{"id":"b4d55fad-5506-4010-916f-f7093ed11f2c","arxiv_id":"2505.08527","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"DFG automatically finds bounding-box prompts for MedSAM and uses the refined masks as pseudo-labels, improving source-free cross-domain medical image segmentation.","lead":"This paper introduces DFG, a method that uses MedSAM to refine pseudo-labels in source-free domain adaptation by automatically searching for an accurate bounding box prompt using features from both the target model and MedSAM. It reports large Dice improvements over prior SFDA methods on cross-modality abdominal and prostate MRI segmentation benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA margin is not yet robust: single-seed, test-tuned hyperparameters plus known stable-interval failures leave DFG's advantage over baselines unquantified.","rationale":"The reader's weakest assumption points to the stable-interval property, and the paper's own failure-case section supports that concern. However, the more load-bearing issue for the central empirical claim is that the reported SOTA margin comes from a single seed with hyperparameters tuned on the same benchmark and no validation split, so run-to-run variance and sensitivity to hyperparameter choice are unknown. The code is available, which makes a multi-seed and perturbation check feasible. The paper has independent support through detailed ablations (Tables IV-VII) and a released implementation, so the concern does not warrant rejection; it strengthens the case for the reader's CONDITIONAL verdict. I therefore recommend keeping the verdict unchanged while requiring the robustness check before the margin over SOTA is accepted.","tokens_in":19613,"tokens_out":8958,"duration_ms":103084,"concrete_test":"Using the released code, rerun DFG and the strongest reimplemented baseline (ProtoContra) on CHAOS->BTCV with 5 random seeds under the paper's fixed hyperparameters, and also run small perturbations around each tuned value (e.g., tau_f in {0.9, 0.99, 0.999}, tau_max in {0.3, 0.35, 0.4}, p_Delta in {0.003, 0.005, 0.01}). Report mean and standard deviation of Dice/ASSD. If the best-configuration margin over ProtoContra is within 2 Dice points of the perturbation range or overlaps at one standard deviation, the SOTA claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the large Dice/ASSD margin over SFDA baselines in Tables I-III. Two facts put that margin at risk. First, all reported numbers are single-seed, and the hyperparameters (tau_f, tau_max, p_Delta, r, tau_Delta, tau_Div; Section V-A3, Fig. 6) are tuned on the same benchmark. Because target labels are unavailable in SFDA, there is no principled validation split, so the reported configuration is effectively selected on the test split and run-to-run variance is unknown. Second, the mechanism that is supposed to justify the gain, MedSAM's stable-interval property (Eq. 5), is explicitly acknowledged to fail in Section V-C7 and Fig. 7: false stable intervals and insufficiently stable MedSAM outputs are documented. The stability heuristic is therefore an empirical prior with an unmeasured failure rate, not a guarantee. If seed variance or hyperparameter sensitivity is comparable to the 8-10 point Dice gap over ProtoContra, the headline advantage cannot be attributed to DFG. If, on the other hand, the margin is stable across seeds and neighboring hyperparameters, the claim holds despite the heuristic's occasional failures.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DFG, a source-free domain adaptation (SFDA) method for medical image segmentation that uses the frozen medical foundation model MedSAM as an external teacher. A feature aggregation (FA) phase adapts the source model to the target feature distribution, after which two box-prompt search procedures, TBS over target model features and MBS over MedSAM features, progressively expand a bounding box until a stable interval in MedSAM output is detected via Eq. (5). The resulting pseudo-labels are post-processed by connectivity analysis (CP) and used to train the target model with a Dice loss. Experiments on three abdominal MRI/CT adaptation tasks and one prostate MRI task report average Dice of 84.9%, 85.2%, 85.7%, and 93.3%, outperforming the compared SFDA baselines.","tokens_in":19884,"tokens_out":7698,"duration_ms":75474,"significance":"If the empirical results are robust, the paper offers a practically useful way to inject a frozen medical foundation model into SFDA, and it does so with an original combination: using both MedSAM features and MedSAM outputs for auto-prompting, rather than fine-tuning SAM on pseudo-labels. The paper includes an ablation study, an efficiency analysis, and an unusually candid failure-case section, and the code is released. However, the headline margin over strong baselines is not yet robustly established: all results are single-run, the many hyperparameters appear to be selected using the test benchmarks, and the central stable-interval assumption is explicitly shown to fail in some cases. The contribution is therefore plausible but needs stronger empirical support before the SOTA claims can be accepted.","major_comments":[{"comment":"The hyperparameters of DFG (tau_f=0.99, tau_Div=2.5, p_Delta=0.005, r=4, tau_max=0.35/0.3, tau_Delta=15/30/45, kappa=1/10, and the margin m) appear to be selected with access to the test labels of the same benchmarks, since Section V-A3 gives no label-free validation criterion and Fig. 6 sweeps values on the CHAOS-to-BTCV task and reports final test Dice. Because SFDA assumes target labels are unavailable, this selection protocol makes the reported margins over the baselines optimistic and their size unquantified. Please introduce a validation protocol that does not use target test labels (e.g., a frozen target split, pseudo-label stability, or an unsupervised criterion), and report the sensitivity of the headline margins over the Fig. 6 ranges on all three benchmarks.","section":"V-A3, Fig. 6, Tables I-III"},{"comment":"All reported numbers are single-run results with no standard deviations or seed counts. Some of the claimed advantages are modest (for example, CT-to-MRI average Dice is 85.2% for DFG versus 80.5% for ProtoContra, and several component gains in Table IV are only one to three points), so run-to-run variance could affect the ranking. Please report mean and standard deviation over at least three seeds for DFG and the reimplemented baselines, and include the top two or three hyperparameter configurations from Fig. 6.","section":"Tables I-VIII"},{"comment":"The stopping rule in Eq. (5) assumes that MedSAM output is stable exactly when the box prompt is near the true organ boundary, but Section V-C7 and Fig. 7 document three failure modes: organ features too variable to be included in the propagation, MedSAM output not stable enough near the ground truth, and a false stable interval appearing before the boundary. Because this stability property is the load-bearing mechanism that converts box search into pseudo-label refinement, the manuscript should quantify its failure rate (e.g., the proportion of organ instances for which the selected box is consistent with the ground-truth box on the target training data) and demonstrate that the average gains are not driven primarily by slices where the heuristic succeeds by chance.","section":"V-C7, Fig. 7, Eq. (5)"},{"comment":"The FVP baseline is not reimplemented: Section V-A3 states that FVP results are taken directly from the original paper, while DPL, AdaMI, UPL, and ProtoContra are reimplemented with tuned batch sizes under the same network and augmentations. If FVP used a different backbone, training schedule, or evaluation protocol, the comparison in Tables I-III is not fully controlled. Please either reimplement FVP under the identical protocol or list the protocol differences and their expected effect on the comparison; also report the chosen batch sizes for all reimplemented baselines and the validation criterion used to select them.","section":"V-A3, Tables I-III"}],"minor_comments":[{"comment":"The phrase 'for the first time explore the potentials of Segment Anything Model for SFDA' should be softened, since the related-work section cites auto-prompting adaptation methods such as MedSAM-U, and the novelty claim is not necessary for the paper's contribution.","section":"Abstract, Section I"},{"comment":"In Eq. (4), the condition 'for all i' in I_j' is too strong if read literally, because a candidate pixel i would need to satisfy both conditions with every pixel already in I_j rather than with at least one neighboring pixel; please clarify the intended neighbor criterion.","section":"IV-C1, Eq. (4)"},{"comment":"The failure-case examples in Fig. 7 are useful, but the figure would be easier to interpret if the final selected box prompt and the ground-truth box prompt were drawn with distinct line styles in a single panel; currently the blue search boxes and the red ground-truth box are difficult to distinguish in the middle row.","section":"V-C7, Fig. 7"},{"comment":"The temperature kappa is set to 10 for the prostate dataset, but its sensitivity is not included in Fig. 6; please add an ablation or state the range over which the final prostate result is stable.","section":"V-A3, Fig. 6"},{"comment":"Please specify how 3D Dice and ASSD are computed (e.g., per-volume averaging and whether slices without foreground are excluded) so that the comparison with FVP and ProtoContra, which may use different averaging conventions, is unambiguous.","section":"V-A4"},{"comment":"The connectivity-based post-processing assumes each foreground class is a single connected component; this is reasonable for the tested organs, but the manuscript should state the limitation for anatomies with multiple disconnected parts, since CP could remove true positive regions.","section":"IV-D, Table VII"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of TMI, and the authors are unusually candid about failure cases, which I view positively. The main reason for major revision is not conceptual soundness but the robustness of the empirical claims: test-set hyperparameter selection and single-seed reporting are common in this literature, but they should not be left unaddressed when the headline is a SOTA margin over strong baselines. If the authors supply multi-seed results and a label-free validation protocol, I would expect the paper to be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result is real but the margin isn't pinned down yet. The paper is the first I know of to get a frozen MedSAM to help SFDA via automatic box prompting, and the dual feature guided search is a genuinely new mechanism. It deserves a serious referee. But the reported 8-10 point Dice gain over ProtoContra needs multi-seed evidence before I'd treat it as robust.\n\nWhat's new: the TBS/MBS split, using target model features for class-wise clustered features and MedSAM features for dispersed ones, the stable-interval criterion for stopping box expansion, and the connectivity postprocessing. The ablations are thorough and the authors honestly show failure cases in Fig. 7. Code is released. That's a solid empirical package.\n\nSoft spots: all numbers are single-seed. The hyperparameters (tau_f, tau_max, p_Delta, r, tau_Delta, tau_Div) are tuned on the same benchmark, effectively on the test split, since SFDA lacks a principled validation set. Fig. 6 shows real sensitivity (tau_f=0.99 vs 0.999, etc.), so the exact margin could shift with seed or neighbor hyperparameters. FVP numbers are taken from the original paper rather than reimplemented, which is a minor caveat. The load-bearing stable-interval property is an empirical prior, not a guarantee; the paper's own Fig. 7 documents false stable intervals and unstable MedSAM outputs. That means the mechanism is messier than the clean narrative in Fig. 1, but it doesn't invalidate the approach.\n\nOverall: this is a useful contribution for SFDA with foundation models, and the failure analysis is refreshingly honest. I'd send it to reviewers with the request to add multi-seed runs with error bars and to specify how hyperparameters were chosen without peeking at target labels. The core idea deserves publication even if the exact numbers need confirmation.","headline":"Real SFDA gains from a frozen MedSAM, but the headline margin needs multi-seed runs and a principled validation split before I'd trust the exact numbers.","tokens_in":20399,"tokens_out":3344,"would_cite":true,"duration_ms":29558,"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":"A frozen MedSAM, prompted by bounding boxes found through dual feature guided search, outperforms existing source-free domain adaptation methods on medical segmentation benchmarks.","keywords":["source-free domain adaptation","medical image segmentation","Segment Anything Model","auto-prompting","bounding box prompt","pseudo-label refinement","feature aggregation","MedSAM"],"falsifier":"Take a held-out target slice with ground-truth labels, expand a ground-truth bounding box outward by 1, 2, and 3 pixels, and count changed MedSAM prediction pixels at each step; if the change is not consistently low while the box stays inside the organ and sharply higher once it crosses the boundary, the stable-interval stopping rule that DFG relies on is not reliable in that domain.","tokens_in":1645,"feed_emoji":"🩻","tokens_out":2363,"duration_ms":71532,"temperature":0.7,"pith_summary":"The paper tries to show that a frozen general-purpose segmentation model, MedSAM, can be turned into a reliable pseudo-labeler for source-free domain adaptation by automatically finding good bounding-box prompts. Existing SFDA methods produce defective boxes because their predictions are biased by the domain gap, and naively passing those boxes to MedSAM adds little or hurts accuracy. The proposed Dual Feature Guided (DFG) method first aggregates target features toward class prototypes, then expands a box prompt through two feature spaces—the target model's and MedSAM's—until MedSAM's output enters a stable interval, and finally keeps only the largest connected component of each predicted organ. If the method is right, a site with only a source model and unlabeled target scans can get segmentation accuracy close to fully supervised training, without sending data back to the source site.","feed_headline":"MedSAM auto-prompting adds ~10 Dice points to source-free segmentation","feed_subtitle":"A frozen medical SAM finds its own bounding-box prompts, lifting average Dice to 84.9–85.7 on cross-modality tasks.","key_machinery":"The carrying mechanism is the dual feature guided box prompt search (DBS). It starts from pixels whose target-model probability is within $p_{\\Delta}$ of the class maximum, then repeatedly adds spatially adjacent pixels whose features are close (cosine similarity above $\\tau_f = 0.99$) to expand a box; in parallel it expands from MedSAM's segmentation seed using a distance threshold proportional to the feature divergence within that seed. At each expansion, MedSAM is run on the current box and the number of changed pixels between consecutive predictions, $\\Delta_M$, is compared; the search stops when a long enough interval with low $\\Delta_M$ is found, meaning the box sits around an organ boundary. A connectivity-based post-processing (CP) then keeps only the largest connected component per class to remove enlarged false-positive regions.","core_discovery":"The paper's central claim is that MedSAM's segmentation output is locally stable when the box prompt is perturbed near the true organ boundary, and that this stability can be used as a stopping signal while growing a box from confidently predicted pixels. To use this signal, the paper observes two feature-distribution properties: target-model features of one class remain partially clustered after feature aggregation, while MedSAM features, though lacking class semantics, separate foreground from background well. DFG therefore grows the box using propagation over the target model's feature space (TBS) when features are clustered, and over MedSAM's feature space (MBS) when features are dispersed, stopping at the first long stable interval in the change of MedSAM prediction pixels. On four benchmarks the refined pseudo-labels and connectivity-based post-processing raise average Dice above all compared SFDA methods, with the prostate result (93.3%) nearly matching the fully supervised upper bound (94.4%).","pith_inferences":["If the box-prompt stability property transfers to point prompts or to faster promptable networks, the same search loop could be adapted for real-time or full 3D-volumetric settings.","Because MedSAM is frozen, the framework should improve automatically when a stronger promptable segmenter is released, without changing the adaptation algorithm.","The largest-connected-component rule embeds a single-object-per-class assumption; organs split into disconnected pieces would be penalized, so a multi-component variant is a natural testable extension.","The reported failure cases suggest the stable-interval detector can be fooled by accidental stability, so an ensemble across box sizes or a learned boundary detector could be tested against the same benchmarks."],"forward_implications":["On CHAOS to BTCV (MRI to CT), average Dice rises from 74.4% with ProtoContra to 84.9% with DFG, and average ASSD drops from 4.87 to 1.99.","On BTCV to CHAOS (CT to MRI), DFG reaches 85.2% average Dice and 1.00 average ASSD, surpassing all compared methods.","On CHAOS to CURVAS (MRI to CT), DFG reaches 85.7% average Dice and 2.27 average ASSD.","On NCI-ISBI to QUBIQ prostate adaptation, DFG attains 93.3% Dice and 2.18 ASSD, close to the fully supervised upper bound of 94.4% Dice.","Finetuning SAM on bad pseudo-labels (ProtoContra plus SAMed) drops average Dice to 67.3%, so keeping MedSAM frozen while searching for boxes is the more effective way to use the foundation model."],"supporting_citations":[{"why":"Establishes that a promptable segmenter generalizes across domains, which motivates using such a model in SFDA.","marker":"[17]"},{"why":"Provides MedSAM, the frozen medical segmenter whose box-prompt stability drives the search and whose encoder features drive MBS.","marker":"[21]"},{"why":"Supplies the prototype-anchored feature alignment loss used in the feature aggregation phase and the strongest SFDA baseline.","marker":"[16]"},{"why":"Provides the conditional transport cost and chain-rule formulation that underlies the feature aggregation loss.","marker":"[35]"},{"why":"Supplies the SAMed fine-tuning strategy used in the comparison that shows fine-tuning SAM on poor pseudo-labels degrades performance.","marker":"[18]"},{"why":"Serves as a pseudo-labeling SFDA baseline whose results DFG is compared against.","marker":"[10]"},{"why":"Serves as an uncertainty-aware pseudo-labeling SFDA baseline in the comparison tables.","marker":"[12]"},{"why":"Serves as an uncertainty-reduction SFDA baseline in the comparison tables.","marker":"[15]"},{"why":"Serves as a Fourier visual prompting SFDA baseline whose reported numbers are compared directly.","marker":"[14]"}],"fun_headline_variants":["SAM stability signal lifts SFDA segmentation by ~10 Dice","Grow box, stop at stability: SAM-guided SFDA","Dual-feature SAM tuning gets SFDA near supervised Dice"],"cache_read_input_tokens":22528,"weakest_assumption_plain":"The entire pipeline depends on MedSAM's output remaining stable while the box prompt fluctuates near the true organ boundary and changing sharply when the box crosses it; if a target domain produces accidental stable intervals or unstable outputs near boundaries, the search stops at the wrong place and the pseudo-labels do not improve.","fun_headline_variants_meta":{"raw":{"variants":["SAM stability signal lifts SFDA segmentation by ~10 Dice","Grow box, stop at stability: SAM-guided SFDA","Dual-feature SAM tuning gets SFDA near supervised Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2482,"prompt_tokens":1034,"completion_tokens":1448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":1393}},"tokens_in":650,"tokens_out":1448,"duration_ms":12958,"temperature":1.0,"reasoning_tokens":1393,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:52:33.458020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out target slice with ground-truth labels, expand a ground-truth bounding box outward by 1, 2, and 3 pixels, and count changed MedSAM prediction pixels at each step; if the change is not consistently low while the box stays inside the organ and sharply higher once it crosses the boundary, the stable-interval stopping rule that DFG relies on is not reliable in that domain.","supporting_citations":[{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Establishes that a promptable segmenter generalizes across domains, which motivates using such a model in SFDA."},{"cited_title":"Segment anything in medical images,","cited_arxiv_id":null,"evidence_quote":"Provides MedSAM, the frozen medical segmenter whose box-prompt stability drives the search and whose encoder features drive MBS."},{"cited_title":"Source- free domain adaptation for medical image segmentation via prototype- anchored feature alignment and contrastive learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the prototype-anchored feature alignment loss used in the feature aggregation phase and the strongest SFDA baseline."},{"cited_title":"Exploiting chain rule and bayes’ theorem to compare probability distributions,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional transport cost and chain-rule formulation that underlies the feature aggregation loss."},{"cited_title":"Source-free domain adaptive fundus image segmentation with denoised pseudo-labeling,","cited_arxiv_id":null,"evidence_quote":"Serves as a pseudo-labeling SFDA baseline whose results DFG is compared against."},{"cited_title":"Upl-sfda: Uncertainty-aware pseudo label guided source-free domain adaptation for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Serves as an uncertainty-aware pseudo-labeling SFDA baseline in the comparison tables."},{"cited_title":"Source- free domain adaptation for image segmentation,","cited_arxiv_id":null,"evidence_quote":"Serves as an uncertainty-reduction SFDA baseline in the comparison tables."},{"cited_title":"Fvp: Fourier visual prompting for source-free unsupervised domain adaptation of medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Serves as a Fourier visual prompting SFDA baseline whose reported numbers are compared directly."}],"review_version":1}