{"id":"4cc7efce-0749-4bc1-b6ce-bf76fa0c556d","arxiv_id":"2412.08575","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SAM-Mix couples a GradCAM classifier to SAM through auto-generated box prompts, reaching 0.948 Dice on liver segmentation from 50 labeled CT slices.","lead":"A new training setup, SAM-Mix, uses a simple image classifier to automatically make prompts for a segmentation foundation model, cutting the amount of manually labeled data needed for liver segmentation in CT scans. On the LiTS benchmark, it reaches higher Dice scores than fully supervised U-Net while training on only 50 labeled slices for 10 epochs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline comparison confounded: pretrained SAM+LoRA at 10 epochs vs randomly initialized fully supervised baselines at 100 epochs; no SAM-LoRA baseline with ground-truth prompts, so the 5.1% gain is not attributable to the CAM guidance.","rationale":"The paper's strongest claim is that SAM-Mix with 50 labeled slices beats fully supervised baselines, and the reported table indeed shows a large margin. However, the comparison conflates three factors: (1) the proposed CAM-guided prompt generation, (2) the use of a pretrained SAM with LoRA, and (3) a much shorter training schedule. The SAM-PP baseline controls for SAM and LoRA to some degree, but not for the prompt source: it uses the same GradCAM boxes, so it does not answer whether a SAM-LoRA model with simple ground-truth box prompts or no prompts would already achieve the reported Dice. The CAM-reliability concern raised by the reader is plausible but less decisive, because SAM-Mix's end-to-end training could compensate for imperfect boxes. The missing ablation is the most load-bearing because it directly tests whether the central claimed superiority is due to the method's novelty or to the underlying foundation model. The public code makes this test straightforward. I therefore keep the reader's CONDITIONAL verdict unchanged, since the paper should address this ablation before the claim is accepted as stated.","tokens_in":7327,"tokens_out":9841,"duration_ms":111070,"concrete_test":"Run a controlled ablation with the released code: fine-tune SAM with the same LoRA settings (rank 8, same 10 epochs, same 50 labeled LiTS slices) using ground-truth bounding-box prompts instead of GradCAM-generated boxes, repeating 5 times and evaluating on the same test split. If this baseline reaches or exceeds Dice 0.948, the CAM-guided prompt generation is not the source of the reported superiority; if it falls substantially below, the CAM guidance is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Table 1 is not controlled for the pretrained foundation model. Under Implementation Details, the authors state that all fully supervised baselines (U-Net, nnU-Net, TransUNet) are randomly initialized and trained for 100 epochs, while the SAM-based models are initialized from SAM's pretrained weights and trained for only 10 epochs. The headline improvement of SAM-Mix-50 (Dice 0.948) over U-Net (0.897) could therefore come largely from SAM's pretrained representation plus LoRA adaptation, rather than from the proposed CAM-guided prompt generation. The SAM-PP rows do not close this gap: SAM-PP also uses SAM and GradCAM boxes, but it is a two-stage pipeline and does not isolate a simple SAM-LoRA fine-tune on the same 50 masks with ground-truth boxes or even full-image boxes. Without such an ablation, the claim that the CAM-based task guidance is responsible for the 5.1% improvement (or the 25.4% cross-domain gain) is unsupported. The reader's CAM-reliability concern is real, but it is secondary: even a noisy CAM can yield useful boxes; the more direct threat is that a much simpler baseline, fine-tuning SAM with LoRA on the same labeled slices, might match or beat SAM-Mix, which would make the proposed mechanism unnecessary for the reported result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAM-Mix, a multitask framework that combines an auxiliary ResNet-38 classifier with SAM for liver segmentation in abdominal CT. Slice-level labels are manufactured from the segmentation masks via C = max(y_i), GradCAM maps are thresholded to produce bounding boxes, and these boxes prompt a SAM image encoder adapted with LoRA. The method is trained end-to-end on 5, 50, or 100 labeled slices and evaluated on LiTS (in-domain) and TotalSegmentator (cross-domain). The authors report that SAM-Mix-50 outperforms fully supervised U-Net trained on the full LiTS training set, with a Dice improvement of about 5 percentage points in-domain and 25 percentage points cross-domain. The paper also compares against a two-stage SAM-PP pipeline and a limited-supervision MultiMix baseline.","tokens_in":7606,"tokens_out":5005,"duration_ms":51497,"significance":"If the reported gains are robust, the method would be a practically useful demonstration that a promptable foundation model can be steered by cheaply manufactured GradCAM boxes, substantially reducing annotation burden. The paper uses public datasets, reports five-seed averages, and provides code, which are strengths. The core idea is simple and the experimental design is internally consistent across Table 1 and Table 2. However, the central comparison is confounded by the use of pretrained SAM weights with LoRA against randomly initialized baselines, and the missing controlled baseline (SAM-LoRA with ground-truth prompts) prevents attributing the improvement to the proposed task-guidance mechanism. The headline arithmetic also contains errors. These issues are load-bearing for the paper's central claim and need to be resolved before the contribution can be assessed.","major_comments":[{"comment":"The central comparison is confounded by the use of a pretrained foundation model. The Implementation Details state that all fully supervised baselines are randomly initialized and trained for 100 epochs, whereas the SAM-based models are initialized from SAM's pretrained weights and trained for only 10 epochs. The reported gain of SAM-Mix-50 over U-Net (Dice 0.948 vs 0.897) could therefore be driven by the pretrained representation plus LoRA adaptation rather than by the proposed GradCAM-guided prompts. A control experiment that fine-tunes SAM with LoRA on the same 50 labeled slices using ground-truth bounding-box prompts (or full-image prompts) is missing; without it, the central claim that the task guidance is responsible for the improvement is unsupported.","section":"§3.2, Table 1"},{"comment":"The headline numerical claims are internally inconsistent and arithmetically incorrect. The abstract states a 5.1% Dice improvement, Section 3.3 states 5.9%, and the difference in Table 1 is 0.948 - 0.897 = 0.051 (5.1 percentage points). Additionally, 50/11437 is 0.44%, not the stated 0.04%. These numbers must be corrected and made consistent across the abstract, the results section, and the tables.","section":"§3.3, abstract"},{"comment":"No direct evidence is provided that the automatically generated bounding boxes are reliable. The auxiliary classifier is trained on labels C = max(y_i) that are derived from the same ground-truth masks, and the GradCAM threshold tau = 0.5 * max(g) is a free parameter (omega) that is never varied. Reporting prompt-box quality (e.g., box IoU or recall of the liver ROI) and an ablation over omega would substantiate the mechanism claim that classification-guided attention, rather than the SAM backbone itself, drives the result. This evidence is load-bearing because the entire novelty rests on the CAM-to-prompt connection.","section":"§2, Eq. (5)-(6)"}],"minor_comments":[{"comment":"The method is described as semi-supervised throughout, but no unlabeled data are used; all training examples carry segmentation masks (5, 50, or 100 slices). The terminology should be changed to limited-supervision or few-shot supervised to avoid confusion with standard semi-supervised learning.","section":"§1, §3.2"},{"comment":"The focal loss hyperparameters alpha and gamma in Eq. (3) are not specified in Implementation Details; please provide the chosen values.","section":"§2, Eq. (3)"},{"comment":"The preprocessing step 'extract only the middle 30% slices' should state the resulting number of slices per scan and clarify whether the 5/50/100 labeled slices are drawn from this filtered set.","section":"§3.2, Data"},{"comment":"The supervision column is incomplete for several rows: nnU-Net, TransU-Net, and the SAM-PP/SAM-Mix variants do not have explicit 'Full' or 'Limited' entries, which makes the table harder to interpret. Please make the supervision status explicit for every row.","section":"Table 1, Table 2"},{"comment":"References [9] and [13] are both from the same group and describe the same method in different venues; please distinguish them at first citation and ensure the author lists are complete and consistent.","section":"§5, References"}],"recommendation":"major_revision","confidential_remarks":"The confounding of pretrained SAM weights with the proposed CAM guidance is the central issue, but it is addressable by adding a controlled SAM-LoRA baseline with ground-truth or full-image prompts. The arithmetic errors and semi-supervised mislabeling should be corrected. I recommend major revision rather than rejection because the core idea is plausible and the missing experiment is well-defined within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the paper's central comparison is confounded. The fully supervised baselines (U-Net, nnU-Net, TransUNet) are randomly initialized and trained for 100 epochs, while SAM-Mix and SAM-PP start from SAM's pretrained weights and train for only 10 epochs. So the 5.1% and 25.4% Dice improvements likely come mostly from SAM's pretrained representation plus LoRA, not from the GradCAM-driven prompt guidance. There is no ablation with a SAM-LoRA baseline using ground-truth boxes or even full-image boxes on the same 50 slices. Without that, the claim that the task guidance is responsible for the gains is unsupported. This is a bigger problem than the reader's CAM-noise concern; even a noisy CAM that produces a rough box may be enough, but a plain SAM-LoRA might match or beat the whole method.\n\nWhat is actually new: the paper inverts the usual S2C direction and uses GradCAM from an auxiliary classifier to generate bounding box prompts for a LoRA-adapted SAM, trained end-to-end. That specific combination is not in the cited literature. The writing is clear, the code is public, and the results are internally consistent across two datasets with five runs each. Those are real merits.\n\nSoft spots beyond the main confound: the paper calls the method semi-supervised, but no unlabeled data are used; the auxiliary labels are derived from the segmentation masks themselves (C = max(y_i)), so the guidance is not independent of the masks; the reported 0.04% labeled-data figure is arithmetically wrong (50/11437 is 0.44%, about ten times larger); and evaluation is restricted to the middle 30% of slices, which limits the claim of generality. The SAM-PP baseline is a two-stage pipeline and does not isolate the effect of end-to-end training versus pretrained initialization.\n\nWho this is for: someone working on annotation-efficient medical segmentation with foundation models will find the idea plausible and the code useful, but should not take the headline numbers at face value. The paper deserves a serious referee because the combination is novel, the experiments are reproducible, and the flaws are fixable. I would send it to review, but require a major revision: add the missing SAM-LoRA ablation with controlled prompts, correct the arithmetic, and either drop the semi-supervised label or actually use unlabeled data.\n\nVerdict: conditional accept after major revision, with the CAM-guidance contribution yet to be demonstrated.","headline":"The proposed CAM-guided SAM pipeline is new and the numbers are consistent, but the headline gains are not attributable to the proposed mechanism because all SAM-based models start from pretrained weights while the fully supervised baselines start from scratch.","tokens_in":8147,"tokens_out":1379,"would_cite":false,"duration_ms":16051,"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":"SAM-Mix beats fully supervised U-Net on liver segmentation using only 50 labeled CT slices and 10 training epochs.","keywords":["medical image segmentation","Segment Anything Model","GradCAM","multitask learning","semi-supervised learning","liver CT segmentation","prompt generation","LoRA"],"falsifier":"On the LiTS test set, compute the intersection-over-union and center-of-mass error between each GradCAM-derived bounding box and the liver ground-truth mask, then correlate them with per-slice Dice. If a substantial fraction of boxes fail to enclose the liver while SAM-Mix still predicts high Dice, the claimed guidance mechanism is contradicted; if boxes are accurate but Dice is low, the segmentation branch would be the limiting factor.","tokens_in":7104,"feed_emoji":"🩻","tokens_out":8615,"duration_ms":76649,"temperature":0.7,"pith_summary":"SAM-Mix claims that a medical image segmenter built on the Segment Anything Model can be trained with almost no dense annotation: an auxiliary classifier trained only on slice-level labels (liver present or absent) produces GradCAM heatmaps, which are thresholded into bounding boxes that automatically prompt SAM. With just 50 labeled slices—0.04% of the LiTS training labels—and 10 epochs, this end-to-end framework reaches a Dice score of 0.948 on in-domain liver segmentation, beating a fully supervised U-Net trained on the whole dataset, and a 25.4 percentage-point higher Dice than U-Net on a cross-domain test set. The paper's point is that classification-guided attention can substitute for dense manual annotation in prompting a foundation segmenter, and that this guidance improves both accuracy and generalization.","feed_headline":"SAM-Mix beats fully supervised U-Net with 50 labeled slices","feed_subtitle":"Automatic GradCAM prompts let SAM segment liver accurately with just 0.04% of training labels.","key_machinery":"The load-bearing mechanism is the CAM-to-prompt connection: a ResNet-38 classifier, trained with focal loss on coarse labels $C=\\max(y_i)$ extracted from the same masks, produces GradCAMs $g=\\sum_i w_i^{(c)} F_{\\mathrm{last},i}$; thresholding at $\\tau = 0.5\\cdot\\max(g)$ yields a binary mask whose bounding box becomes SAM's prompt. The prompt encoder turns this box into sparse embeddings that steer SAM's mask decoder, while a LoRA adapter (rank 8) inserted into SAM's ViT attention makes fine-tuning parameter-efficient. The classifier and segmenter are optimized jointly (focal loss plus Dice loss), so the prompt generator is trained by the same objective as the segmentation.","core_discovery":"The central claim is that connecting an auxiliary classification task to the SAM segmentation branch through automatically generated GradCAM-based bounding-box prompts yields a semi-supervised multitask learner that outperforms fully supervised baselines. When the framework (SAM-Mix) is trained for 90% fewer epochs on only 50 labeled 2D slices, it improves in-domain liver Dice by 5.1 percentage points over the best fully supervised baseline (U-Net) and achieves a 25.4 percentage-point Dice improvement on the cross-domain TotalSegmentator test set, with a 62.26% lower Hausdorff distance. The same configuration with only 5 labeled slices still reaches Dice 0.919 in-domain and 0.807 cross-domain. The authors argue this demonstrates annotation-efficient task guidance: image-level labels, via CAMs, provide enough spatial information to prompt SAM effectively while the end-to-end training keeps the prompts aligned with the segmentation objective.","pith_inferences":["A direct test of the mechanism would measure how often the GradCAM-derived bounding box actually encloses the liver; the paper does not report this, so the causal role of the prompts remains an inference from the end-to-end Dice.","The same task guidance could generalize from bounding-box to point or mask prompts, or to multi-organ segmentation, where a single classifier would need to localize several structures at once—likely stressing the thresholded-CAM assumption.","Because the coarse labels $C=\\max(y_i)$ are derived from the same masks used for segmentation, the method still needs a small seed set of full masks; a further extension is to test whether labels from a different modality or noisy annotations still yield the benefit."],"forward_implications":["Medical image segmentation can be performed accurately with a tiny fraction of the manual annotations, since coarse image-level labels suffice to make SAM segment effectively.","The same framework transfers across scanners and protocols: the 25.4 percentage-point cross-domain improvement implies that CAM-guided prompts improve generalization over full-supervision training.","Training cost drops dramatically: 10 epochs versus 100 for the fully supervised baselines, and LoRA limits the number of trainable parameters, making SAM fine-tuning feasible in low-resource settings.","The approach removes manual or semi-automated prompt generation for SAM, automating what was previously a human-in-the-loop bottleneck.","Even at 5 labeled slices, the method remains competitive (Dice 0.919 in-domain), suggesting that the guidance is robust to extreme label scarcity."],"supporting_citations":[{"why":"Defines the U-Net baseline that SAM-Mix must beat on the LiTS dataset.","marker":"[1]"},{"why":"Provides the Segment Anything Model, the promptable segmenter whose prompt encoder and mask decoder SAM-Mix fine-tunes.","marker":"[2]"},{"why":"Introduces the sparse-label multitask learning setup (classification plus segmentation) that SAM-Mix adapts into a CAM-guided SAM framework.","marker":"[9]"},{"why":"Establishes the inverse CAM-SAM direction (SAM improves CAMs), which motivates the CAM-to-SAM prompt generation used here.","marker":"[12]"},{"why":"Supplies the low-rank adaptation (LoRA) method that makes SAM's adaptation parameter-efficient.","marker":"[15]"},{"why":"Provides the LiTS benchmark used for in-domain training and evaluation.","marker":"[16]"},{"why":"Provides the TotalSegmentator dataset used for the cross-domain generalization test.","marker":"[17]"}],"fun_headline_variants":["50 slices beat full supervision in liver segmentation","SAM-Mix: 0.04% labels outperform fully supervised U-Net","GradCAM prompts slash labels for SAM liver segmentation","25% Dice leap cross-domain with only 50 labeled slices","Annotation-efficient SAM: 5 slices still hit 0.92 Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the auxiliary classifier's GradCAM, thresholded at half its maximum activation, produces a bounding box that reliably encloses the liver; if the CAM fires on non-liver cues or misses the organ, the prompts mislead SAM and the reported gains depend on that luck.","fun_headline_variants_meta":{"raw":{"variants":["50 slices beat full supervision in liver segmentation","SAM-Mix: 0.04% labels outperform fully supervised U-Net","GradCAM prompts slash labels for SAM liver segmentation","25% Dice leap cross-domain with only 50 labeled slices","Annotation-efficient SAM: 5 slices still hit 0.92 Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1757,"prompt_tokens":951,"completion_tokens":806,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":720}},"tokens_in":567,"tokens_out":806,"duration_ms":7992,"temperature":1.0,"reasoning_tokens":720,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:43:54.973379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the LiTS test set, compute the intersection-over-union and center-of-mass error between each GradCAM-derived bounding box and the liver ground-truth mask, then correlate them with per-slice Dice. If a substantial fraction of boxes fail to enclose the liver while SAM-Mix still predicts high Dice, the claimed guidance mechanism is contradicted; if boxes are accurate but Dice is low, the segmentation branch would be the limiting factor.","supporting_citations":[{"cited_title":"Annotation-Efficient Task Guidance for Medical Segment Anything","cited_arxiv_id":"2412.08575","evidence_quote":"Defines the U-Net baseline that SAM-Mix must beat on the LiTS dataset."},{"cited_title":", xN } is a set of abdominal CT slices and Y = {y1, y2,","cited_arxiv_id":null,"evidence_quote":"Provides the Segment Anything Model, the promptable segmenter whose prompt encoder and mask decoder SAM-Mix fine-tunes."},{"cited_title":"Self-supervised, semi-supervised, multi- context learning for the combined classification and seg- mentation of medical images,","cited_arxiv_id":null,"evidence_quote":"Establishes the inverse CAM-SAM direction (SAM improves CAMs), which motivates the CAM-to-SAM prompt generation used here."},{"cited_title":"From SAM to CAMs: Exploring segment anything model for weakly supervised semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the LiTS benchmark used for in-domain training and evaluation."},{"cited_title":"Generalized multi-task learning from substantially unlabeled multi-source med- ical image data,","cited_arxiv_id":null,"evidence_quote":"Provides the TotalSegmentator dataset used for the cross-domain generalization test."}],"review_version":1}