{"id":"2c1200d4-9717-4c6b-b6fa-48bf4066aafa","arxiv_id":"2501.16246","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An unsupervised pipeline that uses CLIP to label slices, CAM to localize tumors, SAM to generate pseudo-masks, and self-training with similarity filtering reaches 85.6% Dice on BraTS2020 whole-tumor segmentation.","lead":"This paper combines CLIP and SAM to segment brain tumors on MRI without any human-annotated training labels, reaching an average Dice score of 85.6% on BraTS2020. The method is a cascade of CLIP-derived labels, class activation maps, SAM prompts, and self-training, and it reports results close to fully supervised segmentation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never measures CLIP's slice-level pseudo-label accuracy, so the first and most load-bearing stage of the pipeline is unvalidated.","rationale":"The paper's strongest claim is that an unsupervised cascade of CLIP, CAM, and SAM approaches fully supervised performance on BraTS2020. The ablation study supports the contributions of AMDA and S3F relative to the pipeline's own baseline, and the comparison to SaLIP and other methods is plausible. However, every component downstream of the image-level labels inherits errors from D1. The authors do not report CLIP's zero-shot accuracy on FLAIR slices, even though ground truth is available for the training split; such a report would not violate the unsupervised protocol because it is an analysis, not a training signal. Without it, the most fragile premise is unverified. We agree with the reader's weakest_assumption. The reader's CONDITIONAL verdict is appropriate, and our stress-test does not change it; if anything, it sharpens the required condition: the authors should report the CLIP label confusion matrix and clarify the AMDA threshold definition. No evidence of dishonesty or internal inconsistency was found; the manuscript is a plausible but under-validated method.","tokens_in":7486,"tokens_out":7549,"duration_ms":71354,"concrete_test":"Run the CLIP labeling step from Sec 2.1.1 on the BraTS2020 training slices and compare yi = argmax(p_i) against a slice-level binary ground truth derived from the provided volumetric tumor masks (any slice containing at least one tumor voxel is positive). Report the confusion matrix, especially false-negative rate and overall accuracy. If accuracy is below ~80% or false-negative rate above ~20%, the first stage is unreliable and the final 85.60% DSC cannot be attributed to the described unsupervised pipeline without further analysis. Also, re-read the AMDA definition in Sec 2.1.2: for α=20, state whether Mi covers the top 20% or top 80% of pixels; if it covers 80%, the hyperparameter curve in Fig. 3(a) should be re-plotted to confirm the model is not simply being trained on mostly-masked images.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CLISC's entire label-generation chain starts with CLIP pseudo-labels in Sec 2.1.1: D1 = {Xi, yi} with yi = argmax(p_i) from two text prompts applied to FLAIR MRI, a modality far from CLIP's natural-image training data. No experiment in the paper reports CLIP's agreement with ground-truth tumor presence on any slice. If the false-negative rate on tumor-bearing slices is substantial, then the classifier, the Layer-CAM, the AMDA refinement, and every SAM prompt inherit missing or spurious positive labels, and the 3D U-Net is trained on pseudo-labels that omit entire tumor slices. The reported final DSC of 85.60% depends on this first stage being reliable, yet the only evidence is the final performance itself, which is circular. A secondary ambiguity is in Sec 2.1.2: Mi = Q0 > Tα with Tα the α-th percentile; for α=20 this masks 80% of pixels, which appears inconsistent with the claim in Fig. 3(a) that a larger α masks all tumor areas. The central claim is plausible and the ablation study is internally consistent, but the missing validation of the CLIP-derived labels is a genuine load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLISC, a three-stage unsupervised pipeline for whole-tumor FLAIR segmentation on BraTS2020. CLIP provides slice-level pseudo-labels that train a ResNet-50 classifier; Layer-CAM with adaptive masking (AMDA) produces ROIs; ROIs generate box/point prompts for SAM; SAM masks are stacked into 3D pseudo-labels; a 3D U-Net is trained with these pseudo-labels and refined with a similarity-based filter (S3F). On a 20% held-out test set, the method reports average DSC 85.60% and HD95 6.72 mm, exceeding five unsupervised baselines by more than 10 DSC points and approaching fully supervised training (87.97% DSC, 7.49 mm HD95).","tokens_in":111,"tokens_out":4714,"duration_ms":101987,"significance":"If valid, the result is significant: it demonstrates that a cascade of frozen foundation models can close much of the gap between unsupervised and fully supervised whole-tumor segmentation, and it outperforms SAM prompted by ground truth. The evaluation is well structured: a held-out test set, five strong baselines, paired t-tests, a size-stratified breakdown, and a three-part ablation. The method is also simple in that CLIP and SAM are used off-the-shelf. However, the load-bearing first stage—CLIP slice-level pseudo-labels—is never validated against any ground truth, and the hyperparameter selection protocol is not explicit; these issues must be addressed before the central claim is fully supported.","major_comments":[{"comment":"The paper never reports the accuracy of the CLIP-derived slice-level pseudo-labels yi = argmax(p_i) against ground-truth tumor presence, even though D1 is the foundation of every subsequent stage (classifier, CAM, AMDA, SAM prompts, pseudo-labels D3/D4). If CLIP's false-negative rate on tumor slices is substantial, the entire pipeline inherits missing or spurious labels, and the final 85.60% DSC would not be evidence that the pipeline is robust. Please report slice-level classification accuracy, a confusion matrix, and a sensitivity analysis (e.g., replacing CLIP labels with oracle labels or perturbing a fraction of labels) on a held-out split.","section":"Sec. 2.1.1, Eq. (1)"},{"comment":"The definition of Mi = Q0_i > Tα, where Tα is 'the α-th percentile,' is ambiguous and appears inconsistent with the stated goal of masking the most discriminative region. For α=20, Mi selects the top 80% of pixels (if Q0 values are sorted ascending), so masking Mi removes most of the image, not just the discriminative region. Please clarify whether Tα should be the (100−α)-th percentile or whether Mi is defined differently, and report the mean fraction of pixels masked for each α in Fig. 3(a).","section":"Sec. 2.1.2"},{"comment":"The choice α=β=20 is not tied to a stated validation procedure. If the test set was used to select these hyperparameters, the reported DSC is optimistically biased. Please state explicitly how Fig. 3 was generated (validation vs test), and report performance on the 10% validation split for all configurations, including the final model.","section":"Sec. 3.1 and Fig. 3"},{"comment":"The S3F filter uses the similarity between the network's own prediction Si and SAM's output S_hat_i, where the SAM prompts are generated from Si. This internal-consistency criterion can be high even when both are confidently wrong, so F(Si,S_hat_i) does not directly measure pseudo-label quality. The ablation shows a gain, but the mechanism remains unclear; please analyze how filtering changes the agreement between retained pseudo-labels and ground truth (if measurable) and discuss the risk of self-confirmation.","section":"Sec. 2.3"}],"minor_comments":[{"comment":"The phrase 'four corner points (background) of the bounding box' is ambiguous; do these points lie inside or outside the box? Please specify their coordinates.","section":"Sec. 2.2"},{"comment":"Define exactly what '5 points' and 'box' mean for SAM prompts in the ablation; a reader cannot reproduce the prompt settings from the current text.","section":"Table 2"},{"comment":"Label the vertical axes; currently the reader cannot tell whether DSC or HD95 is plotted.","section":"Fig. 3"},{"comment":"The manuscript does not mention whether the validation split is used for early stopping or model selection for the 3D U-Net; please clarify.","section":"Sec. 3.1"},{"comment":"No code or pre-trained models are released; given the many non-standard steps (AMDA, S3F), releasing code would substantially aid reproducibility.","section":"General"},{"comment":"There are a few typographical issues in Figure 1 (e.g., duplicated '𝓓𝓓' symbols), which should be cleaned up.","section":"Fig. 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong candidate if the authors can validate the CLIP pseudo-label stage and clarify hyperparameter selection. I would not reject on the current evidence, but the missing validation is a genuine load-bearing gap. Also note that the authors should check whether their percentile definition is correct, as it may indicate a typo in the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on CLISC (arXiv:2501.16246). If the numbers are reproducible, this is an impressive result: 85.6 Dice for whole-tumor FLAIR segmentation on BraTS2020 with zero manual labels, within 2.4 points of full supervision and above SAM prompted with ground-truth boxes. The pipeline is a coherent three-stage cascade — CLIP slice-level labels to train a classifier, LayerCAM with adaptive masking (AMDA) to refine ROIs, SAM prompts from those ROIs to get pseudo-masks, then a 3D U-Net self-trained with a similarity filter (S3F) against SAM. What's genuinely new is the combination plus the two refinements; the individual pieces are known but the integration is sensible and the ablations are clean. Each stage adds a clear, monotonic improvement: AMDA lifts CAM DSC from 51.1 to 60.2, adding box+point prompts gets 74.2, training the U-Net on SAM pseudo-labels jumps to 82.3, and S3F filtering brings it to 85.6. That's good evidence the gains come from the whole system, not one lucky component.\n\nThe main soft spot is the first stage: CLIP's slice-level tumor/non-tumor labels are never validated against ground truth. The paper uses two English text prompts applied to FLAIR MRI — a modality far from CLIP's training data — and the entire downstream chain inherits any error there. No agreement statistics are reported, so a reader can't tell whether the method is robust to label noise or just happens to work on this dataset. That's a real gap, though the final held-out score is an indirect sanity check; if CLIP labels were badly wrong, you'd expect the pipeline to collapse. Still, for transferability, this needs reporting.\n\nSecond, no code is released, and the hyperparameters (alpha=20, beta=20) aren't explicitly tied to the validation split. The ablation plots in Fig. 3 suggest they did tune on validation, but it should be stated. One note: the stress-test raised that alpha=20 masks 80% of pixels, which seems inconsistent with \"larger alpha masks all tumor areas.\" That concern doesn't hold up on reading — larger alpha gives a higher threshold and thus a smaller, more concentrated mask, which can indeed erase the whole tumor. So the paper's explanation is fine.\n\nBottom line: this deserves a serious referee. It's a strong empirical claim with well-designed experiments and a sensible method. I'd recommend sending it to peer review, with the expectation that the authors release code and report CLIP label accuracy on a sample. If those hold, it's a solid contribution for anyone working on annotation-efficient medical segmentation.","headline":"CLISC reports a strikingly strong unsupervised whole-tumor segmentation result, but the unvalidated CLIP labeling stage and missing code are real caveats.","tokens_in":8257,"tokens_out":4433,"would_cite":true,"duration_ms":39603,"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":"An unsupervised cascade that chains CLIP, CAM, and SAM segments brain tumors at 85.60% DSC on BraTS2020 without any manual annotations.","keywords":["unsupervised segmentation","brain tumor","foundation models","CLIP","SAM","class activation mapping","self-training","BraTS2020"],"falsifier":"Run the CLIP label step on the BraTS2020 training slices and compare the argmax labels against ground-truth tumor presence. If the slice-level accuracy is close to chance, or if replacing CLIP labels with ground-truth labels does not materially change the final segmentation DSC, the central claim would be weakened; the paper's own ablations do not include this measurement.","tokens_in":7310,"feed_emoji":"🧠","tokens_out":4699,"duration_ms":38784,"temperature":0.7,"pith_summary":"This paper tries to establish that an unsupervised pipeline can get close to supervised performance on brain tumor segmentation by chaining two foundation models. CLIP supplies image-level tumor/non-tumor labels for every FLAIR slice, a classifier trained on those labels produces enhanced class activation maps, and those maps generate box and point prompts for SAM to produce segmentation pseudo-labels. A 3D U-Net is then trained on the pseudo-labels, with a self-training step that keeps only labels on which SAM and the network agree. On BraTS2020, the pipeline reports 85.60% average whole-tumor DSC and 6.72 mm HD95, about 10 points above five unsupervised baselines and within 2.4 points of fully supervised training. The evaluation is limited to whole-tumor FLAIR segmentation; tumor substructures and other organs are left to future work.","feed_headline":"No-label brain tumor segmentation hits 85.6% DSC","feed_subtitle":"A CLIP-CAM-SAM cascade beats five unsupervised baselines and nears fully supervised accuracy on BraTS2020.","key_machinery":"The load-bearing mechanism is a four-stage pseudo-label cascade. First, CLIP's cosine-similarity scores classify each FLAIR slice as tumor or non-tumor, and those labels train a ResNet classifier. Second, Layer-CAM produces a raw activation map, and an adaptive masking step hides the most discriminative region and retrains the classifier so the next CAM covers more of the tumor. Third, the enhanced CAM defines a bounding box and a center point plus four corner points that prompt SAM, yielding per-slice masks that are stacked into 3D pseudo-labels. Fourth, a 3D U-Net trains on those pseudo-labels, then a self-training round re-prompts SAM with the network's own prediction and keeps only volumes whose pseudo-label agrees with SAM above a percentile threshold.","core_discovery":"CLISC claims that slice-level labels extracted from CLIP with two English text prompts are sufficient to bootstrap a cascade that produces high-quality whole-tumor segmentations. After filtering CLIP labels through a ResNet classifier, Layer-CAM localization with adaptive masking, SAM prompting with a bounding box plus five points, and a similarity-based self-training round, the 3D segmentation network reaches 85.60% DSC on BraTS2020, surpassing zero-shot SAM prompted by ground-truth boxes (80.59% DSC) and approaching fully supervised training (87.97% DSC). The paper treats this as evidence that adapting foundation models through an intermediate CAM stage, rather than using them directly for inference, closes much of the gap between unsupervised and supervised medical image segmentation.","pith_inferences":["The same cascade could be transferred to other organs or modalities where CLIP gives weak slice-level signal, provided the CAM stage can localize the salient region; the paper's future-work sentence points this way.","A testable extension is to replace CLIP with a medical-domain vision-language model and measure how much of the gain comes from better slice labels rather than from CAM or SAM.","The similarity-based filtering between SAM output and student prediction is a generic denoising rule that could be applied to any pseudo-label pipeline, independent of CLIP.","If CLIP slice-label accuracy is low, the entire chain inherits the error; reporting that accuracy against ground-truth labels would make the method's reliability transparent."],"forward_implications":["On BraTS2020 FLAIR whole-tumor segmentation, the unsupervised CLISC pipeline reaches 85.60% average DSC and 6.72 mm HD95, more than 10 DSC points above five state-of-the-art unsupervised baselines.","CLISC beats SAM in zero-shot inference even when SAM is given prompts derived from ground truth: 85.60% versus 80.59% DSC.","The gap to fully supervised training is small: 85.60% versus 87.97% DSC, and CLISC's HD95 is better (6.72 mm versus 7.49 mm).","Ablations show each stage contributes: adaptive masking raises CAM DSC from 51.10% to 60.22%, combined box-and-point SAM prompts reach 74.21% DSC, and self-training with similarity filtering raises the final DSC from 82.29% to 85.60%.","The method also handles small and tiny tumors better than the unsupervised baselines, and for tiny tumors it exceeds fully supervised training."],"supporting_citations":[{"why":"CLIP supplies the slice-level image labels that start the pseudo-label cascade.","marker":"[12]"},{"why":"SAM generates segmentation pseudo-labels from the CAM-derived prompts.","marker":"[10]"},{"why":"Layer-CAM produces the raw activation maps that are enhanced by adaptive masking.","marker":"[14]"},{"why":"Constrained CAM-based unsupervised anomaly segmentation is one of the baselines and the source of the CAM-consistency idea.","marker":"[8]"},{"why":"Sim2Real tumor simulation is a baseline the method must beat.","marker":"[9]"},{"why":"SaLIP is the closest zero-shot SAM-plus-CLIP baseline.","marker":"[11]"},{"why":"Implicit-field anomaly detection is one of the five unsupervised baselines.","marker":"[13]"},{"why":"ResNet50 is the classification network trained on CLIP-derived labels.","marker":"[17]"},{"why":"3D U-Net is the segmentation network trained on SAM-derived pseudo-labels.","marker":"[18]"}],"fun_headline_variants":["Unsupervised brain tumor segmentation hits 85.6% DSC","CLIP+CAM+SAM: no labels, 85.6% DSC on brain tumors","Foundation models segment brain tumors without labels at 85.6% DSC","CLISC: Bootstrap SAM with CLIP cues for 85.6% unsupervised DSC","No annotations needed: CLIP-guided SAM scores 85.6% DSC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes CLIP's zero-shot slice-level tumor/non-tumor labels on FLAIR MRI are accurate enough to train a reliable classifier; the paper never reports this label accuracy against ground truth, and every later stage inherits its errors.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised brain tumor segmentation hits 85.6% DSC","CLIP+CAM+SAM: no labels, 85.6% DSC on brain tumors","Foundation models segment brain tumors without labels at 85.6% DSC","CLISC: Bootstrap SAM with CLIP cues for 85.6% unsupervised DSC","No annotations needed: CLIP-guided SAM scores 85.6% DSC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000874,"raw_usage":{"total_tokens":3788,"prompt_tokens":956,"completion_tokens":2832,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2728}},"tokens_in":572,"tokens_out":2832,"duration_ms":16581,"temperature":1.0,"reasoning_tokens":2728,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:35:57.611220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the CLIP label step on the BraTS2020 training slices and compare the argmax labels against ground-truth tumor presence. If the slice-level accuracy is close to chance, or if replacing CLIP labels with ground-truth labels does not materially change the final segmentation DSC, the central claim would be weakened; the paper's own ablations do not include this measurement.","supporting_citations":[{"cited_title":"Bayesian skip-autoencoders for un- supervised hyperintense anomaly detection in high res- olution brain MRI,","cited_arxiv_id":null,"evidence_quote":"CLIP supplies the slice-level image labels that start the pseudo-label cascade."},{"cited_title":"Brain tumor segmentation based on a hybrid clustering technique,","cited_arxiv_id":null,"evidence_quote":"SAM generates segmentation pseudo-labels from the CAM-derived prompts."},{"cited_title":"Constrained unsupervised anomaly segmentation,","cited_arxiv_id":null,"evidence_quote":"Layer-CAM produces the raw activation maps that are enhanced by adaptive masking."},{"cited_title":"Deep learning models and traditional automated techniques for brain tumor segmentation in MRI: a review,","cited_arxiv_id":null,"evidence_quote":"Constrained CAM-based unsupervised anomaly segmentation is one of the baselines and the source of the CAM-consistency idea."},{"cited_title":"Self-semantic contour adaptation for cross modality brain tumor segmentation,","cited_arxiv_id":null,"evidence_quote":"Sim2Real tumor simulation is a baseline the method must beat."},{"cited_title":"Level set method with automatic selective local statistics for brain tumor segmentation in MR images,","cited_arxiv_id":null,"evidence_quote":"SaLIP is the closest zero-shot SAM-plus-CLIP baseline."},{"cited_title":"Unsupervised region-based anomaly detection in brain MRI with ad- versarial image inpainting,","cited_arxiv_id":null,"evidence_quote":"Implicit-field anomaly detection is one of the five unsupervised baselines."},{"cited_title":"Test-time adap- tation with SaLIP: a cascade of SAM and CLIP for zero- shot medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"ResNet50 is the classification network trained on CLIP-derived labels."},{"cited_title":"Learning transferable visual models from natural lan- guage supervision,","cited_arxiv_id":null,"evidence_quote":"3D U-Net is the segmentation network trained on SAM-derived pseudo-labels."}],"review_version":1}