{"id":"ac21b85f-1f16-4357-b2ae-ea02227588c0","arxiv_id":"2607.10953","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Organ-hierarchical knowledge extracted from radiology reports improves CT vision-language pretraining for zero-shot abnormality diagnosis and retrieval.","lead":"OKA-CT turns free-text CT reports into organ-level structured knowledge and uses it to train better CT image representations. The method improves zero-shot disease detection and report-image matching on two chest CT benchmarks without dense expert labels.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Unvalidated LLM/parser organ slots may corrupt Stage-1 heads and Stage-2 soft targets, so gains may not prove the hierarchy mechanism.","rationale":"The reader correctly located the load-bearing premise: automatic organ slots must be good enough weak labels for both visual supervision and soft-target construction. That premise is stated in §III.B and reused in the Stage-1 heads and Stage-2 target matrix without any external validation. My concern is the same one, sharpened to the concrete risk that unmeasured extraction error plus case-level fallbacks and hand-tuned weights can produce the observed gains without proving the claimed anatomy-grounded mechanism. Ablations and patch-occlusion are supportive but still conditioned on the same extractor, so they do not close the gap. The paper is otherwise careful (honest limitations, two-stage freeze, multi-metric evaluation), so the appropriate stance remains CONDITIONAL rather than REJECT: the empirical superiority is plausible if extraction quality holds, but that condition is currently unchecked. The proposed radiologist-annotated subset + gold/auto/noised retrain is a single decisive test that would either secure or demote the central claim. No stronger internal inconsistency was found; the issue is evidential support for the mechanism, not an algebraic error.","tokens_in":14674,"tokens_out":770,"duration_ms":6918,"concrete_test":"On a stratified 200-report CT-RATE subset, have two radiologists annotate the four slots for lung/heart/esophagus/aorta; compute slot-level P/R/F1 of the RadGraph-XL+Qwen3-8B pipeline. Then retrain Full OKA-CT three ways: (i) gold slots, (ii) auto slots, (iii) auto slots with 30% random field flips. If auto vs gold drops macro AUROC by >2 points or noised slots still match Full OKA-CT within 1 point, the hierarchy-quality premise fails and the strongest claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes the AUROC/retrieval gains to reusable organ-hierarchical knowledge (abnormality, concept, location, attribute slots) that both supervises mask-pooled visual tokens in Stage 1 (Eqs. 3–5) and defines soft semantic neighborhoods for organ contrastive loss in Stage 2 (Eqs. 9–12, Fig. 4). That claim requires the RadGraph-XL + Qwen3-8B pipeline (§III.B, Fig. 3) to produce slots that are accurate enough as weak labels. The paper never measures extraction precision/recall against expert organ-level annotations, never reports inter-annotator agreement or error modes (negation, laterality, multi-organ findings), and never abates with deliberately noised or shuffled slots. Case-level disease labels are admitted as fallbacks when organ concepts are missing, so Stage-1 organ heads can partly re-learn global labels rather than true anatomy-grounded evidence. Soft-target weights (α=β=0.8, γ=0.2, σ=0.15) are hand-chosen; if slots are noisy, off-diagonal “weak positives” become false positives that dilute contrastive geometry. Ablations (Tables III–IV) show hierarchy helps relative to concept-only or Stage-2-only, but still use the same unvalidated extractor, so they do not isolate mechanism from label quality. Without an independent quality check, the reported 84.9 / 72.2 AUROCs and retrieval lifts remain consistent with better multi-task supervision or soft-label smoothing rather than anatomy-grounded hierarchy.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"OKA-CT is a two-stage CT-report vision-language pretraining framework that converts free-text radiology reports into organ-conditioned hierarchical knowledge (abnormality status, disease/finding concepts, location, and severity attributes) via RadGraph-XL parsing and Qwen3-8B structuring. Stage 1 trains a 3D ResNet-18 with global disease supervision plus organ-mask-pooled visual tokens supervised by the four hierarchy slots (Eqs. 1–5); Stage 2 freezes the visual encoder and performs global report-CT contrastive learning plus organ-level alignment with hierarchy-derived soft targets that treat non-paired cases sharing organ-level findings as weak positives (Eqs. 6–13, Fig. 4). A query-based global branch aggregates volumetric evidence. On CT-RATE the method reports zero-shot abnormality diagnosis AUROC 84.9 (vs. 77.8 for fVLM) and strong retrieval gains; on RAD-ChestCT AUROC is 72.2. Ablations (Tables III–IV), organ-group deltas, and patch-occlusion maps support the contribution of the hierarchy.","tokens_in":15177,"tokens_out":1233,"duration_ms":9625,"significance":"If the gains are truly driven by reusable organ-hierarchical report knowledge rather than generic multi-task or soft-label effects, the work offers a practical and scalable way to inject anatomy-grounded structure into CT VLP without dense expert organ annotations. The two-stage reuse of the same slots for visual supervision and soft-target contrastive neighborhoods is a clear design contribution relative to prior global (CT-CLIP, Merlin) and local/anatomy-aware (fVLM, CT-GLIP) methods. Empirical strengths include multi-metric diagnosis and retrieval results, cross-dataset transfer, stepwise ablations, Stage-1-only analysis, and qualitative occlusion evidence of anatomy-sensitive localization. These make the paper a solid candidate for the medical VLP literature once the weakest assumption—slot quality—is better supported.","major_comments":[{"comment":"§III.B and Fig. 3: The central claim attributes gains to organ-hierarchical slots z^o_i = {a,c,ℓ,h} used in Stage-1 heads (Eqs. 3–5) and Stage-2 soft targets (Eqs. 9–12). The paper never reports extraction precision/recall, error modes (negation, laterality, multi-organ findings), or agreement against expert organ-level annotations. Without this, Tables I–III cannot isolate the hierarchy mechanism from noisy multi-task labels or soft-label smoothing. A modest validation set or controlled noise/shuffle ablation of slots is needed to support the mechanism claim.","section":"§III.B, Fig. 3, Eqs. (3)–(5), (9)–(12)"},{"comment":"§III.B and Stage-1 objective (Eq. 4): Case-level disease labels are used as fallbacks when organ concepts are missing. This can allow organ heads to re-learn global labels rather than anatomy-grounded evidence, weakening the claim that Stage 1 injects organ-conditioned visual geometry. Please quantify fallback frequency and report Stage-1 performance with vs. without fallbacks, or restrict organ supervision to report-derived slots only.","section":"§III.B, Eq. (4), Table IV"},{"comment":"Tables I–II: Several baseline numbers are collected from prior comparisons and some metrics are missing (e.g., COLIPRI-C). No error bars, multiple seeds, or significance tests are reported for the 7.1-point AUROC gain over fVLM or the large report-image retrieval lift. Given free parameters λ1, λ2, α, β, γ, σ and temperatures, at least seed variance or a sensitivity sweep on the soft-target weights would strengthen confidence that the gains are stable.","section":"Tables I–II, §IV.B"}],"minor_comments":[{"comment":"Fig. 5: Occlusion heatmaps are informative but only qualitative; a quantitative summary (e.g., fraction of score drop inside organ masks vs. outside) would better support the anatomy-sensitivity claim.","section":"Fig. 5, §IV.D"},{"comment":"Notation: organ set O is restricted to lung/heart/esophagus/aorta; the discussion notes this limitation, but the abstract and introduction should more clearly scope claims to chest CT with these four organs.","section":"Abstract, §I, §V"},{"comment":"Eq. (11): The weighted contrastive loss uses raw target mass rather than row-normalized targets; a short justification or comparison to normalized soft targets would help readers.","section":"Eq. (11)"},{"comment":"Typos/clarity: “V ocabFine” spacing in Table II; “Stage~1” tilde usage is inconsistent with the rest of the text; some figure captions are dense and could be tightened.","section":"Table II, Abstract"}],"recommendation":"major_revision","confidential_remarks":"The technical idea is timely and the empirical package is above average for CT VLP, but the unvalidated LLM/parser slots are a load-bearing gap that a careful medical-imaging venue will expect to see addressed. I would not reject on novelty grounds; major revision with a slot-quality check or noise ablation should be sufficient. Fit is appropriate for a methods-focused medical imaging or CV journal with a clinical ML track."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a solid CT-report VLP methods paper. The new piece is not “organ alignment” in the abstract—that already exists—but reusing a four-slot report hierarchy (abnormality, concept, location, attribute) for both Stage-1 mask-pooled visual heads and Stage-2 soft semantic targets, plus a light query global branch. They ship clear numbers: 84.9 AUROC on CT-RATE, 72.2 on RAD-ChestCT, big report-image retrieval lift, and occlusion maps that look more anatomy-aligned than CT-CLIP.\n\nWhat they do well: the design is readable, the ablations are stepwise (Stage-2 only → concept Stage-1 → full hierarchy → soft targets), Stage-1 alone already helps the visual classifier, and the limitations section is honest about weak slots, four chest organs, and domain shift. Citations cover the right prior art (CT-CLIP, Merlin, fVLM, CT-GLIP, RadGraph-XL, soft-positive contrastive work). Math is standard BCE/CE plus weighted contrastive; nothing load-bearing is hand-wavy in the equations.\n\nSoft spots, in proportion: the stress-test lands. They never measure RadGraph-XL + Qwen3-8B slot precision against expert organ labels, and case-level disease labels are fallbacks, so Stage-1 organ heads can partly relearn global labels. Soft-target weights are hand-set. No error bars, and some baselines are taken from prior tables. That means the gains are real as system results, but “anatomy-grounded hierarchy” is not fully isolated from multi-task supervision and soft-label smoothing. Scope is chest-only with four organs.\n\nWho it’s for: people building CT foundation models and report-supervised 3D VLP. Not a foundational theory paper; a useful engineering-science result if extraction quality holds under broader checks.\n\nI’d send it to peer review. Ask for extraction QA, a noise/shuffle control on slots, uncertainty estimates, and clearer separation of hierarchy vs multi-task effects. Worth engaging if you work in this lane.","headline":"Real CT-VLP gains from a coherent two-stage organ-hierarchy package; the extraction-quality gap is real but does not erase the empirical result.","tokens_in":15769,"tokens_out":543,"would_cite":true,"duration_ms":8641,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Organ-level structure from radiology reports, reused as supervision, makes CT vision-language models more accurate and anatomy-sensitive.","keywords":["Medical vision-language pretraining","Organ-hierarchical knowledge","Report knowledge extraction","Structured contrastive learning","CT-report alignment","Zero-shot abnormality diagnosis","Anatomy-grounded representation"],"falsifier":"If independent expert review of the extracted organ slots shows high error rates on the same organs, and replacing those slots with random or shuffled slots collapses the reported AUROC and retrieval gains, the central claim fails.","tokens_in":15561,"feed_emoji":"🫁","tokens_out":870,"duration_ms":7375,"temperature":0.7,"pith_summary":"Radiology reports describe findings organ by organ—abnormality status, disease concepts, location, and severity—but most CT vision-language pretraining either matches whole scans to whole reports or matches local patches to text fragments. This paper argues that the organ hierarchy itself is usable supervision. OKA-CT extracts that hierarchy automatically from free-text reports, then reuses it in two stages: first to train the CT encoder with organ-conditioned visual targets, and second to guide report–CT contrastive learning so that cases sharing the same organ-level findings become weak positives rather than hard negatives. A light query branch still builds a whole-scan embedding. On two public chest CT datasets the method raises zero-shot abnormality diagnosis and retrieval scores over prior CT vision-language baselines and, under patch occlusion, responds more strongly to disease-associated anatomy. The practical claim is that anatomy-grounded report structure is a scalable signal for better CT representations without dense manual labels.","feed_headline":"Organ hierarchy from reports lifts CT zero-shot diagnosis to 84.9 AUROC","feed_subtitle":"OKA-CT reuses report structure as weak supervision so CT models align better with disease anatomy.","key_machinery":"Organ-hierarchical knowledge (slots a, c, ℓ, h per organ) extracted by report parsing plus LLM structuring, reused as Stage-1 organ-conditioned heads on mask-pooled features and as Stage-2 soft target matrices that turn shared organ-level findings into weak positives in contrastive loss.","core_discovery":"OKA-CT shows that organ-hierarchical knowledge extracted from free-text radiology reports—abnormality, concept, location, and attribute slots—can be reused as weak supervision for both organ-conditioned visual pretraining and structured report–CT contrastive learning, producing CT representations that improve zero-shot diagnosis, retrieval, and sensitivity to disease-linked anatomical regions.","pith_inferences":["If extraction quality is the main bottleneck, better report parsers or lighter human verification of slots could raise performance more cheaply than scaling the visual backbone alone.","The soft-target idea may transfer to other paired medical modalities where free-text findings are also organ- or system-organized (e.g., MRI reports).","Domain shift between CT-RATE and RAD-ChestCT remains large; calibrating organ-slot distributions across institutions may be as important as the architecture itself."],"forward_implications":["CT vision-language models can improve zero-shot abnormality diagnosis by treating report organ hierarchy as reusable weak supervision rather than only whole-report text.","Non-paired CT–report pairs that share organ-level findings should be soft positives in contrastive learning, not uniform hard negatives.","Global CT embeddings become more sensitive to disease-associated anatomical regions when organ-conditioned supervision is injected before report alignment.","The same organ-slot pipeline can, in principle, extend beyond the four chest organs studied here once masks and report structuring cover a larger anatomy set."],"fun_headline_variants":["Organ hierarchy from reports lifts CT zero-shot AUROC to 84.9","OKA-CT reuses report organ structure for 84.9 AUROC CT diagnosis","Anatomy-grounded CT VLP via organ hierarchy reaches 84.9 AUROC","Report organ knowledge improves CT zero-shot diagnosis to 84.9","OKA-CT organ-conditioned alignment yields 84.9 AUROC zero-shot"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The automatically extracted organ slots from free-text reports are accurate enough as weak labels that supervising visual features and defining soft semantic neighborhoods improves rather than corrupts the learned evidence space.","fun_headline_variants_meta":{"raw":{"variants":["Organ hierarchy from reports lifts CT zero-shot AUROC to 84.9","OKA-CT reuses report organ structure for 84.9 AUROC CT diagnosis","Anatomy-grounded CT VLP via organ hierarchy reaches 84.9 AUROC","Report organ knowledge improves CT zero-shot diagnosis to 84.9","OKA-CT organ-conditioned alignment yields 84.9 AUROC zero-shot"]},"model":"grok-4.5","effort":"low","cost_usd":0.005536,"raw_usage":{"total_tokens":1534,"prompt_tokens":821,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":55360000,"prompt_tokens_details":{"text_tokens":821,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":621,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":821,"tokens_out":92,"duration_ms":4514,"temperature":1.0,"reasoning_tokens":621,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T08:05:17.194212+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"If independent expert review of the extracted organ slots shows high error rates on the same organs, and replacing those slots with random or shuffled slots collapses the reported AUROC and retrieval gains, the central claim fails.","supporting_citations":[],"review_version":1}