"},{"comment":"The pseudo-label loss is written with a confidence weight conf^{G_t}_{h,w}, while the preceding text and the pseudo-label selection in Eq. (9) use the preadapted model G'_s. Using the target model's own confidence for reweighting would be a different training signal, and the notation is inconsistent. Please correct the equation and state explicitly which model produces the confidence weights.","section":"3.5, Eq. (12)"},{"comment":"The trainable-layer selection is specified as the first four convolutional layers for the U-Net backbone, but the paper also reports DeepLabv3 results. The analogous style-related layers for ResNet50/DeepLabv3 are never defined, so the DeepLab rows in Tables 1-3 cannot be reproduced. Please specify the layer split for each backbone. I do not regard the heuristic nature of the style/content split itself as disqualifying, since Fig. 9 explores alternate layer selections, but the missing DeepLab specification is a concrete reproducibility gap.","section":"3.6, 4.1.3"},{"comment":"All quantitative results appear to be single-run; no error bars, seeds, or per-comparison statistical tests are given. The Wilcoxon test in Fig. 6 aggregates 3D Dice over four tasks with roughly 5 test volumes per task, which is not a substitute for per-table significance tests under matched test slices. Please provide at least three seeds with standard deviations, and clarify whether the aggregated test uses paired samples from the same test split.","section":"4.1.3, 4.5"}],"minor_comments":[{"comment":"The abstract and introduction state that code is available online, but no URL is provided; please include a repository link.","section":"Abstract, Section 1"},{"comment":"The text says that two sets of experiments (with and without fine-tuning of style-related layers) were conducted, but Table 5 reports only one set; the additional results should be shown or the sentence corrected.","section":"4.6.2, Table 5"},{"comment":"The method name is written 'TT-FSUDA' in the brain tumor section, while the rest of the paper uses 'TT-SFUDA'; please unify the naming.","section":"4.4"},{"comment":"Both Section 4.6.2 and Section 4.6.3 refer to 'Fig. 8' for different content; renumber the figures and update the cross-references.","section":"4.6.2, 4.6.3"},{"comment":"The phrase 'pseudo-label generalization' should read 'pseudo-label generation.'","section":"4.6.3"},{"comment":"The module name is written inconsistently as both 'DDFP' and 'DFFP'; please adopt one consistent abbreviation.","section":"3.4"},{"comment":"In Section 4.5, 'restuls' is a typo for 'results.'","section":"4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready in its present form because the headline comparisons against FVP and FSM are not controlled, and the reproducibility information (seeds, code URL, DeepLab layer specification) is incomplete. The proposed method is plausible, and the concerns are addressable within the scope of a revision, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper's real contribution is a per-image data-dependent frequency prompt for SFDA, which is genuinely new relative to FVP and ProSFDA's domain-level prompts. They combine it with BN pre-adaptation and style-layer fine-tuning in a way that is coherent and reasonably well ablated. The Dice gains on the controlled tasks (abdominal CT->MRI, cardiac MRI->CT) are large, and the efficiency argument (only shallow layers and prompt trainable) is plausible and supported by the FLOPs figure. I'd give them credit for the ablation structure: loss components, prompt design, BN init, and trainable layer selection are each separately tested.\n\nThe main soft spot is exactly the one the stress-test flags: the SOTA claim is built on comparisons where FVP and FSM numbers are quoted from [38]'s random data split, not re-run under the paper's 8:2 split. The paper discloses this with the asterisk, but it still puts the headline numbers on shaky ground. A gap of 0.8987 vs 0.8382 for ProContra is controlled, but the 0.8987 vs 0.7330 for FVP is not. That's load-bearing: the abstract says \"outperforms existing state-of-the-art methods.\" To support that, they need either to re-run those baselines on the same test slices or to soften the claim and make the split difference explicit in the main results. No error bars or multi-seed results make it worse; the aggregated Wilcoxon is thin with roughly 5 test volumes per task.\n\nThe style/content layer split is a mechanism-level risk, but it's less concerning because they ablate different selections in Fig. 9. The delta_cls inconsistency (40 vs 0.4) looks like a typo but should be fixed. And the \"code available online\" statement has no URL, which is a concrete reproducibility problem.\n\nWho's this for? Anyone working on source-free adaptation for medical segmentation, especially prompt-based style translation. They'll want to build on the per-image prompt idea. It's not a field re-organizer, but it's a solid incremental method with a genuinely new prompt formulation.\n\nRecommendation: send to a serious referee, with major revision. The referee should require re-running FVP/FSM (or any other quoted baselines) under the same split, reporting multi-seed statistics or error bars, providing the code link, and fixing the delta_cls typo. If those are addressed, the paper's central claim becomes credible.","headline":"A genuinely new per-image frequency prompt for SFDA, but the headline SOTA comparisons are not yet controlled: FVP/FSM baselines come from another split and no error bars are reported.","tokens_in":23862,"tokens_out":3440,"would_cite":true,"duration_ms":33028,"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":"This paper claims that a source-free domain adaptation pipeline combining batch-normalization preadaptation, data-dependent frequency prompts, and style-layer-only fine-tuning outperforms existing methods on cross-modality medical image…","keywords":["source-free domain adaptation","medical image segmentation","frequency prompt","data-dependent prompt","batch normalization calibration","pseudo-labeling","cross-modality domain adaptation","style-related layer fine-tuning"],"falsifier":"Compare DDFP against an adaptation method that trains the full network using the same pseudo-labels and prompted images on a dataset whose domain shift includes anatomical shape or label-distribution changes; if full-network training clearly wins, the fixed shallow/deep split is the limiting assumption. Concretely, on a target dataset with substantially different organ proportions or pathology, compute Dice of the frozen-deep version against the same method with deeper layers trainable.","tokens_in":22811,"feed_emoji":"🩻","tokens_out":7075,"duration_ms":70662,"temperature":0.7,"pith_summary":"This paper tries to establish that source-free domain adaptation (SFDA) for medical image segmentation works best when the domain gap is attacked at three different moments: recalibrate the source model's batch-normalization statistics on target data before training begins; translate each target image into a source-like style with a per-image, data-dependent prompt in the frequency domain; and fine-tune only the shallow style-related layers of the target network. The claimed payoff is that this combination outperforms existing SFDA methods on cross-modality abdominal, cardiac, and brain tumor segmentation, including an average Dice of 0.8987 versus 0.8382 for the best prior method on abdominal CT-to-MRI with a U-Net. A curious reader would care because in medical settings the labeled source data may be unavailable for privacy reasons, and the paper argues the previous fixes—one fixed style prompt for the whole domain and whole-model fine-tuning—leave both style transfer and training efficiency on the table.","feed_headline":"Per-image frequency prompts lift cross-modality segmentation Dice","feed_subtitle":"Source-free abdomen and heart MRI/CT tests beat prior adaptation methods using only prompted images and reliable pseudo-labels.","key_machinery":"The load-bearing mechanism is the data-dependent frequency prompt generation module. For each target image it takes the FFT amplitude and phase spectra, processes them through separate 1x1-convolution networks, concatenates the features with a learnable domain-dependent frequency prompt, passes them through a fusion network, and extracts a per-image prompt that is blended with the shared prompt via a skip connection. This prompt is applied multiplicatively to the amplitude spectrum, and an inverse FFT reconstructs a source-like target image. The other two carrying components are the BN preadaptation step, which recalibrates the source model's running statistics on target data before any training, and the style-related layer fine-tuning strategy, which freezes everything except the first four convolutional layers and the prompt parameters.","core_discovery":"The paper's central claim is that a single domain-level style prompt is the bottleneck, and that conditioning the prompt on each image's Fourier amplitude and phase spectra removes that bottleneck. The DDFP module computes a shared trainable frequency prompt and then adjusts it per image through small 1x1-convolution networks before multiplying the result onto the image's amplitude spectrum and reconstructing the image with an inverse FFT. The same framework gets a second boost by using a batch-normalization-preadapted model, rather than the raw source model, to initialize the target model and to produce filtered pseudo-labels. With only the first four convolutional layers trainable, the paper reports top average Dice on both abdominal and cardiac cross-modality tasks and superior results on brain tumor Flair/T2 adaptation.","pith_inferences":["The paper fixes the first four convolutional layers as style-related; the natural follow-up I would draw is to make that split adaptive, for example by selecting trainable layers from target-data BN-statistic drift or gradient sensitivity, so the method could handle shifts that penetrate deeper into the network.","Because the prompts act on the low-frequency part of the amplitude spectrum, a minimal testable baseline would be a per-image histogram or affine intensity matching in the frequency domain; if that baseline matched DDFP, the learned prompt's advantage would narrow considerably.","The paper evaluates Dice and ASD and notes blurred boundaries, so an implicit extension is to add boundary- or contour-aware supervision to the prompted-image pipeline, which could improve the ASD numbers without changing the core mechanism."],"forward_implications":["On the paper's own numbers, a fixed source model can reach average Dice around 0.90 on abdominal CT-to-MRI and 0.85 on cardiac MRI-to-CT without ever seeing source data again, provided target-model training uses prompted images and reliable pseudo-labels.","Adaptation cost drops because only a few shallow layers and small prompt modules are trainable, so the trainable FLOPs are much lower than whole-model fine-tuning.","The gain should be largest when the domain gap is large: BN preadaptation gives a big boost in hard directions such as abdominal MRI-to-CT and a small boost in easy directions.","A single shared domain prompt is not enough; the paper's ablation suggests conditioning the prompt on each image's amplitude and phase spectra is what closes the remaining gap.","The style/content layer split means the method should work best for shifts that are mostly low-level appearance changes, which is precisely the cross-modality setting tested here."],"supporting_citations":[{"why":"Supplies the frequency-domain visual prompting baseline that DDFP extends, and the source of the FVP comparison numbers.","marker":"[38]"},{"why":"Supplies the BN-statistic loss objective and a Fourier style mining baseline whose two-stage design DDFP refines.","marker":"[42]"},{"why":"Supplies the contrastive SFDA baseline that DDFP is measured against on Dice in the main tables.","marker":"[47]"},{"why":"Supplies the BN statistic recalibration procedure that the BN preadaptation step adapts for initialization and pseudo-label generation.","marker":"[22]"},{"why":"Supplies the prompt-based SFDA baseline in the spatial domain that motivates testing prompts in the frequency domain.","marker":"[13]"},{"why":"Supplies the premise that shallow features encode style and deep features encode content, which justifies freezing all but the first four layers.","marker":"[30]"},{"why":"Supplies the U-Net backbone and the layer structure used to define trainable style-related layers versus frozen content layers.","marker":"[32]"},{"why":"Supplies the CHAOS MRI/CT abdominal dataset used for the main evaluations and ablations.","marker":"[17]"},{"why":"Supplies the MMWHS cardiac CT/MRI dataset used for the cardiac evaluations.","marker":"[49]"},{"why":"Supplies the BraTS brain tumor dataset used for the third evaluation.","marker":"[27]"}],"fun_headline_variants":["Per-image frequency prompts adapt medical segmentation without source data","Data-dependent frequency prompts improve source-free domain adaptation","Frequency prompts per image beat prior source-free adaptation methods","Cross-modality segmentation lifted by per-image frequency prompts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes the domain gap lives in low-level style and that shallow layers (the first four convolutions of a U-Net) carry it while deep content layers can be frozen, so if a shift also alters content-level features the frozen layers cannot be repaired.","fun_headline_variants_meta":{"raw":{"variants":["Per-image frequency prompts adapt medical segmentation without source data","Data-dependent frequency prompts improve source-free domain adaptation","Frequency prompts per image beat prior source-free adaptation methods","Cross-modality segmentation lifted by per-image frequency prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2743,"prompt_tokens":1005,"completion_tokens":1738,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":1676}},"tokens_in":621,"tokens_out":1738,"duration_ms":12770,"temperature":1.0,"reasoning_tokens":1676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:20:33.756829+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare DDFP against an adaptation method that trains the full network using the same pseudo-labels and prompted images on a dataset whose domain shift includes anatomical shape or label-distribution changes; if full-network training clearly wins, the fixed shallow/deep split is the limiting assumption. Concretely, on a target dataset with substantially different organ proportions or pathology, compute Dice of the frozen-deep version against the same method with deeper layers trainable.","supporting_citations":[{"cited_title":"Fvp:Fouriervisualpromptingforsource-freeunsupervised domain adaptation of medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the frequency-domain visual prompting baseline that DDFP extends, and the source of the FVP comparison numbers."},{"cited_title":"Source free domain adaptationformedicalimagesegmentationwithfourierstylemining","cited_arxiv_id":null,"evidence_quote":"Supplies the BN-statistic loss objective and a Fourier style mining baseline whose two-stage design DDFP refines."},{"cited_title":"Source-free domain adaptation for medical image segmentation via S","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive SFDA baseline that DDFP is measured against on Dice in the main tables."},{"cited_title":"Two at once: Enhancing learning and generalization capacities via ibn-net, in: Proceedings of the european conference on computer vision (ECCV), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the premise that shallow features encode style and deep features encode content, which justifies freezing all but the first four layers."},{"cited_title":"Chaos challenge-combined (ct-mr) healthy abdominal organ segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the CHAOS MRI/CT abdominal dataset used for the main evaluations and ablations."},{"cited_title":"Multi-scale patch and multi-modality atlases for whole heart segmentation of mri","cited_arxiv_id":null,"evidence_quote":"Supplies the MMWHS cardiac CT/MRI dataset used for the cardiac evaluations."}],"review_version":1}