{"id":"fc31712f-b658-4457-a87c-09d8e7376d34","arxiv_id":"2606.24214","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"MorVess improves pulmonary vessel segmentation by jointly predicting vessel masks, distance maps, and thickness maps using a 2.5D SAM adapter and global-local fusion for better small-vessel recovery and connectivity.","lead":"The paper introduces MorVess, a morphology-aware deep learning framework for pulmonary vessel segmentation in CT scans that jointly predicts masks along with distance and thickness maps while adapting the SAM foundation model via a 2.5D adapter. A smart generalist might read it to understand practical ways to embed geometric shape constraints into pretrained vision models for medical imaging tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Joint distance/thickness map prediction may introduce unvalidated errors into main segmentation task","rationale":"The reader's weakest assumption directly identifies the unverified link between auxiliary map quality and final segmentation gains. Because the assessment was abstract-only, the same assumption remains the single most load-bearing point even after noting the full-text placeholder; no other internal inconsistency is visible from the given material.","tokens_in":1703,"tokens_out":301,"duration_ms":16665,"concrete_test":"Run the ablation that disables the distance and thickness prediction heads and their loss terms while keeping the SAM adapter and fusion block fixed; recompute Dice/clDice/HD95 on both benchmarks. If the drop is <1% on clDice/HD95, the morphology supervision is not load-bearing for the headline result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on jointly predicting masks + distance + thickness maps to supply explicit geometric supervision for boundaries and topology. This only improves small-vessel recovery and clDice/HD95 if the auxiliary heads produce sufficiently accurate maps whose losses reinforce rather than compete with the primary mask loss. The abstract provides no separate error metrics for the distance/thickness outputs, no ablation removing those heads, and no analysis of whether their predictions are used at inference or only during training. If the auxiliary predictions are noisy, the reported gains could be attributable to the 2.5D SAM adapter or global-local fusion instead.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes MorVess, a morphology-aware framework for pulmonary vessel segmentation in CT that jointly predicts binary masks, distance maps, and thickness maps to enforce geometric constraints on boundaries and topology. It incorporates a lightweight 2.5D adapter to bridge 3D context with 2D SAM representations and a global-local fusion block to aggregate multi-level semantics and geometric cues. The central claim is that this yields superior Dice, clDice, and HD95 scores on two challenging pulmonary CT benchmarks, with particular gains in small-vessel recovery and global connectivity.","tokens_in":1828,"tokens_out":455,"duration_ms":16439,"significance":"If the empirical claims are substantiated, the approach of embedding explicit geometric supervision via auxiliary distance and thickness heads into a SAM-adapted architecture could meaningfully advance topology-preserving segmentation for sparse, multi-scale tubular structures in medical imaging. The open-source code release is a positive factor for reproducibility.","major_comments":[{"comment":"Abstract: The central claim of superior Dice, clDice, and HD95 performance (with substantial gains in small-vessel recovery) is asserted without any numerical values, baseline comparisons, ablation studies, or error analysis, which prevents assessment of whether the data and method support the claims.","section":"Abstract"},{"comment":"Method (joint prediction of auxiliary maps): No separate quantitative metrics (e.g., MAE or correlation) are reported for the distance and thickness map heads, nor is there an ablation removing these heads; this leaves open whether the auxiliary predictions reinforce or compete with the primary mask loss and whether they are used at inference.","section":"Method"},{"comment":"Experiments: The absence of any reported results, tables, or figures quantifying the claimed improvements on the two benchmarks means the load-bearing assertion of better small-vessel recovery and connectivity cannot be evaluated.","section":"Experiments"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":"The manuscript as presented (abstract plus description) supplies no empirical content, which is unusual for a CV methods paper and makes it difficult to judge scope fit or novelty relative to existing multi-task vessel segmentation work."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. The comments correctly identify that the current manuscript version lacks explicit numerical results, ablations, and auxiliary-task metrics in the sections highlighted. We will revise the manuscript to incorporate these elements, thereby strengthening the empirical support for our claims.","responses":[{"response":"We agree that the abstract should contain concrete numerical evidence. In the revised manuscript we will insert the key quantitative results (Dice, clDice, HD95) together with the main baseline comparisons and a brief reference to the ablation findings that demonstrate the contribution of the morphology-aware components.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The central claim of superior Dice, clDice, and HD95 performance (with substantial gains in small-vessel recovery) is asserted without any numerical values, baseline comparisons, ablation studies, or error analysis, which prevents assessment of whether the data and method support the claims."},{"response":"We will add MAE and Pearson correlation values for both auxiliary heads on the validation sets. We will also include an ablation that removes the distance and thickness heads while keeping all other components fixed. The auxiliary maps are used exclusively during training to provide geometric supervision; at inference only the binary mask is output. This clarification and the new quantitative results will be inserted into the Method section.","revision_made":"yes","referee_comment":"[Method] Method (joint prediction of auxiliary maps): No separate quantitative metrics (e.g., MAE or correlation) are reported for the distance and thickness map heads, nor is there an ablation removing these heads; this leaves open whether the auxiliary predictions reinforce or compete with the primary mask loss and whether they are used at inference."},{"response":"We will insert the full quantitative tables (including per-method Dice, clDice, HD95, and small-vessel-specific metrics) and the corresponding figures for both benchmarks. These tables will also report the ablation results mentioned above so that the contribution of each design choice can be directly assessed.","revision_made":"yes","referee_comment":"[Experiments] Experiments: The absence of any reported results, tables, or figures quantifying the claimed improvements on the two benchmarks means the load-bearing assertion of better small-vessel recovery and connectivity cannot be evaluated."}],"tokens_in":1404,"tokens_out":498,"duration_ms":15802,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's core move is a multi-task setup that predicts vessel masks together with distance and thickness maps, using a lightweight 2.5D adapter on SAM and a global-local fusion block to keep topology intact on pulmonary CT.\n\nWhat stands out as new is the concrete choice to supervise boundaries and diameter transitions directly through those two geometric heads rather than relying only on mask loss. The code release is also useful for anyone who wants to test the same priors.\n\nThe main weakness is that the abstract asserts better Dice, clDice, and HD95 without any actual scores, baseline tables, or ablation results. There are also no reported errors on the distance or thickness outputs themselves, so it is impossible to tell whether those heads reinforce the mask or simply add noise. The stress-test point about unvalidated auxiliaries therefore lands: if the extra predictions are inaccurate, the reported connectivity gains could come from the adapter or fusion alone.\n\nThis work is aimed at people already doing medical vessel segmentation on CT. A reader in that niche might pick up the architecture details and the geometric supervision idea, but only if the full experiments include proper controls and separate metrics for the auxiliary tasks.\n\nI would send it to peer review so the authors can supply the missing numbers and ablations; the idea is straightforward enough that referees can judge it quickly once the evidence is there.","headline":"MorVess adds explicit distance and thickness map prediction plus a 2.5D SAM adapter to vessel segmentation, but the abstract supplies no numbers or ablations to show the auxiliaries actually help.","tokens_in":2345,"tokens_out":361,"would_cite":false,"duration_ms":8480,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"MorVess jointly predicts vessel masks with distance and thickness maps to improve pulmonary vessel segmentation in CT scans.","keywords":["pulmonary vessel segmentation","morphology-aware segmentation","distance map","thickness map","CT imaging","vascular topology","deep learning","foundation model adaptation"],"falsifier":"An ablation experiment on the same two CT benchmarks in which removing the distance-map and thickness-map prediction heads produces no drop or an increase in clDice and HD95 scores.","tokens_in":2619,"feed_emoji":"🫁","tokens_out":685,"duration_ms":16215,"temperature":0.7,"pith_summary":"The paper introduces MorVess to address limitations in segmenting sparse and tortuous pulmonary vessels where small branches are lost and topology is hard to maintain. Standard models optimize only binary masks and lack explicit geometric constraints, leading to poor recovery of fine structures. MorVess adds joint prediction of distance maps for centerline consistency and thickness maps for diameter transitions, alongside a 2.5D adapter and fusion block to incorporate 3D context into foundation model features. This produces higher Dice, clDice, and HD95 scores on two CT benchmarks with better small-vessel recovery and connectivity. A sympathetic reader would care because accurate vessel maps support clinical quantification of lung structure.","feed_headline":"MorVess recovers more small pulmonary vessels via joint map prediction","feed_subtitle":"Predicting masks together with distance and thickness maps raises Dice and connectivity scores on lung CT benchmarks.","key_machinery":"Joint prediction of vessel masks, distance maps, and thickness maps that supplies explicit supervision for boundaries and topology, together with a 2.5D adapter and global-local fusion block.","core_discovery":"MorVess is a morphology-aware segmentation framework that integrates differentiable geometric priors with large-scale foundation model adaptation. It jointly predicts vessel masks, distance maps, and thickness maps to supply explicit supervision for vascular boundaries, centerline consistency, and smooth diameter transitions. A lightweight 2.5D adapter bridges 3D spatial context and 2D SAM representations while a global-local fusion block aggregates multi-level semantics and geometric cues. On two challenging pulmonary CT benchmarks the method yields superior Dice, clDice, and HD95 scores and substantially improves small-vessel recovery and global connectivity.","pith_inferences":["The same joint-map supervision could transfer to segmentation of other tubular anatomy such as coronary or cerebral vessels.","Clinical workflows that rely on vessel diameter measurements might obtain more consistent results without additional post-processing steps.","The framework suggests a route for embedding geometric priors into other foundation-model adaptations in medical imaging."],"forward_implications":["Small branches become recoverable because distance and thickness supervision enforce centerline and diameter consistency.","Global connectivity improves because the geometric maps reduce fragmentation under voxel-wise loss alone.","The 2.5D adapter allows pretrained 2D foundation models to handle 3D tubular structures without full 3D retraining.","Quantitative vessel analysis gains reliability from the explicit diameter and boundary predictions."],"fun_headline_variants":["MorVess predicts vessel masks distance and thickness maps","MorVess integrates geometric priors with SAM for vessel segmentation","Joint map prediction aids MorVess small vessel recovery","MorVess uses 2.5D adapter for SAM based vessel segmentation"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That jointly predicting distance and thickness maps will supply effective explicit supervision for vascular boundaries and topology without the auxiliary predictions introducing errors.","fun_headline_variants_meta":{"raw":{"variants":["MorVess predicts vessel masks distance and thickness maps","MorVess integrates geometric priors with SAM for vessel segmentation","Joint map prediction aids MorVess small vessel recovery","MorVess uses 2.5D adapter for SAM based vessel segmentation"]},"model":"grok-4.3","cost_usd":0.007209,"raw_usage":{"total_tokens":3344,"prompt_tokens":706,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":72087000,"prompt_tokens_details":{"text_tokens":706,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2571,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":706,"tokens_out":67,"duration_ms":17360,"temperature":1.0,"reasoning_tokens":2571,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T00:44:05.726978+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An ablation experiment on the same two CT benchmarks in which removing the distance-map and thickness-map prediction heads produces no drop or an increase in clDice and HD95 scores.","supporting_citations":[],"review_version":1}