{"id":"4762ed9c-a540-4021-a227-0604fc427dd6","arxiv_id":"2607.18014","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Freezing SAM2's encoder, fine-tuning its decoder/memory, and adding TSDF-trained global volume tokens yields 0.78 mean Dice on a new 34-dataset MRI benchmark, up from 0.58 zero-shot.","lead":"The paper assembles a 10,392-volume, 34-dataset MRI segmentation benchmark and adapts SAM2 by freezing its encoder, fine-tuning only the decoder and memory modules, and adding whole-volume context tokens trained with a boundary-geometry (TSDF) objective. It reports a mean Dice rise from 0.58 (zero-shot SAM2) to 0.78, with the lowest variance across datasets and no drop on 8 held-out datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Comparison to prior methods may be uncontrolled: SAMRI-3D is fine-tuned on the benchmark's seen datasets while SAMed-2, Medical-SAM2, and SAM-Med3D are likely evaluated zero-shot, so the 'surpassing' claim may reflect training-set access rather than architectural superiority.","rationale":"The reader's weakest assumption was the untested cut-slice prompting sensitivity. That is a valid practical concern, but I see a more fundamental issue with the paper's central comparative claim. The abstract's headline is that SAMRI-3D 'surpasses' existing SAM-based medical models. For that claim to be supported, those models must be evaluated under comparable conditions. The paper's protocol section (§4.1) describes training/validation only on the seen datasets, but it is ambiguous whether this applies to the comparison methods. The main text explicitly calls SAM2-Small 'zero-shot' but not the other three methods, suggesting they were evaluated as-is with their public checkpoints. If so, the comparison is not controlled: SAMRI-3D has been fine-tuned on 26 datasets that include the same anatomical domains and sequences as the test sets, while the alternatives have not. The statistically significant margins then reflect training-set access, not necessarily architectural superiority. This does not undermine the reported absolute Dice values or the benchmark's utility, but it does invalidate the 'surpassing' wording unless the comparisons are clarified or re-run. The cut-slice concern is real but secondary: it affects real-world deployment, while the comparison issue affects whether the paper's central claim is scientifically established at all. The paper is otherwise transparent (e.g., candid about GVT's per-dataset insignificance, p=0.22), which is why I do not recommend rejection—only a conditional revision to specify and, if necessary, correct the evaluation protocol.","tokens_in":75713,"tokens_out":8299,"duration_ms":91502,"concrete_test":"Check the released evaluation scripts/configs to determine whether SAMed-2, Medical-SAM2, and SAM-Med3D were fine-tuned on the 26 seen training folds. If they were not, fine-tune at least SAMed-2 (the closest prior) on the same 26 seen train splits with the same prompt protocol (GT box on cut slice) and evaluate on the same test folds. If fine-tuned SAMed-2 reaches or exceeds SAMRI-3D's 0.78 mean Dice, the 'surpassing' claim is not supported; if it remains below, the claim stands. Also require a statement of the evaluation protocol for all comparisons in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SAMRI-3D 'surpasses' SAMed-2, Medical-SAM2, and SAM-Med3D (Abstract; §4.3, Table 2) depends on a like-for-like evaluation protocol. The paper states in §4.1 that models are trained and validated only on the 26 seen datasets' train/validation subsets, but it never states whether the three comparison methods received this training or were evaluated zero-shot with their public checkpoints. Only SAM2-Small is explicitly labeled 'zero-shot' in §4.3. If, as is standard, the prior methods were not fine-tuned on the SAMRI-3D seen datasets, then SAMRI-3D had access to the evaluation distribution (same anatomy, sequences, and modalities) while the alternatives did not. The 'strong statistical significance' in Table S12 then only demonstrates that training on the target distribution improves performance—an unsurprising result—rather than that SAMRI-3D's architecture or GVT is superior. Because the abstract's headline is a comparative superiority claim, this missing control is load-bearing. The paper must either disclose that the comparisons are zero-shot and qualify the claim, or report fine-tuned comparisons.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SAMRI-3D, a benchmark and method for 3D MRI segmentation built on SAM2. The benchmark comprises 10,392 volumes from 34 datasets (26 seen, 8 unseen; 27 public, 7 in-house) across 12 anatomical domains. The method freezes SAM2.1's Hiera image encoder and prompt encoder, fine-tunes the mask decoder and memory modules, and adds Global Volume Tokens (GVT): 16 persistent tokens that summarize the whole volume and are appended to SAM2's streaming memory bank at every propagation step. The GVT tokens are trained with an auxiliary TSDF-reconstruction objective (boundary-weighted Smooth-L1 loss, Eq. 4), and the TSDF decoder is discarded at inference. The paper reports mean Dice 0.78 ± 0.14 across the 34 datasets, versus 0.58 for zero-shot SAM2-Small and 0.76 for a fine-tuned no-GVT baseline, and compares favorably with SAMed-2 (0.69), Medical-SAM2 (0.49), and SAM-Med3D (0.37), with no generalization gap on the 8 held-out datasets (0.79 unseen vs. 0.78 seen). The GVT-vs-baseline gain is significant at per-sample and per-label levels but not at the per-dataset level (p = 0.22), a result the paper reports honestly.","tokens_in":76013,"tokens_out":12888,"duration_ms":122020,"significance":"If the claims hold, the paper is a substantial contribution to medical foundation-model adaptation: it provides the largest MRI-only SAM2 evaluation to date with explicit seen/unseen splits; a parameter-efficient adaptation (frozen encoder, roughly 150K added trainable parameters) that approximately halves the error of zero-shot SAM2; and a falsifiable mechanistic claim — that TSDF-supervised global tokens help most on low-contrast sequences — which the per-sequence analysis (Fig. 3) directly tests. The reporting is unusually careful: per-dataset medians with quartiles, Wilcoxon tests at three aggregation levels, disclosure that the GVT advantage is not per-dataset significant, and a plan to release benchmark, code, and models. These strengths are contingent on the comparison protocol being made explicit (Major Comment 1).","major_comments":[{"comment":"§4.3/Table 2: the claim that SAMRI-3D 'surpasses' SAMed-2, Medical-SAM2, and SAM-Med3D 'with strong statistical significance' is not verifiable because the evaluation protocol for these three external methods is never stated. §4.1 describes training only on the 26 seen datasets' train/validation subsets, and §4.3 labels only SAM2-Small as 'zero-shot.' If the three external methods were evaluated zero-shot with public checkpoints, the comparison is confounded by training-set access: SAMRI-3D was trained on the same datasets (hence the same anatomy, sequences, and modalities) on which it is evaluated, so the small p-values in Table S12 would show the benefit of target-distribution fine-tuning rather than architectural superiority. This is load-bearing for the headline comparative claim. Please disclose each method's checkpoint, prompt format, preprocessing, and fine-tuning status, and eith","section":"§4.3, Table 2; Abstract; §4.1"},{"comment":"The method is anchored on the 'cut slice' (the slice with the largest target foreground), and segmentation propagates outward from it; at inference, boxes and/or points are supplied by the user or an upstream detector. The paper never evaluates sensitivity to prompt-slice choice. A user prompting on a slice with small or diffuse foreground may anchor the entire propagation chain poorly, and the reported 0.78 Dice is established only under the favorable cut-slice protocol. Please add an analysis that varies the anchor (e.g., slices with decreasing foreground area, or cut slices displaced by ±k slices from the optimal one) and reports the resulting Dice/HD, to demonstrate that accuracy does not critically depend on selecting the optimal slice.","section":"§3.1, Fig. 1A"},{"comment":"The statement that 'two further SAM2-based adaptations produced near-zero Dice on this benchmark and are omitted' is insufficient as reported. Near-zero Dice across 34 datasets is a strong signal of a protocol mismatch (different prompting requirements, evaluation convention, or preprocessing) rather than a genuine performance measurement. Name the methods, specify the protocol used (checkpoint, prompt, preprocessing, datasets on which they were run), and report the numbers; otherwise the omission invites suspicion that the comparison set was filtered.","section":"§4.3"}],"minor_comments":[{"comment":"The abstract and §3.3 say the TSDF objective adds 'zero added cost' and that 'GVT adds no test-time cost.' Strictly, only the TSDF decoder is discarded. The GVT encoder (≈120K parameters, one extra forward pass over the pooled slice embeddings) and the appended K=16 tokens in the memory bank are active at inference, so the overhead is small but nonzero. Please rephrase and quantify the actual inference-time cost.","section":"§3.3 / Supp. B / Table S2"},{"comment":"The per-sample unit is not defined. With an 80/10/10 split of 10,392 volumes the test set is ≈1,039 volumes, yet n≈3,376 for the 'sample-level' tests; clarify whether the unit is volume-label pairs. Also note that samples from the same volume are not independent, so the per-dataset level (where the GVT gain is p=0.22) is the safe summary of the GVT effect.","section":"Table S12"},{"comment":"The same model is named 'GVT-baseline' in Table 4 and 'SAMRI-3D-2dloss' in Tables S4 and S6–S8; unify the naming to avoid confusion.","section":"Table 4 vs. Tables S4, S6–S8"},{"comment":"'Global V olume Tokens' contains a spurious space ('V olume') in the abstract and in the contributions list.","section":"Abstract; §1"},{"comment":"The caption does not define the 'baseline' being compared (SAMRI-3D-2dloss) nor the pooling used to compute per-sequence medians; please add these definitions.","section":"Fig. 3 caption"},{"comment":"The Discussion explicitly defers comparison with fully-supervised methods such as nnU-Net. For a paper claiming best accuracy on its own benchmark, a nnU-Net reference is important for contextualizing the absolute Dice values; if not addressed in the revision, the omission should be stated as a limitation in the conclusions.","section":"Discussion, §5"}],"recommendation":"major_revision","confidential_remarks":"The internal comparisons (fine-tuned baseline and GVT ablations) are controlled and the reporting is honest; I would not reject. The main fix is disclosing the evaluation protocol for the three external methods (Major Comment 1), which I judge to be revision-scale rather than fatal. Also verify at acceptance: the release plan should state which of the 7 in-house datasets can actually be shared, since the benchmark's reproducibility depends on the public subset."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The benchmark is the real contribution here: 10,392 volumes across 34 MRI datasets with seen/unseen splits is the most substantial MRI-only evaluation anyone has built for SAM2, and it is reported carefully — medians, quartiles, Wilcoxon at three levels, per-sequence and per-domain breakdowns. If the code and data actually ship, this becomes a useful testbed for the field. The method side is also reasonable: freezing the Hiera encoder, fine-tuning only the decoder/memory, and adding a small set of persistent volume tokens with a training-only TSDF objective is a clean recipe, and the GVT idea is genuinely not in the cited SAM2-adaptation literature.\n\nThe soft spot is the central comparison. The paper says SAM2-Small is the zero-shot reference, but it never states whether SAMed-2, Medical-SAM2, and SAM-Med3D were fine-tuned on the 26 seen datasets or evaluated with their public checkpoints. Given standard practice, they were almost certainly zero-shot. If so, the abstract's 'surpassing recent SAM-based medical models' is mostly a statement about having trained on the evaluation distribution, not about architectural superiority. The \"strong statistical significance\" in Table S12 then proves the unsurprising fact that target-domain fine-tuning beats zero-shot. This needs to be disclosed and qualified, or the baselines need to be fine-tuned/reported with the same training access.\n\nOther issues are smaller. The two unnamed near-zero baselines should be named; if they fail, say so explicitly. The unseen split is anatomically overlapping (multiple knee datasets from OAI/iMorphics), so the no-drop claim on held-out data is weaker than it sounds. The pipeline depends on prompting the cut slice — the slice with the largest foreground — and no sensitivity analysis to prompt-slice choice is offered. And the TSDF objective was chosen based on the same test set used for headline metrics; the reported difference between TSDF and binary reconstruction is not statistically significant at the dataset level, so the claim that TSDF is the \"most effective\" is not robust. The paper is honestly written about several of these, which helps, but they still need addressing.\n\nWho should read it: anyone working on SAM2 adaptation for medical volumes, and anyone building MRI evaluation benchmarks. It deserves a serious referee — the benchmark is worth that — but a revision should be required to clarify or fix the comparison protocol, add prompt-slice sensitivity, and validate the GVT choice on a properly separated validation set.","headline":"A genuinely large MRI benchmark and a sensible method paper, but the headline 'surpassing' claim rests on comparisons that are probably not like-for-like and should be fixed before this is citable as a win.","tokens_in":76572,"tokens_out":2883,"would_cite":true,"duration_ms":36183,"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-encoder SAM2 adaptation plus TSDF-supervised global volume tokens raises 3D MRI segmentation from 0.58 to 0.78 mean Dice on 34 datasets, with no drop on 8 held-out datasets.","keywords":["MRI segmentation","SAM2 adaptation","3D medical image segmentation","global memory tokens","signed distance field","foundation models","parameter-efficient fine-tuning","zero-shot generalization"],"falsifier":"Run SAMRI-3D on a fixed set of volumes but prompt on slices chosen progressively further from the maximum-foreground slice (for example, near the top or bottom of the target, or on a slice where the target is fragmented), holding everything else identical; if mean Dice drops steeply as the anchor moves off the cut slice, the cut-slice assumption is load-bearing. A second check: retrain GVT with a high-resolution binary mask or a different geometric prior in place of the TSDF target — the paper's own ablation already shows binary reconstruction adds little, so a positive result there would weak","tokens_in":75550,"feed_emoji":"🧠","tokens_out":9112,"duration_ms":79614,"temperature":0.7,"pith_summary":"SAMRI-3D aims to show that SAM2, a video-segmentation foundation model, can be turned into a strong general 3D MRI segmenter without retraining its image encoder. The paper reports that freezing SAM2's image encoder and fine-tuning only its decoder and memory modules lifts mean Dice from 0.58 to 0.76 on the largest MRI-only benchmark assembled to date — 10,392 volumes across 34 datasets — and that adding Global Volume Tokens trained against a truncated signed distance field (TSDF) target pushes this to 0.78 with the lowest variance across datasets. The core claim is that whole-volume boundary geometry, not appearance, is the information missing from SAM2's local streaming memory, and that a training-only TSDF reconstruction injects exactly that information. If true, this matters because it offers a parameter-efficient route to promptable MRI segmentation that generalizes to datasets never seen in training, at zero added inference cost.","feed_headline":"Frozen SAM2 hits 0.78 Dice on 3D MRI, no unseen-data drop","feed_subtitle":"Decoder-only fine-tuning plus boundary-geometry memory tokens add whole-volume context at zero inference cost.","key_machinery":"Global Volume Tokens (GVT): K=16 persistent memory tokens built by compressing the pre-computed embeddings of all slices into one whole-volume summary through Perceiver-style cross-attention, then appended to SAM2's streaming memory bank so that every slice, however far from the prompted anchor slice, sees global context. The load-bearing twist is their training objective: a training-only decoder reconstructs a Truncated Signed Distance Field of the target (boundary-weighted Smooth-L1, truncation radius 32 pixels), forcing the tokens to encode boundary geometry rather than redundant binary occupancy. The auxiliary decoder is discarded at inference, so the gain costs nothing at test time. Sli","core_discovery":"On the paper's own terms, the central claim is this: an MRI-specific adaptation of SAM2 that freezes the Hiera image encoder, fine-tunes only the lightweight decoder and memory modules, and augments the memory bank with 16 persistent Global Volume Tokens supervised by a Truncated Signed Distance Field reconstruction achieves the best mean accuracy (0.78 Dice) and lowest variance across 34 MRI datasets, surpassing recent SAM-based medical models, and — uniquely among strong baselines — shows no generalization gap on 8 unseen datasets (0.79 unseen vs 0.78 seen). The TSDF objective, discarded at inference, is what lets the tokens survive the memory attention's softmax competition and carry boun","pith_inferences":["Not tested in the paper: sensitivity to the cut-slice anchor. The entire propagation chain starts from a prompt on the slice where the target is largest, so a natural follow-up is to sweep prompts across slices and measure how much of the 0.78 depends on that choice.","Beyond MRI: the recipe — persistent global tokens plus a geometric self-supervision target — is modality-agnostic and could be applied to other long-sequence inputs (CT, ultrasound, endoscopic video) where a streaming memory otherwise evicts distant context.","Reading the statistics: the TSDF gain over the strong baseline is statistically significant per-sample and per-label but not per-dataset (p=0.22); the paper itself positions the model as the lowest-variance, best-mean option, and that is the framing a reader should weigh — small, consistent gains rather than large jumps on any single dataset.","A testable prediction follows from the per-sequence analysis: on high-contrast sequences with already-sharp boundaries (DESS, T1CE) the GVT gain should shrink toward zero — which is what the paper reports, and what an independent re-run could verify."],"forward_implications":["Fine-tuning only the decoder and memory modules, with the image encoder frozen, is sufficient for large gains on MRI — the paper shows SAM2 backbone size has almost no effect on zero-shot MRI accuracy, so parameter-efficient adaptation is the scalable path.","Whole-volume context injected through the memory bank and supervised by boundary geometry (TSDF) beats both no-token and binary-reconstruction variants on mean accuracy and variance, at under 1.1% parameter overhead and zero test-time cost.","The model transfers to eight datasets never seen during training with no performance drop (0.79 unseen vs 0.78 seen), which is the strongest zero-shot evidence in the comparison.","The improvement is concentrated exactly where it is needed: low-contrast sequences such as T1 and FLASH with diffuse boundaries, matching the paper's invisible-boundary motivation.","Because the TSDF decoder is discarded at inference, GVT adds no runtime cost to deployment — the reported gains are essentially free at test time."],"fun_headline_variants":["SAMRI-3D: 0.78 Dice on MRI, zero unseen-drop","Frozen encoder + token trick lifts MRI segmentation to 0.78","MRI segmentation: SAM2 fine-tune beats SAM-Med3D by 0.41","Whole-volume tokens erase MRI generalization gap","SAMRI-3D tops 34 MRI datasets, no unseen-data penalty"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported accuracy assumes someone (or an upstream detector) supplies a reliable prompt on the cut slice — the slice with the largest target foreground — because every other slice's prediction propagates outward from that anchor; the paper does not test how accuracy degrades when the prompt lands on a slice where the target is small, fragmented, or diffuse.","fun_headline_variants_meta":{"raw":{"variants":["SAMRI-3D: 0.78 Dice on MRI, zero unseen-drop","Frozen encoder + token trick lifts MRI segmentation to 0.78","MRI segmentation: SAM2 fine-tune beats SAM-Med3D by 0.41","Whole-volume tokens erase MRI generalization gap","SAMRI-3D tops 34 MRI datasets, no unseen-data penalty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1372,"prompt_tokens":874,"completion_tokens":498,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":401}},"tokens_in":618,"tokens_out":498,"duration_ms":4584,"temperature":1.0,"reasoning_tokens":401,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:21:19.039595+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SAMRI-3D on a fixed set of volumes but prompt on slices chosen progressively further from the maximum-foreground slice (for example, near the top or bottom of the target, or on a slice where the target is fragmented), holding everything else identical; if mean Dice drops steeply as the anchor moves off the cut slice, the cut-slice assumption is load-bearing. A second check: retrain GVT with a high-resolution binary mask or a different geometric prior in place of the TSDF target — the paper's own ablation already shows binary reconstruction adds little, so a positive result there would weak","supporting_citations":[],"review_version":1}