{"id":"57c78b0e-7748-4f2f-aca3-ace23a6905be","arxiv_id":"2506.09403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SRPL-SFDA refines source-model pseudo-labels with SAM and consistency-based reliability selection, reporting near-supervised adaptation on two MRI segmentation benchmarks.","lead":"An SFDA method for medical image segmentation uses SAM to refine pseudo-labels from a source model, selecting reliable regions by comparing SAM outputs under three intensity enhancements. It reports Dice scores close to target-supervised training on fetal brain and prostate MRI datasets, though model selection uses target validation labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SFDA performance is selected using target-domain validation labels (Sec. 4.2); the 'close to supervised training' claim is therefore an upper bound and not established for a truly source-free protocol.","rationale":"The reader's weakest_assumption (CMSO) is a genuine methodological weakness, but it is not the most load-bearing issue for the paper's headline claim. Even if the consistency assumption is imperfect, the method could still outperform the baselines, and the ablation in Table 5 provides some empirical support that the reliable-region supervision helps. By contrast, the evaluation protocol directly determines whether the reported numbers support the abstract's unqualified claim. Section 4.2 states that the hyperparameter and checkpoint are selected using the target-domain validation set, and the Discussion admits this. With only 4 and 5 validation volumes, this selection can overfit, and no source-free alternative selection criterion is offered. Therefore the reported Dice values are conditional on access to target labels, which violates the source-free premise and weakens the comparison to supervised training. This is why I recommend keeping the conditional verdict but re-anchoring the stated concern on the evaluation protocol rather than on CMSO.","tokens_in":19595,"tokens_out":9450,"duration_ms":99279,"concrete_test":"Rerun SRPL-SFDA on both datasets with model selection performed without target labels, e.g., using a fixed epoch count (say 300) or a target-training-set statistic such as mean prediction entropy or pseudo-label confidence, and compare test Dice against Tables 2 and 3. Also report whether comparable SFDA baselines were selected with the same target-validation protocol. If the Dice gap to \"Target only\" widens by more than 1.5 points, or if SRPL-SFDA no longer significantly outperforms UPL-SFDA, the central claim is an artifact of target-label-based selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SRPL-SFDA approaches supervised performance is not supported under a strictly source-free protocol. In Sec. 4.2, the authors state: \"According to the best performance on the validation set of the target domain, the hyper-parameter setting was λ=10.0, and the corresponding checkpoint was employed for inference.\" This means target-domain labels are used for both hyperparameter tuning and checkpoint selection, directly contradicting the source-free setting. The validation sets are tiny: 4 volumes for prostate D,E,F and 5 for fetal brain (Table 1), so selecting from 300 epochs on 4-5 labeled volumes can overfit and inflate test Dice. Moreover, the supervised references \"Target only\" and \"Fine-tune\" use target labels for training, so comparing to them while peeking at target labels for selection is not a fair assessment of the unsupervised adaptation gain. The Discussion acknowledges this limitation but the Abstract and headline claim do not qualify it. Without target validation labels, the reported 82.22% and 94.33% Dice are an upper bound, and the \"close to supervised\" claim is not established.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SRPL-SFDA, a source-free domain adaptation method for medical image segmentation. The method has three main components: Test-Time Tri-branch Intensity Enhancement (T3IE), which applies histogram equalization and two gamma corrections to target images before inference; SAM-guided pseudo-label refinement, where the source model's averaged T3IE prediction is used to form a bounding-box prompt and the three T3IE channels are concatenated as an SAM-compatible RGB input; and Consistency of Multiple SAM Outputs (CMSO), which identifies reliable pseudo-label pixels as those where the three SAM outputs agree, followed by a Reliability-Aware Pseudo-Label Supervision and Regularization (RPSR) loss. Experiments are reported on two MRI datasets, prostate (sites A/B to D/E/F, plus site C) and fetal brain (TrueFISP to HASTE). The reported results show that SRPL-SFDA outperforms four SFDA baselines and approaches the Dice of supervised target-only training and fine-tuning, with ablations on pseudo-label quality, loss components, hyper-parameter lambda, and SAM prompt types. The code is publicly available.","tokens_in":19773,"tokens_out":4296,"duration_ms":52715,"significance":"If the reported numbers are valid under a genuinely source-free protocol, the paper makes a useful empirical contribution: it demonstrates a concrete recipe for improving pseudo-labels with a frozen SAM through intensity enhancement and consistency-based reliable-region mining, and the ablation studies are internally coherent and informative. The release of code and the model-agnostic framing are additional strengths. However, the headline claim of approaching supervised performance is currently conditional on using target-domain validation labels for hyper-parameter and checkpoint selection, which is disclosed in Section 4.2. This weakens the source-free claim as stated and must be addressed before the result can be taken at face value.","major_comments":[{"comment":"The implementation details state: 'According to the best performance on the validation set of the target domain, the hyper-parameter setting was lambda=10.0, and the corresponding checkpoint was employed for inference.' This uses target-domain labels for both hyper-parameter tuning and model selection, which is not a source-free protocol. The abstract's claim that SRPL-SFDA 'is close to that of supervised training in the target domain' is therefore an upper bound under oracle selection, not an established result for a method that never sees target labels. I request that the authors either (a) re-run the main comparisons without any target-label-based selection, for example by fixing lambda via the source validation set or by reporting the last-epoch model and the sensitivity of test Dice to lambda, or (b) clearly reframe all headline results as 'oracle-selected' and remove or qualify the comparison to supervised training. This is load-bearing because it concerns the central claim of the paper.","section":"Sec. 4.2"},{"comment":"The CMSO reliability criterion assumes that agreement among R_He, R_gammaD, and R_gammaS certifies correctness. However, all three SAM outputs are produced from the same bounding-box prompt B, which is derived from the same source-model pseudo-label Y. A biased or incomplete box prompt is therefore shared by all three outputs, so agreement can reflect consistency rather than correctness. The Discussion acknowledges this possibility but provides no quantitative evidence that the consensus region is actually more accurate than the non-consensus region. Please add an analysis on the target validation set that compares segmentation accuracy inside Omega_C versus Omega_U, or an experiment that perturbs the box prompt and shows that consensus remains correlated with correctness. Without this, the RPSR supervision in Eq. (13) may be training on trusted noise in precisely the way the method is designed to avoid.","section":"Sec. 3.2.1, Eq. (7)"},{"comment":"The target validation sets are very small: 4 volumes for prostate sites D/E/F and 5 volumes for fetal brain. Selecting lambda and a checkpoint from 300 epochs on such small sets risks overfitting the validation set, and the reported test Dice may therefore be inflated relative to what a truly unsupervised protocol would achieve. The paper reports a single random split with no repeated-seed or repeated-split variability. Please report results across multiple data splits or random seeds, or at least show that the test performance is flat around lambda=10.0 and that checkpoint selection is not a sharp peak on the validation curve. This is important for assessing whether the reported gains over baselines are robust rather than an artifact of small-sample oracle selection.","section":"Sec. 4.2, Table 1"}],"minor_comments":[{"comment":"The text at the start of this subsection is corrupted: 'We conducThese values represent the overall distributioe performance' is an incomplete and duplicated fragment that makes the prompt-type ablation description unreadable. Please rewrite this paragraph.","section":"Sec. 4.4.4"},{"comment":"The final method in Table 5 reports Dice of 84.46% on prostate and 92.58% on fetal brain, whereas the same method in Tables 2 and 3 reports 82.22% and 94.33%. It appears Table 5 is evaluated on the target validation set while Tables 2 and 3 are on the test set, but this is not stated in the captions or text. Please clarify the evaluation protocol for each table, as the discrepancy is currently confusing.","section":"Tables 2, 3, and 5"},{"comment":"The 'SAM(X)-BB prompt' baseline uses ground-truth bounding boxes expanded by 5-10 pixels to simulate user interaction. This is an oracle-like upper reference for SAM's prompt-based capability rather than an SFDA baseline, and the comparison should be described as such: it shows how much of the gain comes from the pseudo-label-derived box rather than from user-provided box prompts.","section":"Sec. 4.3"},{"comment":"The claim that SRPL-SFDA is 'close to supervised training' is based on Dice gaps of 0.80-3.08 percentage points, but no significance test is reported against 'Target only' or 'Fine-tune.' Please either add such tests or soften the wording to describe the observed Dice gap without implying statistical equivalence.","section":"Sec. 4.3"},{"comment":"The related work discusses DPL, CLR, and CCMT but the experimental comparison includes only PTBN, TENT, AdaMI, and UPL-SFDA. Please state explicitly why the additional pseudo-label-based SFDA methods are not included in the comparison, or add them if feasible.","section":"Sec. 2.2"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the target-domain validation leakage disclosed in Section 4.2. The paper is otherwise well-executed and the ablation study is coherent, but the abstract's central claim of 'close to supervised training' cannot stand under a strictly source-free protocol unless the experiments are rerun without oracle-based selection or the claims are explicitly conditioned on using target validation labels. I would be willing to revisit after the authors address this concern with additional experiments or a careful reframing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent method paper with a real contribution: using SAM as a zero-shot pseudo-label refiner inside an SFDA loop, with T3IE to make medical images SAM-compatible and CMSO to mark reliable regions. The ablation chain is coherent, the gains over strong baselines are consistent on two datasets, and the authors are honest that RPL-SFDA is their own preliminary work. That honesty matters, because self-extension here is a feature, not a bug.\n\nWhat is actually new: nobody else, as far as I can tell, runs SAM at test time on three T3IE-augmented channels, uses the source model's own box to prompt it, and then supervises only the consistent regions. The T3IE trick is simple and plausible, and Table 4 tells a clean story: each component adds a couple of Dice points. The code is linked. That's enough to make the paper useful.\n\nThe soft spots are real but not fatal. The biggest one is in Section 4.2: lambda=10 and the final checkpoint are picked using the target-domain validation set. That is standard practice in a lot of the literature, but the paper sells itself as source-free, and with validation sets of 4-5 volumes, selection over 300 epochs can easily inflate test numbers. So the headline \"close to supervised training\" is an upper bound under a strictly unsupervised protocol. The Discussion admits this, but the Abstract doesn't qualify it. I would want to see results with no target validation labels, plus multiple seeds.\n\nThe other load-bearing assumption is CMSO, Eq. (7): the three SAM outputs share the same box prompt B, which comes from the initial pseudo-label Y. Agreement can therefore certify consistency rather than correctness. The authors address this in the Discussion by saying the box itself is derived from the enhanced images, which is true, but a shared biased box can still propagate the same error to all three outputs. I don't think this kills the method—Table 5 shows RPL beats PL(R), which is evidence the consistency filter helps—but it deserves a direct test, e.g., perturbing the box or comparing against a single SAM output.\n\nOverall: the paper is a serious empirical contribution to SFDA for medical segmentation. The math is simple and the method is reproducible. I'd send it to review, and I'd tell the reviewers to focus on the target-validation leak and the shared-box confound, not on the SAM machinery. Whoever reads it for the pseudo-labeling angle will get something useful.","headline":"A solid empirical SFDA recipe with a disclosed but real target-validation leak in model selection; worth refereeing, but the 'close to supervised' claim needs a fully unsupervised protocol.","tokens_in":20344,"tokens_out":1751,"would_cite":true,"duration_ms":18986,"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":"A SAM-guided pseudo-label pipeline for source-free medical image segmentation reaches Dice within about one point of supervised target-domain training on two multi-site MRI benchmarks.","keywords":["source-free domain adaptation","medical image segmentation","Segment Anything Model","pseudo-label refinement","test-time intensity enhancement","consistency-based reliability","MRI domain shift","reliability-aware training"],"falsifier":"On a held-out target test set with ground truth, compute the Dice of the consensus region $\\Omega_C$ against the true mask while also recomputing SRPL-SFDA with the box prompt $B$ shifted by 10 pixels: if $\\Omega_C$ is not more accurate than the full refined pseudo-label $R$, or if the 'reliable' labels move with the box, the CMSO reliability claim is falsified.","tokens_in":19333,"feed_emoji":"🩻","tokens_out":14843,"duration_ms":137673,"temperature":0.7,"pith_summary":"The paper claims that a frozen Segment Anything Model (SAM) can be turned into a reliable pseudo-label refiner for source-free domain adaptation (SFDA) in medical image segmentation, so that adapting a pretrained model to a new clinical site needs neither the source images nor target annotations. The method applies three intensity transforms to each target image, uses their averaged source-model predictions to draw a bounding box, and sends the three enhanced copies concatenated as an RGB-like image to SAM for refined boundaries. A consistency check over three SAM outputs selects a reliable pixel region, and a reliability-aware loss trains the target model there while entropy-minimizing the remaining ambiguous regions. On prostate and fetal-brain MRI benchmarks the adapted model reaches 82.22% and 94.33% Dice, compared with 83.02% and 95.53% for training directly on labeled target data. The authors note that their checkpoint selection uses a labeled target-domain validation set, a practical constraint acknowledged in the paper.","feed_headline":"SAM-corrected labels nearly match supervised medical segmentation","feed_subtitle":"No source images needed: adapted accuracy lands within about a point of supervised training.","key_machinery":"The load-bearing mechanism is the T3IE triple used twice. For each target image, histogram equalization, domain-adaptive gamma correction (aligning image mean to dataset mean), and SAM-compatible gamma correction (pushing pixel statistics toward natural-image mean 0.5 and standard deviation 0.29) produce three enhanced copies. Averaged source-model predictions from these copies yield the initial pseudo-label and the box prompt; the copies concatenated as $X_{\\text{RGB}}$ yield SAM's refined mask. The same three copies, run through SAM separately with an identical box, produce the CMSO agreement map: a pixel is reliable ($\\Omega_C$) exactly when $R_{\\text{He}}=R_{\\gamma_D}=R_{\\gamma_S}$, and unreliable otherwise. The RPSR loss completes the training: partial cross-entropy plus partial Dice on $\\Omega_C$, and partial entropy minimization on $\\Omega_U$ with weight $\\lambda=10$.","core_discovery":"The central claim is that SAM's zero-shot, prompt-based segmentation, although trained on natural images, can correct medical pseudo-labels tightly enough that source-free adaptation nearly matches supervised target training. The source model's averaged predictions over the three T3IE-enhanced copies of a target image give an initial pseudo-label; its bounding box becomes the prompt for SAM, and the same three enhanced copies concatenated into a pseudo-RGB image become SAM's input. The refined mask $R$ is then filtered by consistency: each enhanced copy is sent to SAM separately with the same box, and only pixels $\\Omega_C$ where all three outputs agree are used as supervision, through a combination of partial cross-entropy and partial Dice, while the remaining pixels $\\Omega_U$ are regularized by entropy minimization. With this recipe the method reaches 82.22% Dice on prostate sites D/E/F versus 83.02% for target-only supervised training and 94.33% on fetal brain versus 95.53%, and it reports lower average surface distance than the prior SFDA methods on the prostate benchmark.","pith_inferences":["Beyond the paper: the same consensus-based selection could be applied to other promptable segmenters such as medical-SAM variants, with only the input-transform statistics needing recalibration.","Beyond the paper: the SAM-compatible gamma target of mean 0.5 and standard deviation 0.29 is tuned to natural images; on CT, PET, or ultrasound the same T3IE idea would likely need modality-specific reference statistics.","Beyond the paper: since the paper selects checkpoints with a labeled target validation set, a deployment-ready variant would need a label-free selection proxy, such as consensus-region entropy or agreement on the unlabeled training set."],"forward_implications":["A clinic deploying a segmentation model could adapt it to a new scanner using only unlabeled target scans, avoiding the transfer of patient data from the original site.","Future promptable segmenters can be dropped into the same recipe without retraining, since SAM itself is used frozen and only the target segmentation model is updated.","The threshold-free consensus rule means reliability estimation does not need a tuned confidence cutoff, making the method less sensitive to hyperparameter choice.","Because boundary-sensitive losses run only on the consensus region, the adapted model can learn sharper edges than the source model provides, which matters for contour-based clinical tasks."],"supporting_citations":[{"why":"Supplies SAM, the frozen promptable segmenter whose zero-shot outputs are used as refined pseudo-labels.","marker":"Kirillov et al., 2023"},{"why":"The earlier RPL-SFDA method this work extends; supplies the base pseudo-label pipeline and the strongest prior baseline.","marker":"Liu et al., 2024"},{"why":"UPL-SFDA, the uncertainty-aware pseudo-label SFDA baseline that the reported results are compared against.","marker":"Wu et al., 2023c"},{"why":"MS-Net multi-site prostate MRI dataset, which defines the source and target domain splits for the prostate experiments.","marker":"Liu et al., 2020"},{"why":"Fetal brain dataset with TrueFISP/HASTE protocols, providing the second source-target pair and the UPL-TTA baseline.","marker":"Wu et al., 2023a"},{"why":"Test-time augmentation uncertainty estimation, the idea behind using multiple perturbed SAM outputs for reliability.","marker":"Wang et al., 2019"},{"why":"Natural-image pixel statistics (mean 0.5, standard deviation 0.29) that set the SAM-compatible gamma correction target.","marker":"Sener et al., 2022"},{"why":"TENT entropy-minimization baseline and the origin of the entropy regularizer applied to unreliable pseudo-label regions.","marker":"Wang et al., 2021"},{"why":"AdaMI adaptive mutual-information SFDA method, another baseline in the comparison tables.","marker":"Bateson et al., 2022"},{"why":"PTBN prediction-time batch-normalization baseline, the first comparison method in the SFDA evaluation.","marker":"Nado et al., 2020"}],"fun_headline_variants":["SAM-corrected pseudo-labels near supervised segmentation","Source-free segmentation nears supervised accuracy with SAM","SAM guides reliable pseudo-labels, closing supervised gap","Reliable SAM pseudo-labels rival supervised segmentation","Zero-shot SAM refines pseudo-labels, nearing supervised accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that when the three SAM runs agree on a pixel, that pixel is correct, but all three runs share the same bounding-box prompt derived from the source model's pseudo-label, so agreement can certify shared bias rather than anatomical truth.","fun_headline_variants_meta":{"raw":{"variants":["SAM-corrected pseudo-labels near supervised segmentation","Source-free segmentation nears supervised accuracy with SAM","SAM guides reliable pseudo-labels, closing supervised gap","Reliable SAM pseudo-labels rival supervised segmentation","Zero-shot SAM refines pseudo-labels, nearing supervised accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00112,"raw_usage":{"total_tokens":4727,"prompt_tokens":1079,"completion_tokens":3648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":3572}},"tokens_in":695,"tokens_out":3648,"duration_ms":27306,"temperature":1.0,"reasoning_tokens":3572,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:49:16.553235+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out target test set with ground truth, compute the Dice of the consensus region $\\Omega_C$ against the true mask while also recomputing SRPL-SFDA with the box prompt $B$ shifted by 10 pixels: if $\\Omega_C$ is not more accurate than the full refined pseudo-label $R$, or if the 'reliable' labels move with the box, the CMSO reliability claim is falsified.","supporting_citations":[{"cited_title":", author Mintun, E","cited_arxiv_id":null,"evidence_quote":"Supplies SAM, the frozen promptable segmenter whose zero-shot outputs are used as refined pseudo-labels."},{"cited_title":", author Wu, J","cited_arxiv_id":null,"evidence_quote":"The earlier RPL-SFDA method this work extends; supplies the base pseudo-label pipeline and the strongest prior baseline."},{"cited_title":", author Dou, Q","cited_arxiv_id":null,"evidence_quote":"MS-Net multi-site prostate MRI dataset, which defines the source and target domain splits for the prostate experiments."},{"cited_title":", author Li, W","cited_arxiv_id":null,"evidence_quote":"Test-time augmentation uncertainty estimation, the idea behind using multiple perturbed SAM outputs for reliability."},{"cited_title":", author Chatterjee, D","cited_arxiv_id":null,"evidence_quote":"Natural-image pixel statistics (mean 0.5, standard deviation 0.29) that set the SAM-compatible gamma correction target."},{"cited_title":", author Shelhamer, E","cited_arxiv_id":null,"evidence_quote":"TENT entropy-minimization baseline and the origin of the entropy regularizer applied to unreliable pseudo-label regions."},{"cited_title":", author Kervadec, H","cited_arxiv_id":null,"evidence_quote":"AdaMI adaptive mutual-information SFDA method, another baseline in the comparison tables."}],"review_version":1}