{"id":"ace501e3-f9b5-4d29-881c-68a6a7b19fd6","arxiv_id":"2507.19282","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SAM2-Aug, a fine-tuned SAM2 model that incorporates prior MRI and prior tumor annotations plus prompt augmentation, outperforms comparison models on liver, abdomen, and brain tumor segmentation for adaptive radiation therapy.","lead":"This paper adapts SAM2, a general image segmentation model, to automatically outline tumors in MRI scans used for adaptive radiation therapy, by feeding in the previous scan and its outline along with the current image. Tests on liver, abdominal, and brain tumor datasets suggest the method segments more accurately and generalizes better than several existing models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time bounding boxes appear to come from ground-truth contours, so the headline Dice gains over interactive SAM-based baselines may reflect an information leak rather than ART-workflow performance.","rationale":"The paper's central claim is a comparative performance claim: SAM2-Aug outperforms prompt-driven baselines and shows generalization. The comparison requires equivalent and clinically realistic prompts for all interactive methods. The reader flagged the prior-annotation availability assumption as the weakest assumption; I agree that is a real limitation, but it does not threaten the paper's internal comparison as much as the test-time bounding-box provenance does. If test-time boxes for MedSAM, SAM-Med2D, and SAM2 are derived from ground-truth contours, then their reported Dice scores are optimistic, and the reported gains of SAM2-Aug are not a fair assessment of its advantage. The paper does not state the source of test-time boxes, and since the code is not released, this is unverifiable. My reading is that this is the single most load-bearing concern because the headline claim is explicitly framed as outperforming other methods; a confounded test protocol would change the numerical conclusions. The paper's self-acknowledged limitations are honest, and the ablation study on input channels and prompt types is good evidence that prior images and prior masks help. The registration-only baseline winning over many deep models on One-Seq-Liver is an important stress-test the authors themselves surface, and is consistent with the prior-mask carrying most of the signal. Still, absent clarification of the evaluation protocol, the verdict should remain CONDITIONAL, with the condition being a fair, specified comparison protocol. I do not think the concern warrants REJECT, because the effect sizes are large and the internal ablations are internally consistent.","tokens_in":18470,"tokens_out":2107,"duration_ms":18569,"concrete_test":"Re-run the evaluation of MedSAM, SAM-Med2D, and SAM2 on the One-Seq-Liver, Mix-Seq-Abdomen, and Mix-Seq-Brain test sets using bounding boxes generated from the prior mask propagated by rigid registration (the same prior information already used by SAM2-Aug), rather than boxes derived from the ground-truth current mask. Also re-run the SAM2-Aug experiments with those prior-derived boxes under the same 0-5 pixel test-time perturbation. If the Dice margins over MedSAM and SAM2 shrink to within one standard deviation, the central comparative claim is not supported; if the margins persist, the fair-comparison concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that SAM2-Aug outperforms prompt-driven models such as MedSAM, SAM-Med2D, and SAM2 across three datasets. For that comparison to be valid, all interactive methods must receive equivalent, clinically realistic prompts. The paper states generally that SAM-based methods were compared under their original settings with bounding boxes, ten points, or boxes plus masks, but it never specifies how the test-time bounding boxes were obtained for MedSAM, SAM-Med2D, or SAM2. In ART, a current-fraction bounding box is not known a priori; it must be generated by user interaction, detection, or propagation from the prior. If the comparison boxes were derived from the ground-truth current mask, the reported Dice values for MedSAM and SAM2 are optimistic upper bounds, and SAM2-Aug's margin over them is not a fair reflection of its value in the clinical workflow. SAM2-Aug itself is also evaluated with test-time boxes that are described only as randomly expanded/contracted by 0-5 pixels; no mention is made of box degradation beyond 5 pixels at test time, despite the robustness experiment in Figure 8 using 1-10 pixels and reporting improved stability from augmentation. The most load-bearing issue is therefore not the availability of prior annotations (the reader's identified weakest assumption), but the unverified provenance of the prompt boxes at evaluation time. The paper's Appendices or implementation notes do not clarify this, and the GitHub link is listed as forthcoming, so the claimed head-to-head advantage over MedSAM and SAM2 depends on an unstated and possibly unfair test protocol. A concrete check: re-evaluate MedSAM and SAM2 with boxes that are derived from the same prior-based propagation or a detection output used for SAM2-Aug, and confirm the reported margins persist.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SAM2-Aug, a SAM2-based tumor segmentation framework for adaptive radiation therapy (ART). The key ideas are (1) augmenting the input with the prior MR image and prior annotation as additional channels, and (2) augmenting prompts during fine-tuning via random bounding-box expansion/contraction, mask erosion/dilation, and stochastic prompt dropout. The model is fine-tuned on a One-Seq-Liver dataset (19 training patients) and then tested on a held-out liver test set plus two unseen datasets: Mix-Seq-Abdomen and Mix-Seq-Brain. The authors report mean Dice scores of 0.86, 0.89, and 0.90 on the three datasets, claiming superiority over convolutional, transformer, and SAM-based baselines, with ablations over input channels, prompt types, and fine-tuned modules.","tokens_in":18729,"tokens_out":4828,"duration_ms":47408,"significance":"If the reported results are robust, the work addresses a practically important and timely problem: leveraging the natural availability of prior-fraction annotations in ART. The held-out evaluation on unseen MR sequences and tumor sites is a notable strength, as is the ablation structure that isolates the contributions of prior input, prompt augmentation, and fine-tuning strategy. The paper is clearly written and the experimental design is mostly thoughtful. However, the central comparison with interactive SAM-based baselines hinges on how test-time prompts were generated, and this is not specified precisely enough to rule out an information leak that would make the headline gains optimistic.","major_comments":[{"comment":"The provenance of test-time bounding boxes for the prompt-based baselines is not specified. The text states in §4.3 that MedSAM and SAM-Med2D were given bounding boxes, SAM-Med3D/Turbo ten points, and SAM2 and SAM2-Aug boxes plus masks, but it does not state how these boxes were obtained. In §3.2, the testing stage for SAM2-Aug says all bounding boxes were 'expanded or contracted by 0 to 5 pixels', which suggests they were derived from the ground-truth current mask. If the same is true for MedSAM, SAM-Med2D, and SAM2, the reported Dice values are optimistic upper bounds and the comparison does not reflect a realistic ART workflow, where the current-fraction box must come from user interaction or propagation from the prior. Please state the box generation protocol explicitly and re-evaluate all interactive methods with a realistic prompt source (e.g., prior-mask propagation with registration, or simulated user boxes), reporting results for that protocol.","section":"§4.3 and §3.2"},{"comment":"The 'SAM2 (without prompt augmentation)' baseline is not sufficiently defined. It is unclear whether this baseline receives the same three-channel input (current MR, prior MR, prior segmentation) and the same prior-mask prompt as SAM2-Aug, or only the current MR and a current box. If the baseline lacks the prior input, the 0.08 Dice gap in Table 3 conflates the proposed input augmentation with the prompt augmentation that the text credits for the improvement. Please specify the exact input and prompt configuration of every baseline, ideally adding this information as a column or footnote in the comparison tables.","section":"§5.1, Table 3"},{"comment":"No statistical significance testing is reported. The margins over the closest competitor are small on some datasets (e.g., brain Dice 0.90 vs 0.88 for SAM2 in Table 5; liver Dice 0.86 vs 0.83 for MedSAM and registration-only in Table 3), and test scans are clustered by patient. Please add paired (per-case) significance tests, confidence intervals, or effect-size measures to support the claim that SAM2-Aug outperforms all compared methods.","section":"§5, Tables 3–5"},{"comment":"The robustness of the method to imperfect or missing prior information is not tested. The test protocol deliberately keeps the prior annotation unaugmented and performs no prompt dropout, 'assuming physician-approved prior information will always be available'. This assumption may fail in ART when the prior mask is outdated after large anatomical change or when registration is inaccurate. Since the entire approach depends on the prior, please add experiments with, e.g., simulated registration errors, prior masks from different protocols, or randomly dropped priors at test time, and report how performance degrades.","section":"§3.2 and §6"}],"minor_comments":[{"comment":"There are two tables numbered Table 7: one for prompt-type ablations and one for fine-tuning ablations. Please renumber the second table.","section":"§5.2"},{"comment":"In the sentence about robustness and generalizability, the model is referred to as 'SAM-Aug2'; this should be 'SAM2-Aug'.","section":"§4.1"},{"comment":"The caption repeats 'The best values are shown in bold.' twice; one instance should be removed.","section":"Table 3 caption"},{"comment":"In Table 3, the baseline 'SAM2 (without prompt augmentation)' is listed without a footnote describing whether it uses the three-channel input from §3.3 and whether it receives the prior-mask prompt. A footnote would prevent misinterpretation.","section":"§5.1"},{"comment":"The first limitation discusses partial prompts and SAM2's memory mechanism, but the experiments reported in the paper always use full slices with complete prompts. Clarify whether this limitation refers to experiments not reported or to an anticipated issue, and if the former, report the corresponding results.","section":"§6, limitation (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the general idea is interesting. The decisive issue is the unverified prompt provenance in the interactive-baseline comparison: if the boxes were ground-truth-derived, the central claim of superiority over MedSAM/SAM2 is not established. This is fixable by clarifying and re-running the comparison with clinically realistic prompts, but until then the headline result should be treated as an upper bound. I would also encourage the authors to add significance testing, as several reported margins are small."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"If you only read one thing: the central idea here is good engineering, and the ablations are well structured, but the paper never says where the test-time bounding boxes come from. If they come from the ground-truth current contour, the reported margins over MedSAM and SAM2 are optimistic and need re-examination.\n\nWhat's new and useful: incorporating the prior MR and prior annotation as extra input channels to SAM2, plus prompt augmentation (bbox expansion/contraction, mask erosion/dilation, stochastic dropout) during fine-tuning. The combination is not in the literature for ART, and the three-dataset generalization test is a real strength. The ablations over input channels, prompt types, and fine-tuned modules are careful and tell a coherent story: prior context helps, prompt diversity helps, and the full configuration wins. The Dice gains over the no-augmentation SAM2 baseline (0.78 to 0.86 on liver, 0.84 to 0.89 on abdomen, 0.88 to 0.90 on brain) are modest but consistent.\n\nThe soft spot is the evaluation protocol for interactive baselines. Section 4.3 says they followed \"original settings\" with bounding boxes for MedSAM and SAM2, but never states how those boxes were obtained. The testing-stage description in Section 3.2 says boxes were \"expanded or contracted by 0 to 5 pixels,\" which reads like the starting point is the ground-truth box. If so, MedSAM and SAM2 are being given near-perfect prompts, which inflates SAM2-Aug's advantage. The robustness test in Figure 8 does degrade boxes up to 10 pixels, but only for SAM2-Aug, not for the comparison methods. The reader's worry about prior annotation availability is real but secondary; the paper at least states that assumption explicitly. Missing code/data is a practical limitation.\n\nWho this is for: anyone working on auto-segmentation in MR-guided ART, especially those using SAM-family models. The paper deserves a serious referee, but the evaluation section needs hard clarification. Ask for the exact prompt generation protocol for every SAM-based method, and re-run the comparison with non-oracle boxes (e.g., propagated from the prior via registration). If the margins persist, this is a solid contribution. If not, the remaining claim is just \"adding prior context helps,\" which is still of some value but much weaker.\n\nRecommendation: send it to peer review, but insist on the prompt-provenance question being answered before it is accepted.","headline":"A practical ART segmentation paper with a clever prior/prompt augmentation scheme, but the head-to-head comparison is undermined by unspecified test-time box provenance.","tokens_in":19317,"tokens_out":1996,"would_cite":false,"duration_ms":19876,"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":"SAM2-Aug claims that feeding SAM2 the previous fraction's MRI and tumor contour, plus randomized prompt perturbations, lifts tumor segmentation to Dice 0.86–0.90 across liver, abdomen, and brain without retraining.","keywords":["Segment Anything Model 2","adaptive radiation therapy","tumor segmentation","prior knowledge","prompt augmentation","MRI","MR-LINAC","generalization"],"falsifier":"A direct test would be to run SAM2-Aug on the same three datasets while systematically corrupting the prior: dropping the prior mask, adding synthetic misregistration shifts of 1 to 10 mm, or substituting a prior from a different MRI sequence. If Dice and HD95 degrade to the level of baseline SAM2 (0.78 liver, 0.84 abdomen, 0.88 brain) or below, the paper's causal claim that prior context drives the gain would be falsified.","tokens_in":18274,"feed_emoji":"🩺","tokens_out":7543,"duration_ms":66062,"temperature":0.7,"pith_summary":"This paper argues that SAM2, a generalist image-and-video segmentation model, can become accurate for tumor delineation in adaptive radiation therapy if it is given the right context and the right prompts. The proposed method, SAM2-Aug, feeds the model a three-channel input made of the current MRI, the previous fraction's MRI, and the previous fraction's tumor annotation, so the network can use the tumor's known location and shape as a scaffold. It also randomizes prompts during fine-tuning, expanding or contracting bounding boxes, eroding or dilating prior masks, and sometimes dropping them, so the model learns to tolerate imperfect user input. Fine-tuned on only 69 liver MRI scans, SAM2-Aug reports Dice scores of 0.86 for liver, 0.89 for abdominal, and 0.90 for brain targets, beating convolutional, transformer, and other SAM-based baselines without retraining on the abdominal and brain sets. The practical payoff is that re-contouring tumors at every treatment fraction, a time-consuming bottleneck in adaptive radiation therapy, could be automated using information the clinic already has.","feed_headline":"Prior MR scans lift tumor segmentation to 0.90 Dice","feed_subtitle":"A model fine-tuned only on liver MRIs also segments abdominal and brain tumors without retraining.","key_machinery":"The load-bearing object is the three-channel prior-augmented input combined with prompt augmentation. In the adaptive radiation therapy workflow, fraction n's scan is rigidly registered to fraction (n-1)'s scan, so the previous tumor mask can be overlaid; the current MRI, prior MRI, and prior mask become a three-channel tensor fed to SAM2. On the prompt side, the current-frame bounding box is randomly expanded or contracted by 0 to 5 pixels and the prior mask is randomly eroded or dilated, with a stochastic selection that can drop either prompt; at test time only the box is perturbed and the prior mask is always used. The adaptation mechanism is joint fine-tuning of SAM2's image encoder, prompt encoder, and mask decoder on a small liver dataset, which then transfers to other sites.","core_discovery":"The paper's central claim is that prior knowledge from a patient's own earlier treatment fraction is sufficient to lift a promptable foundation model above specialist segmentation networks for tumor targets. Concretely, concatenating the current MR image, the prior MR image, and the prior tumor annotation into a three-channel input, combined with prompt augmentation (random bounding-box perturbation and prior-mask erosion/dilation with stochastic dropout), and fine-tuning SAM2's image encoder, prompt encoder, and mask decoder, yields mean Dice scores of 0.86 on liver ITV, 0.89 on abdominal ITV, and 0.90 on brain CTV. The generalization claim is that the same liver-trained weights transfer to different MRI sequences and tumor sites without retraining. The ablation evidence supports the mechanism: the full three-channel input adds 0.11 Dice over current-MR-only input, augmented box-plus-mask prompts add 0.1 over mask-only prompts, and joint fine-tuning of the three modules beats fine-tuning any subset or all modules.","pith_inferences":["If the prior annotation is unreliable, missing, from a different sequence, or misregistered, the contextual channels could actively mislead the model; a natural extension is measuring how Dice degrades as registration error or prior quality worsens.","The method's success suggests a broader design principle: for any sequential imaging task, a foundation model can be turned into an online segmenter by conditioning on the most recently approved label, which could extend to CBCT-guided radiation therapy and other longitudinal imaging.","A testable extension would be to make the prior-mask prompt itself a learnable parameter or to fuse the three channels with attention rather than concatenation, which might close the remaining gap to registration-only performance on near-stationary anatomy."],"forward_implications":["If the claim holds, clinics can reuse the contour already approved for the previous fraction as the main guide, so each new fraction's tumor contour could be generated automatically rather than drawn from scratch.","The same liver-trained weights should transfer to other abdominal and brain sites, meaning one small fine-tuning dataset could support a cross-site auto-segmentation tool.","Prompt augmentation should make the pipeline robust to realistic sloppiness in user-drawn boxes, since the model is trained on boxes that are wrong by up to five pixels and on eroded or dilated masks.","The ablation results suggest prior-mask prompts are especially valuable for small tumors, so the approach should help in the low-contrast, small-target cases where manual contouring is most variable.","Because the inputs are simple three-channel concatenations, the method can ride on existing SAM2 infrastructure without architectural changes, making adoption comparatively cheap."],"supporting_citations":[{"why":"Supplies the SAM2 backbone whose image encoder, prompt encoder, and mask decoder SAM2-Aug fine-tunes.","marker":"[50]"},{"why":"Defines the prompt-based segmentation paradigm and the box, mask, and point prompts that the augmentation strategy perturbs.","marker":"[28]"},{"why":"Provides MedSAM, the strongest prompt-driven medical baseline that SAM2-Aug must beat on Dice and boundary metrics.","marker":"[22]"},{"why":"Provides SAM-Med2D, a medical SAM variant used as a baseline and an example of fine-tuning on medical images.","marker":"[29]"},{"why":"Provides SAM-Med3D and SAM-Med3D-Turbo, volumetric SAM baselines whose poor transfer contrasts with SAM2-Aug.","marker":"[30]"},{"why":"Shows the video-as-volume SAM2 medical adaptation that motivated treating MR fractions as sequential context.","marker":"[31]"},{"why":"Demonstrates that prior-fraction contours improve online auto-segmentation, the premise SAM2-Aug extends to SAM2.","marker":"[41]"},{"why":"Inspires the prompt augmentation idea with point-prompt augmentation for SAM.","marker":"[42]"},{"why":"Supports the claim that prompt quality and structure strongly affect SAM2 segmentation, motivating the prompt augmentation design.","marker":"[52]"},{"why":"Describes the MR-linac adaptive radiation therapy workflow that supplies the prior-fraction data used by the method.","marker":"[53]"}],"fun_headline_variants":["Prior MR scans add 0.11 Dice to tumor segmentation","Liver-trained SAM2-Aug segments brain and abdomen tumors","Prior images lift SAM2 tumor segmentation to 0.90 Dice","SAM2-Aug: prior knowledge boosts tumor Dice to 0.90"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a physician-approved tumor contour from the previous treatment fraction or from simulation is always available, can be rigidly registered to the current scan, and still resembles the current tumor; if that prior is missing, heavily deformed, or from a different imaging protocol, the three-channel input and mask prompt can mislead the model rather than help it.","fun_headline_variants_meta":{"raw":{"variants":["Prior MR scans add 0.11 Dice to tumor segmentation","Liver-trained SAM2-Aug segments brain and abdomen tumors","Prior images lift SAM2 tumor segmentation to 0.90 Dice","SAM2-Aug: prior knowledge boosts tumor Dice to 0.90"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3256,"prompt_tokens":1021,"completion_tokens":2235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":2163}},"tokens_in":637,"tokens_out":2235,"duration_ms":15790,"temperature":1.0,"reasoning_tokens":2163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:55:36.245437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to run SAM2-Aug on the same three datasets while systematically corrupting the prior: dropping the prior mask, adding synthetic misregistration shifts of 1 to 10 mm, or substituting a prior from a different MRI sequence. If Dice and HD95 degrade to the level of baseline SAM2 (0.78 liver, 0.84 abdomen, 0.88 brain) or below, the paper's causal claim that prior context drives the gain would be falsified.","supporting_citations":[],"review_version":2}