{"id":"ea1684c2-e5c3-41f2-986e-d23dd1033d4c","arxiv_id":"2501.14718","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A heat map from a cancer-grade classifier used as a SAM prompt gives small gland-segmentation gains on GlaS, with no error bars and only old baselines compared.","lead":"This paper combines a benign/malignant classifier with the Segment Anything Model, using a Grad-CAM++ heat map as a prompt to improve gland segmentation in colorectal cancer images. It reports small gains over fine-tuned SAM on the public GlaS dataset, but the improvements are modest and the SOTA claim is weakened by comparisons to much older baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed prompt benefit is not established: Table 1 shows prompted SAM-H F1 on Test A decreases (0.932 to 0.929), the prompted model adds structural changes beyond the prompt, and heatmap alignment from a 224px ViT to 400px SAM input is unspecified.","rationale":"I agree with the reader's conditional verdict and partially with the identified weakest assumption. The reader focused on heatmap alignment and the missing prompt-removal ablation; I would sharpen this: the Table 1 comparison itself contains a counterexample (Test A F1 for SAM-H) and the architecture differs from the baseline by more than the prompt. Nonetheless, the paper is honest in reporting the numbers, describes its modifications clearly, and the direction of most comparisons is positive, so this is a robustness/evidence concern, not an indication of a wrong method. The correct response is to require the missing control and confidence intervals before accepting the attribution claim. Since the reader already issued CONDITIONAL on essentially this basis, my verdict is unchanged. Credit should be given for using a standard public benchmark and reporting both the baseline and prompted results in full, which makes the inconsistency visible.","tokens_in":5519,"tokens_out":4676,"duration_ms":53669,"concrete_test":"Run the Prompted SAM-H pipeline with the adapter input set to (i) the Grad-CAM++ heatmap, (ii) a constant map equal to the mean heatmap value, and (iii) a randomly permuted heatmap with the same pixel statistics, keeping all weights and the training schedule identical, across at least five seeds; report paired bootstrap 95% confidence intervals for F1, Object Dice, and Hausdorff on Test A and B. If (ii) or (iii) performs within the CI of (i), the prompt information is not the driver. As a separate alignment check, generate the heatmap at 224x224 and upsample to 400x400 using bilinear versus nearest-neighbor interpolation, and also shift the prompt by 50 pixels; if segmentation is invariant to these changes, the claimed spatial conditioning is not occurring.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 is the only evidence for the central claim, and it does not show a consistent prompt gain. Prompted SAM-H F1 on Test A is 0.929 vs. 0.932 for SAM-H, a decrease; SAM-L Object Dice on Test B is tied at 0.846; other gains are small differences in the third decimal. No repeats, seeds, or confidence intervals are reported. More importantly, the prompted model is not the fine-tuned SAM plus a prompt: it adds a trainable prompt adapter, a second prompt encoder and mask decoder for contours, a stepwise training schedule, and contour-based post-processing (Sections 2.3-2.7). Any of these changes could explain the differences. No ablation feeds the adapter a blank, constant, or shuffled heatmap, so the grade information is not isolated as the cause. The mechanism is also underspecified: the classifier is deit-base-patch16-224, while the SAM input is 400x400, and Section 2.2 only states that the heatmap 'maintains the same size as the original input image.' No resampling, positional-embedding interpolation, or alignment procedure is described. If the heatmap is generated at 400x400 without modifying DeiT's positional embeddings, the 224x224 pretraining assumption is violated; if it is upsampled from 224x224, the prompt may be too coarse to condition gland boundaries. The claimed causal role of cancer grade is therefore not isolatable from the current experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-branch model for colorectal gland segmentation evaluated on the GlaS dataset. A ViT classification branch is fine-tuned for benign/malignant gland classification, and a Grad-CAM++ heatmap derived from this classifier is used as a prompt for a modified Segment Anything Model (SAM). The segmentation branch contains a shared image encoder, separate prompt encoders and mask decoders for gland and contour prediction, a trainable prompt adapter, a stepwise training schedule, and contour-based post-processing. The paper claims that the prompted SAM variants improve over fine-tuned SAMs and that the best variant, prompted SAM-H, achieves SOTA results on GlaS. The evidence for these claims is presented in Table 1 (prompted vs. fine-tuned SAM) and Table 2 (comparison with GlaS contest baselines).","tokens_in":5836,"tokens_out":4660,"duration_ms":41410,"significance":"The underlying idea is interesting and potentially useful: using a grade classifier's saliency map as a conditioning signal for SAM is a plausible mechanism for injecting clinical prior knowledge into gland segmentation in histopathology. If the prompt benefit were robustly established, the paper would make a useful contribution to the medical-image segmentation literature. The paper also provides a simple prompt-adapter design and a dual-branch SAM modification that are easy to understand. However, as presented, the central claim is not supported by the experimental evidence: the improvements in Table 1 are small, inconsistent, and reported without uncertainty estimates; no ablation isolates the grade-prompt component; the spatial alignment of the heatmap is underspecified; and the SOTA comparison is limited to dated contest baselines. The significance of the work is therefore conditional on a substantially strengthened experimental evaluation.","major_comments":[{"comment":"The central quantitative claim is not established because the improvements are inconsistent and no uncertainty is reported. Prompted SAM-H F1 on Test A decreases from 0.932 to 0.929, SAM-L Object Dice on Test B is tied at 0.846, and most other differences are at the third decimal place. With test sets of only 60 and 20 images, and no repeated runs, seeds, confidence intervals, or significance tests, these differences cannot be distinguished from noise. The authors should report multiple runs with means and standard deviations and apply paired statistical tests (e.g., Wilcoxon signed-rank or bootstrap) on object-level metrics.","section":"§3.3, Table 1"},{"comment":"The experimental design does not isolate the grade prompt as the cause of any observed improvement. The prompted model is not simply fine-tuned SAM plus a prompt; it also adds a trainable prompt adapter (§2.3), a second prompt encoder and mask decoder for contour prediction, a stepwise training schedule (§2.5), and contour-based post-processing that removes overlapping gland and contour regions (§2.7). Any of these components could explain the differences in Table 1. No ablation feeds a blank, constant, shuffled, or adversarial heatmap into the same architecture, so the causal role of the cancer-grade information is not tested. Such ablations are necessary before claiming that cancer grade, rather than added capacity or post-processing, drives the reported gains.","section":"§2.3–§2.7"},{"comment":"The alignment between the grade heatmap and SAM's input space is underspecified. The classifier is deit-base-patch16-224, while the SAM input is stated to be 400×400 pixels, and the text only says that the heatmap \"maintains the same size as the original input image.\" The paper must state how the Grad-CAM++ output is resized or generated at 400×400, including how DeiT's positional embeddings are handled. Without this, the method cannot be reproduced, and it is unclear whether the prompt is generated under assumptions compatible with the pretrained ViT or is a coarse upsampling that would not carry boundary-level gland information.","section":"§2.2 and §3.2"},{"comment":"The claim of SOTA results is not supported by the comparison in Table 2. The table lists only GlaS contest entries from 2017, and the paper's own conclusion cites Hi-gMISnet with an F1 score of 0.932 on the overall test set, yet this method is not included in the comparison. The authors should either qualify the claim to \"improves over the fine-tuned SAM baseline\" or extend the comparison to modern methods evaluated under the same protocol and split.","section":"Tables 1–2 and §4"}],"minor_comments":[{"comment":"The adapter description is difficult to follow: the text says the heatmap is concatenated with the original image and the four-channel feature is reduced to one channel, then \"the heat map is added with the output from the second convolution.\" Please clarify the exact residual structure and the channel dimensions at each step.","section":"§2.3"},{"comment":"Hyperparameters such as learning rates, batch size, number of epochs, validation split, and the exact patch-extraction and inference-overlap strategy are not reported, which limits reproducibility.","section":"§3.1"},{"comment":"There is a typo in the conclusion: \"purposed method\" should be \"proposed method.\"","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a short conference-style paper, and its central claim currently rests on a single table without error bars or ablations. For a journal submission, the evaluation needs to be substantially expanded: multiple runs, statistical tests, prompt-control ablations, and a more modern and complete comparison. The idea is worth pursuing, and the issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a plausible incremental idea with an under-evidenced central claim. The grade-prompt is a neat twist, but Table 1 doesn't prove it's the prompt doing the work.\n\nWhat's new and good: using a Grad-CAM++ heat map from a benign/malignant ViT as a prompt for a fine-tuned SAM, plus a prompt adapter and a contour branch to separate touching glands. That's a sensible extension of CGS-Net and the SAM-adaptation literature, and the paper compares against fine-tuned SAM at three scales, which is more than many short papers do. The writing is clear and GlaS is the right benchmark.\n\nThe soft spots are mostly about evidence. Table 1 shows prompted SAM-H F1 on Test A drops from 0.932 to 0.929, SAM-L Object Dice on Test B ties at 0.846, and the rest of the gains are in the third decimal. No error bars, no repeats, no significance tests, and the test sets are 60 and 20 images. More importantly, the prompted model adds a trainable adapter, a second prompt encoder/decoder, stepwise training, and contour post-processing. Without an ablation feeding a blank or shuffled heatmap, you can't attribute the difference to grade information. That's the missing experiment.\n\nI also share the alignment worry. The classifier is deit-base-patch16-224, so its Grad-CAM++ is naturally 224x224, while SAM sees 400x400. The paper says the heatmap 'maintains the same size as the original input image' but doesn't say how. If it's just upsampling, the prompt may be too coarse for gland boundaries. A sentence or two and a sensitivity check would settle it.\n\nThe 'SOTA' phrasing is too strong, since the paper's own conclusion cites Hi-gMISnet with an overall F1 of 0.932, which is at or above the prompted SAM-H's 0.929 on Test A. Comparing only to 2017 GlaS baselines in Table 2 isn't the full picture.\n\nNone of this is fatal. The idea is coherent, the writing is honest, and the weaknesses are fixable. It deserves a serious referee, one who will insist on the ablation and alignment details. I wouldn't cite it yet, but I'd bring it to our reading group as an example of prompt-based SAM adaptation in pathology. My recommendation: send to peer review, expect major revision.","headline":"Grade-prompted SAM is a plausible but under-evidenced increment: the prompt effect is not isolated from added architecture and no ablation supports it.","tokens_in":6336,"tokens_out":3447,"would_cite":false,"duration_ms":28514,"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":"Cancer grade, encoded as a Grad-CAM++ heat map from a ViT classifier and fused into SAM through a prompt adapter, improves gland segmentation on GlaS beyond fine-tuned SAM alone.","keywords":["gland segmentation","Segment Anything Model","cancer grade prompt","Grad-CAM++","benign malignant classification","colorectal histology","GlaS dataset"],"falsifier":"Fix all training settings and replace the Grad-CAM++ heat map with a constant-valued map of the same size before it enters the prompt adapter; if the scores remain close to the prompted SAM results, the grade signal is not the active ingredient.","tokens_in":5317,"feed_emoji":"🔬","tokens_out":9148,"duration_ms":69224,"temperature":0.7,"pith_summary":"Cancer grade is a categorical clinical label, and this paper asks whether it can be turned into a spatial prompt for gland segmentation. The proposed model fine-tunes a Vision Transformer to classify benign versus malignant glands, explains that decision with Grad-CAM++, and feeds the resulting heat map through a small prompt adapter into a fine-tuned Segment Anything Model (SAM), a foundation model for image segmentation. The segmentation branch predicts gland and contour masks simultaneously, and the overlap between them is removed to separate touching glands. On the GlaS colorectal benchmark, prompted SAM-H reports F1 0.929/0.841, object Dice 0.921/0.881, and object Hausdorff 41.189/74.300 on test sets A and B, improving over fine-tuned SAM alone at all three scales and over the listed contest entries. If the idea holds, it suggests that a simple image-level label can act as a cheap, effective segmentation prompt in medical imaging.","feed_headline":"A cancer-grade heat map lifts SAM's gland scores on GlaS","feed_subtitle":"Prompted SAM-H beats fine-tuned SAM and every GlaS contest entry listed on F1, Dice, and Hausdorff.","key_machinery":"The central object is the grade prompt: a one-channel Grad-CAM++ heat map produced by a fine-tuned ViT that has just classified the gland as benign or malignant. The prompt adapter processes this heat map by concatenating it with the original image and passing the four-channel tensor through two convolutional layers with batch normalization and ReLU, then adds the result to the signal fed into SAM's gland prompt encoder. The segmentation branch splits into gland and contour predictions, each with its own prompt encoder and mask decoder, and post-processing removes the predicted contour overlap from the gland mask so adjacent glands stay separate. Training is staged: the classification branch is fine-tuned first, then the shared image encoder, prompt adapter, and gland decoder, then the contour decoder with the image encoder frozen, and a U-Net-style weight map weights the pixel-wise MSE loss.","core_discovery":"The central claim is that a binary cancer-grade label can carry useful spatial information sufficient to guide a foundation segmentation model. The benign/malignant ViT classifier emits a Grad-CAM++ heat map, and that one-channel map, concatenated with the original image and compressed by two convolutional layers in a prompt adapter, becomes the prompt given to SAM's gland mask decoder. With this grade prompt, prompted SAM-H reaches F1 0.929/0.841, object Dice 0.921/0.881, and object Hausdorff 41.189/74.300 on GlaS test sets A and B, outperforming fine-tuned SAM-H and every GlaS benchmark entry listed in the paper. The paper attributes the gain to the prior knowledge of gland type embedded in the heat map, which lets the segmentation model anticipate the appearance of the target.","pith_inferences":["A direct test the paper leaves undone: replace the heat map with a blank or shuffled one while freezing all adapter parameters; if segmentation stays at the prompted level, the grade information is not the cause of the gain.","The recipe is portable to other histology tasks where a coarse image-level label exists, such as nuclei or tissue-type segmentation, and the paper names cell segmentation as future work.","Because SAM already supports point, box, and text prompts, the grade heat map could be combined with those prompt types rather than used alone, which may give even stronger guidance without new annotations."],"forward_implications":["Prompted SAM matches or beats unprompted fine-tuned SAM at every scale tested (ViT-B, ViT-L, ViT-H), so the benefit is not tied to one model size.","The joint gland-and-contour prediction with overlap removal gives a practical way to separate touching glands, a known failure mode in gland segmentation.","A single forward path yields both a benign/malignant classification and a gland/contour segmentation, so the two clinical tasks are solved together.","The prompted SAM-H numbers top the GlaS benchmarks listed in the paper, so the grade-prompt recipe is competitive with task-specific gland segmentation networks."],"supporting_citations":[{"why":"Supplies the Segment Anything Model architecture, including the image encoder, prompt encoder, and mask decoder that the paper fine-tunes.","marker":"[5]"},{"why":"Supplies Grad-CAM++, the method that turns the ViT grade classifier's decision into the heat map used as the segmentation prompt.","marker":"[13]"},{"why":"Supplies the GlaS challenge images, annotations, evaluation metrics, and the benchmark entries the prompted SAM is compared with.","marker":"[15]"},{"why":"Supplies the Vision Transformer architecture that the classification branch is built on and fine-tunes.","marker":"[12]"},{"why":"Supplies the U-Net weight-map idea used to weight the pixel-wise MSE loss during training.","marker":"[14]"},{"why":"Supplies the pre-trained deit-base-patch16-224 model used as the initialization for the classification branch.","marker":"[16]"}],"fun_headline_variants":["Cancer grade as prompt lifts SAM's gland scores","Grade-guided SAM beats fine-tuned base on GlaS","SAM-H with grade prompt sets new GlaS SOTA","Benign or malignant? Prompt SAM for better glands","A heat map hint helps SAM segment glands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole improvement rests on the assumption that the benign/malignant heat map, not the extra network layers added at the same time, is what makes segmentation better; the paper does not test this by removing the heat map.","fun_headline_variants_meta":{"raw":{"variants":["Cancer grade as prompt lifts SAM's gland scores","Grade-guided SAM beats fine-tuned base on GlaS","SAM-H with grade prompt sets new GlaS SOTA","Benign or malignant? Prompt SAM for better glands","A heat map hint helps SAM segment glands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1148,"prompt_tokens":843,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":230}},"tokens_in":459,"tokens_out":305,"duration_ms":3833,"temperature":1.0,"reasoning_tokens":230,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:51:20.346085+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix all training settings and replace the Grad-CAM++ heat map with a constant-valued map of the same size before it enters the prompt adapter; if the scores remain close to the prompted SAM results, the grade signal is not the active ingredient.","supporting_citations":[{"cited_title":"Primar- ily trained on a broad dataset, it demonstrates impressive zero-shot performance involving natural images","cited_arxiv_id":null,"evidence_quote":"Supplies the Segment Anything Model architecture, including the image encoder, prompt encoder, and mask decoder that the paper fine-tunes."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the Vision Transformer architecture that the classification branch is built on and fine-tunes."},{"cited_title":"Segment anything in medical images,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net weight-map idea used to weight the pixel-wise MSE loss during training."},{"cited_title":"Segment anything model for medical image analysis: an experimental study,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained deit-base-patch16-224 model used as the initialization for the classification branch."}],"review_version":1}