{"id":"763c59aa-804a-4724-9ff3-6607a66562c2","arxiv_id":"2411.18169","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PDZSeg overlays visual prompts on images and fine-tunes DINOv2 with LoRA to segment the dissection zone in endoscopic submucosal dissection, outperforming baselines on the new ESD-DZSeg dataset.","lead":"This paper presents PDZSeg, a model that adds visual prompts (scribbles, boxes, points) directly onto endoscopic images and fine-tunes a DINOv2 foundation model to outline the dissection zone in robot-assisted ESD surgery. It also introduces a new ex-vivo porcine ESD dataset and reports gains over standard segmentation baselines when prompts are provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table 1 contradicts the headline claim: with the long-scribble prompt, Point-Rend beats PDZSeg on dissection-zone IoU (74.73 vs 74.06) and mean IoU (86.95 vs 86.60), so 'outperforms SOTA' is unsupported.","rationale":"The reader's conditional verdict is well-founded because the paper's contribution and clinical value hang on the claim of outperforming SOTA. I identified a more direct, internal failure than the prompt-reliability concern: Table 1 itself contains a disconfirming row. This does not mean the method is useless or the dataset is invalid; it means the paper must either correct the claim or add statistical support. The prompt-shortcut concern remains relevant and is complementary: without tests with misplaced or noisy prompts, it is unclear whether the model uses prompts to refine boundaries or merely follows the drawn region. Both concerns point to the same practical remedy: qualify the claim and add targeted experiments. The final verdict stays conditional because the work is a legitimate extension with a new dataset, but the abstract and conclusion need revision before acceptance.","tokens_in":9398,"tokens_out":7393,"duration_ms":66035,"concrete_test":"Using the released code and the ESD-DZSeg test split, rerun the long-scribble condition for Point-Rend and PDZSeg with at least 5 random seeds and report dissection-zone IoU as mean±std. If the Point-Rend advantage (74.73 vs 74.06) persists or falls within noise, the conclusion must be revised to say PDZSeg is competitive but not uniformly superior, or the claim must be restricted to the prompt modalities where PDZSeg actually wins.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PDZSeg outperforms state-of-the-art segmentation methods on ESD-DZSeg. Table 1 is the direct in-domain comparison for this claim. In the exact condition the paper identifies as best, long scribbles, Point-Rend exceeds PDZSeg on every reported metric: dissection-zone IoU 74.73 vs 74.06, Dice 85.54 vs 84.30, no-go-zone IoU 99.17 vs 99.15, mean IoU 86.95 vs 86.60, and mean Dice 92.56 vs 91.93. Since this is not an edge case but the highest-performing prompt mode, the abstract and conclusion statement that PDZSeg 'outperforms state-of-the-art segmentation approaches' is false as stated. All numbers come from a single test split with no error bars or significance tests, so even the conditions where PDZSeg ranks first are not statistically established. The contribution therefore reduces to 'competitive in some prompt modes and inferior in the best mode,' which is materially weaker than the paper's central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PDZSeg, a dissection-zone segmentation model for robot-assisted endoscopic submucosal dissection. A DINOv2 ViT backbone is fine-tuned with LoRA and connected to an all-MLP decoder; user-provided visual prompts (point, short/long scribble, bounding box) are overlaid directly on the RGB image. The authors introduce the ESD-DZSeg benchmark built from 1,849 ex-vivo porcine images from 21 procedures and report in-domain, prompt-availability, and robustness experiments. They conclude that PDZSeg outperforms state-of-the-art segmentation methods and effectively uses visual prompts for dissection-zone suggestion.","tokens_in":9581,"tokens_out":7335,"duration_ms":66051,"significance":"The dataset and code release are useful community assets, and the prompt-overlay design is simple enough to be applied to existing baselines. The intended application—real-time contour guidance from an experienced surgeon's prompt—is clinically meaningful. The visual prompt is a legitimate test-time input, so I do not view the performance gain from adding it as circular; the concern is whether the model is using the prompt as intended. However, the headline performance claim is contradicted by Table 1 in the long-scribble condition, and the experimental design lacks statistical validation. The paper is best viewed as a preliminary benchmark and method proposal rather than a demonstrated SOTA result.","major_comments":[{"comment":"Table 1, long-scribble rows: in the condition the authors themselves identify as the best prompt type, Point-Rend outperforms PDZSeg on every reported metric (dissection-zone IoU 74.73 vs 74.06, Dice 85.54 vs 84.30, no-go-zone IoU 99.17 vs 99.15, mean IoU 86.95 vs 86.60, mean Dice 92.56 vs 91.93). Consequently, the abstract and Section 5 statement that PDZSeg 'outperforms state-of-the-art segmentation approaches' is not supported by the data and should be qualified or corrected.","section":"Abstract and Table 1"},{"comment":"All quantitative claims rest on a single train/test split of four videos (369 test images) with no error bars, confidence intervals, or significance tests, and many favorable differences are small (e.g., 65.45 vs 64.15 dissection-zone IoU with short-scribble prompts). The paper should add repeated runs with variance estimates and paired tests, or explicitly present the results as preliminary observations.","section":"Section 4.2, Tables 1-4"},{"comment":"The experiments never vary prompt quality or location, so the paper does not test whether the model refines tissue boundaries from the prompt or simply segments in the vicinity of the overlaid cue. Adding evaluations with noisy, misplaced, or contradictory prompts would support the claimed mechanism and the clinical-safety motivation.","section":"Sections 3.2.1 and 4.2"},{"comment":"The robustness experiment applies five image corruptions at a single fixed severity level (3) and reports no severity-response analysis or uncertainty quantification, so the conclusion that PDZSeg is robust is not established beyond this specific setting.","section":"Section 4.2, Table 4"}],"minor_comments":[{"comment":"The implementation states that Vit-Base has a 'feature dimension of 784', but DINOv2 ViT-Base uses an embedding dimension of 768; please correct or clarify the intended dimension.","section":"Section 4.1"},{"comment":"Several table captions contain typos: 'T able 1', 'T able 2', and 'T able 3' appear with an extra space, and Section 3.1 writes 'Dissecon Zone Segmentation' instead of 'Dissection Zone Segmentation'.","section":"Various section headers"},{"comment":"The Medical-SAM-Adapter baseline reports only dissection-zone IoU and Dice and only under the point-prompt condition; all other cells are '/'. Please complete the table or state explicitly why those results are unavailable.","section":"Table 1"},{"comment":"The attention equations write Wqx, Wkx, and Wvx without subscript formatting; these should be W_q x, W_k x, and W_v x for readability.","section":"Section 3.2.2"},{"comment":"The caption says 'The first two columns display the ground truth and the segmentation masks predicted by our model', but the layout appears to show one ground-truth column followed by multiple prediction columns; please re-check the caption.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the dataset is the real contribution, and the method is a sensible combination of known pieces—DINOv2, LoRA, an MLP decoder, and prompt-overlaid RGB. The central claim is not supported by the paper's own table. In the long-scribble condition, which the paper identifies as best, Point-Rend gets dissection-zone IoU 74.73 vs PDZSeg's 74.06, Dice 85.54 vs 84.30, and the same pattern on mean IoU and mean Dice. So \"outperforms state-of-the-art segmentation approaches\" is false as stated; the accurate claim is \"competitive with, and sometimes ahead of, baselines under point, box, and short-scribble prompts.\" That is a material difference.\n\nWhat is actually new: ESD-DZSeg, 1,849 annotated frames from 21 robot-assisted ESD ex-vivo porcine videos, split by video with 4 test videos. That is a new benchmark for dissection zone segmentation and it fills a real gap. The prompt-overlay design is simple and practical, and the behavior is plausible—prompts help, long scribbles help most, and mixed-prompt training degrades gracefully. The robustness table at severity 3 shows PDZSeg handling corruption better than Point-Rend in most cells. I believe the experiments are internally consistent; no invented metrics and no fitting to the test set.\n\nSoft spots, in proportion. First, no error bars, no significance tests, single split, test set from only four videos. The 2–3 point differences over Point-Rend could easily be noise. Second, the \"outperforms\" sentence needs to be rewritten or the long-scribble comparison addressed; as it stands a referee will catch it immediately. Third, there is no test with noisy or misplaced prompts, which is the obvious failure mode—the model may just be cropping attention to the prompted region rather than learning tissue boundaries. I share the reader's concern there. Fourth, no user study despite claims about improving \"user experience\"; that is descriptive, not evidence. Fifth, the dataset is not publicly released, only code, so the benchmark is not yet reproducible. Minor: the robustness evaluation uses one fixed severity level.\n\nWho this is for: surgical vision researchers, especially those working on interactive or ESD segmentation. It deserves a serious referee, but at best a conditional accept—the central claim must be corrected, significance or per-video breakdowns added, and prompt-error analysis included before publication. I would send it to review, and I would not cite it in its current form.","headline":"A useful new benchmark and a sensible prompting recipe, but the central claim is false on the paper's own best setting: long-scribble Point-Rend beats PDZSeg on every metric in Table 1.","tokens_in":10188,"tokens_out":3519,"would_cite":false,"duration_ms":28373,"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":"PDZSeg adapts a vision foundation model to turn surgeon scribbles and bounding boxes into precise dissection-zone contours for robot-assisted ESD.","keywords":["dissection zone segmentation","visual prompting","scribble prompts","bounding box prompts","robot-assisted ESD","DINOv2","LoRA","endoscopic submucosal dissection"],"falsifier":"Take the trained PDZSeg and evaluate it on the same test set with prompts systematically displaced outside the dissection zone, partially erased scribbles, or scribbles that cross the true boundary; if IoU remains high and the predicted mask does not move with the prompt, the overlay is acting as a localization hint rather than as boundary information, which would falsify the claimed mechanism.","tokens_in":9151,"feed_emoji":"✏️","tokens_out":7637,"duration_ms":62699,"temperature":0.7,"pith_summary":"The paper introduces PDZSeg, a segmentation model that takes a visual prompt—a point, a scribble, or a bounding box drawn by the surgeon—overlays it directly on the RGB endoscopic image, and predicts the dissection zone during robot-assisted endoscopic submucosal dissection (ESD). To make this work, the authors fine-tune the DINOv2 foundation model with LoRA adapters on a new benchmark dataset, ESD-DZSeg, built from 21 ex-vivo porcine robotic ESD procedures and 1,849 expert-annotated images. They report that prompted versions outperform their unprompted baseline and, on average across prompt types, outperform the comparison segmentation models, with the long-scribble prompt giving the best dissection-zone IoU and robust behavior under image corruption. If accepted, the work offers a practical way for an experienced surgeon to guide a less experienced one by simply sketching the safe dissection region, and it provides the first benchmark for this task.","feed_headline":"Scribble prompts lift ESD dissection-zone IoU from 47.6 to 74.1","feed_subtitle":"Overlaying an expert's sketch lets a foundation model mark the safe cutting boundary in robot-assisted endoscopy.","key_machinery":"The central mechanism is visual referral by image overlay: instead of a learned prompt embedding, the prompt is drawn into the RGB input itself, so the model sees the user's mark as part of the scene. The encoder is DINOv2, a self-supervised vision transformer, kept largely frozen and adapted through LoRA, a low-rank adaptation that injects small trainable matrices into the query and value projections of each attention block. The decoder is an all-MLP network that fuses multi-level features from transformer blocks 3, 6, 9, and 12 and maps them to a two-class mask (dissection zone and no-go zone), so the prompt acts as a location and boundary cue.","core_discovery":"On the paper's own terms, the discovery is that visual prompts can be folded into segmentation without a separate prompt-encoding branch: painting the user's marker onto the image and fine-tuning a foundation model is enough to steer the model toward the dissection zone. The evidence is the ESD-DZSeg results: with a long scribble, the model reaches 74.06 IoU and 84.30 Dice on the dissection zone, versus 47.55 IoU and 63.01 Dice for the same architecture without a prompt, and the model retains higher IoU than a strong baseline under Gaussian noise, smoke, contrast, brightness, and motion blur when the same prompt is provided. The paper also positions the mixed-prompt experiments as showing that partial or heterogeneous prompt availability during training still yields strong segmentation performance.","pith_inferences":["The overlay design suggests the prompt may function as a soft region prior rather than a strict boundary constraint; a direct ablation that removes or erodes the scribble inside the target region would clarify whether the model is localizing or contour-refining.","Because all images come from ex-vivo porcine tissue with a fixed endoscope and robotic platform, transfer to live human ESD is untested; a cross-domain evaluation with different centers, in-vivo tissue, and varying scope optics would be the most direct next experiment.","The paper evaluates only correct, expert-style prompts; robustness to misplaced, partial, or adversarial prompts is not measured, so a natural safety extension is to pair the model with a prompt-quality check or allow the surgeon to correct the suggested contour.","If prompt-conditioned robustness holds in vivo, the approach could be extended from a suggestion tool to an intraoperative alarm that detects when the electric knife approaches the predicted boundary."],"forward_implications":["With a long-scribble prompt, PDZSeg's dissection-zone IoU rises from 47.55 (no prompt) to 74.06 on ESD-DZSeg, showing that the type of visual prompt strongly modulates performance.","The visual-prompt overlay also improves standard segmentation baselines such as DeepLabv3, whose dissection-zone IoU climbs from 45.95 to 69.24 with a bounding box, so the prompting scheme is not tied to the proposed architecture alone.","Training with mixed prompt availability (prompted-to-unprompted ratios of 6:4, 5:5, and 4:6) keeps performance well above unprompted-only training, supporting deployment where the surgeon does not prompt every frame.","Under five corruption types, the long-scribble-prompted model keeps dissection-zone IoU between 64.91 and 73.52, while the Point-Rend baseline drops to 2.89 under Gaussian noise, indicating prompt-conditioned robustness.","The paper introduces the ESD-DZSeg dataset, 1,849 images from 21 ex-vivo porcine robotic ESD videos, as a benchmark for dissection zone segmentation and visual prompt interpretation."],"supporting_citations":[{"why":"Supplies the dual-arm robotic ESD platform and the ex-vivo porcine surgical videos from which the ESD-DZSeg dataset is built.","marker":"[3]"},{"why":"Provides the pre-trained DINOv2 vision foundation model used as the image encoder that PDZSeg adapts with LoRA.","marker":"[4]"},{"why":"Defines the low-rank adaptation method used to fine-tune the encoder efficiently on the specialized dataset.","marker":"[5]"},{"why":"Point-Rend, a baseline segmentation model compared against PDZSeg in the prompt and corruption experiments.","marker":"[8]"},{"why":"DeepLabv3, a baseline segmentation model that the prompted overlay also improves in the comparisons.","marker":"[10]"},{"why":"Segment Anything, cited as the source of prompt-driven segmentation inspiration and the need for prompt embeddings.","marker":"[12]"},{"why":"Medical-SAM-Adapter, the medical-adaptation baseline used in the with-point-prompt evaluation.","marker":"[13]"},{"why":"The corruption and perturbation benchmark used to design the five image corruption robustness tests.","marker":"[20]"}],"fun_headline_variants":["Scribble prompt boosts ESD dissection-zone IoU to 74.1","Visual prompts steer foundation model for safer endoscopy","Scribble prompts lift IoU 26 points for ESD dissection zone","User sketch steers AI to mark safe cutting line in ESD","PDZSeg: draw on image, get IoU 74.1 for dissection zone"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that overlaying a prompt teaches the model to refine tissue boundaries rather than to segment only tissue near the prompt, and that an expert's placed prompt reliably indicates the dissection zone; neither is tested, and the dataset is ex-vivo porcine from one center.","fun_headline_variants_meta":{"raw":{"variants":["Scribble prompt boosts ESD dissection-zone IoU to 74.1","Visual prompts steer foundation model for safer endoscopy","Scribble prompts lift IoU 26 points for ESD dissection zone","User sketch steers AI to mark safe cutting line in ESD","PDZSeg: draw on image, get IoU 74.1 for dissection zone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000869,"raw_usage":{"total_tokens":3764,"prompt_tokens":943,"completion_tokens":2821,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":2723}},"tokens_in":559,"tokens_out":2821,"duration_ms":17403,"temperature":1.0,"reasoning_tokens":2723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:26:21.734717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained PDZSeg and evaluate it on the same test set with prompts systematically displaced outside the dissection zone, partially erased scribbles, or scribbles that cross the true boundary; if IoU remains high and the predicted mask does not move with the prompt, the overlay is acting as a localization hint rather than as boundary information, which would falsify the claimed mechanism.","supporting_citations":[{"cited_title":"The International Journal of Robotics Research 43(3), 281–304 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the dual-arm robotic ESD platform and the ex-vivo porcine surgical videos from which the ESD-DZSeg dataset is built."}],"review_version":1}