Same Evidence, Different Verdicts: How Delivery of CT Images and DICOM Metadata Changes VLM Quality Control Decisions
Same Evidence, Different Verdicts: Delivery of CT Images and DICOM Metadata Changes VLM Quality-Control Decisions
The authors tested 9 VLMs on 60 pairs of NLST chest CT: when a fake DICOM header arrived after the model had answered, it overturned 91%–100% of correct refusals; when delivered with the same prompt, it was mostly resisted.
- In clinical pipelines, CT images and DICOM metadata may be inconsistent; a truncated series, if it enters lung cancer screening volume follow-up as complete, would bias measurements; VLMs have been proposed for input quality control, but when images and metadata conflict, which one they believe and whether delivery order has an effect remain unclear.
- From 60 pairs of NLST CT, the authors construct 240 defect cases with exact geometric ground truth, have 9 VLMs make ADMIT/REFUSE decisions under conditions including image alone, image with a fake header in the same prompt, and a fake header delivered in a second round after answering (including as a tool result), and test the mitigation effect of one warning sentence.
- When the fake header and image share the same prompt, models are mostly conservative and refuse; when delivered after answering, it overturns most correct refusals (Haiku aligned 2%→91%, Nemotron 43%/97%→100%), and this holds in both directions; a neutral follow-up almost never overturns, and one warning sentence reduces second-round overturning to ≤9%; the pattern is the same on the 24 real mismatch pairs.
- The authors note: only one coronal slice is taken per series; the z-shift conflict in the native view partly depends on text labels; the header is a synthetic one-line summary; header-first was tested only on Claude; GPT/Gemini were not tested. The experiments cover 9 VLMs, 60 synthetic pairs plus 24 real pairs, with some models having only 15 patients.
- Threat model
- No adversarial attacker: study the behavior of VLMs in clinical imaging quality control. The victim system is a clinical/agentic pipeline that uses a VLM for CT input admission checks; the source of conflict is a one-line DICOM header summary derived by the pipeline (possibly outdated or erroneous), which can be delivered in the same prompt as the image, or as a user message or tool result after the model has answered.
- Evaluated models
- Nemotron-3.5-VL-120BClaude Haiku 5.5Claude Sonnet 5.5Gemma-4-31BLlama-3.2-90B-VisionLlama-3.2-11B-VisionMedGemma-1.5-4BMuse-Glimmer-30BQwen3-VL-8B
- Benchmarks
- NLST same-study reconstruction pairs(60 对合成缺陷 + 24 对真实不匹配,作者构建的测试床)
- Metrics
- SFR(silent failure rate)FRR(false refusal rate)Youden's JAUC(pixel-only 参考检查)
A paper by University of Georgia researcher Nicolas Avelar finds that in clinical imaging quality control, vision-language models' judgments on the same evidence depend on delivery order, with later-arriving sources more likely to overturn earlier judgments. On 60 NLST chest CTs with coverage defects, when the same erroneous DICOM header claiming the series was complete was given in the same prompt as the image, it overturned only about 2% of Claude Haiku 5.5's defect detections, whereas when delivered after the model answered, it overturned 91%, and Nemotron-3.5-VL rose from 43% to 100%; when the order was reversed, the image also overturned 98–100% of the Claude model's header errors. The effect is not general sycophancy: a neutral "please confirm" did not overturn any of 1,072 correct refusals, and the effect persisted when the header was delivered as a tool result (52–99%). A prompt saying "if the image conflicts with the header, trust the image" reduced second-round overturning to ≤9%, and the pattern held on 24 pairs of real mismatched data.
Why it was selectedThe paper tests nine VLMs with exact geometric ground truth, finds that the delivery order of the same evidence can flip quality-inspection conclusions, and provides a reusable validation procedure.
Paper analysis
What problem does this paper attempt to solve?
When a VLM performs clinical imaging input quality control, which to trust when the image conflicts with DICOM metadata, and whether delivery order changes the decision.
Core question: when CT images conflict with DICOM metadata, which one does the VLM's input admission decision believe, and does the delivery order of the evidence change the conclusion.
- Scenario and importance: lung cancer screening compares low-dose CT across rounds, and volume follow-up assumes each series covers the anatomical extent it claims; a series truncated in transit, if admitted as complete, would silently remove anatomy from the comparison and bias measurements (Sec. 1).
- Shortcomings of existing evaluations: VLMs are increasingly evaluated for image-level judgments (CT quality, robustness to corrupted visual evidence), but almost only on images alone; the authors argue that image-only evaluation may miss this type of failure (Sec. 1).
- Core observation: pipeline-derived header summaries may be stale or wrong (missing slices, still reporting complete coverage after a series is relinked); in agentic pipelines, metadata may arrive as a tool result after the visual judgment. The authors find that having the model see the image first and the metadata afterward makes it more likely to abandon a correct detection (summary).
- The paper proposes: a testbed with exact geometric ground truth (240 synthetic defect cases, 24 real mismatch pairs, 9 VLMs), a validation protocol for VLM input checks (Sec. 4), and a mitigation consisting of one warning sentence.
- Conflict setup (not an adversarial attack): the paper has no attacker; the source of the conflict is a one-line header summary derived by the pipeline from count/range fields, which falsely claims completeness for defect cases and falsely claims truncation for clean cases; the victim system is a clinical imaging pipeline that uses a VLM for ADMIT/REFUSE admission decisions; the decision rule is that the two series cover the same z-range with no missing slices (Sec. 2).
What related research exists?
Related work covers three categories: text overriding vision, medical sycophancy and fabricated evidence, and context interference; the authors state that this paper differs in the source of conflict, ground truth, and comparison design.
Text overriding vision and answers flipped by follow-up questions
- Deng et al. [4], Restrepo et al. [5]: in VLMs, text often overrides vision (cited by the paper as prior work).
- Laban et al. [6] (2023): "Are you sure?"-style follow-ups flip LLM answers (FlipFlop experiment).
Sycophancy and fabricated evidence in medical settings
- Ping et al. [7], Yuan et al. [8] (EchoBench): medical LLMs/VLMs abandon a correct answer they had committed to when the user pushes back.
- Ahmed et al. [9]: fabricated user documents on radiology VQA, and recognizing a document as fabricated did not prevent adopting it.
- The paper notes that as a follow-up message, fabricated user evidence is less persuasive [7], whereas this paper's metadata is more persuasive (Introduction).
Clinical context interference and modality conflict order
- Lee et al. [10] (MC-CXR): misleading clinical context disrupts correct image-only decisions.
- Lee et al. [11]: swapping the order of conflicting text and image within the same prompt can flip which side the model follows.
Medical imaging quality and robustness evaluation
- Chen et al. [2] (IQAGPT): CT image quality assessment with VLMs and ChatGPT.
- Jiang et al. [3] (Med-VIGIL): evaluating trustworthy medical VLMs under corrupted visual evidence; the authors cite it in their validation protocol recommendation to support verifying inputs outside the model.
Positioning and baseline: the authors state three differences from the above work—the conflict source is pipeline metadata rather than an arguing user; ground truth is computed exactly from scan geometry; and the same evidence is compared in both directions before and after the model commits (fake vs. true header, correct vs. incorrect first answer). The paper has no methodological baseline model; the control is a learning-free pixel-only rule check (comparing the two panels' data range and intensity profile, threshold 3%), used to confirm that defects are visible in the image (Sec. 2).
How does the paper address this problem?
Builds a CT coverage defect testbed with exact geometric ground truth, measures the admission decisions of 9 VLMs under four delivery conditions, and provides a validation protocol.
Overall approach: ground "which to trust when image and metadata conflict" in the coverage-check task, where ground truth can be computed exactly, and systematically compare the same evidence's decisions under different delivery conditions (same prompt, second turn after answering, tool result, header first).
Data and defect construction
- Data source: NLST same-exam reconstruction pairs (NCI Imaging Data Commons, CC BY 4.0); from 262 pairs, sample 60 admissible pairs (same frame of reference, same z coverage, slice thickness ≤2.5 mm), stratified by manufacturer (GE 24, Toshiba 19, Siemens 14, Philips 3), and take the matched mid-coronal plane (Sec. 2).
- Three defect types: for each pair's Series B, apply truncation (remove 10–35% of slices from one end), internal gap (5–20%), equal-length z-shift (5–20%, with A and B each losing one end so slice counts stay equal), with magnitude drawn per pair (fixed seed); together with clean pairs, 240 cases in total.
- Real mismatches: 24 NLST pairs with header-computed z-IoU < 0.95 (not used in the main grid; median 1−IoU 0.13), rendered directly without editing.
Rendering and conditions
- Two views: aligned (both series share the z axis, missing data shown as black) and native (each stacked by slice order with its own z labels, so defects must be inferred from labels and anatomy).
- Five conditions: header only; image only (aligned, native); image + correct header; image + conflicting header (aligned, native).
- Ground truth and prompt: ground truth is computed from per-slice Image Position (Patient); the header is a one-line summary derived by the pipeline, claiming completeness for defect cases and truncation for clean cases; the prompt states the admission rule and requires a single line DECISION: ADMIT/REFUSE plus a reason.
Models and metrics
- 9 VLMs: Llama-3.2-11B/90B-Vision, Gemma-4-31B, Nemotron-3.5-VL-120B, Muse-Glimmer-30B via NVIDIA API (temperature 0); MedGemma-1.5-4B, Qwen3-VL-8B open weights with greedy decoding (one A10G); Claude Haiku 5.5, Sonnet 5.5 via Anthropic API (default sampling, one sample per case). The Gemma two-turn arm and the Llama-90B native arm cover only 15 patients due to API throughput.
- Metrics: SFR (proportion of defect cases admitted) and FRR (proportion of admissible cases refused) reported separately, with patient-clustered bootstrap 95% CI (2,000 resamples); Youden's J = 1 − SFR − FRR; order comparisons use only defects that the model refused from the image alone and for which both orders have parseable answers, and comparisons are limited to within models.
Two turns, mitigation and control
- Two-turn design: turn 1 is the model's own recorded image-only answer; turn 2 delivers the fake header, the true header, or only "Please confirm your final decision" as a user message. Header-first (Claude only): turn 1 gives only the fake header, turn 2 gives the image.
- Mitigation: add a sentence before the header stating that the header summary may be wrong or corrupted, that the image should be trusted, and to trust the image on conflict; tested in both the same-prompt and second-turn positions.
- Pixel-only reference: a learning-free rule compares the two panels' row-wise data range and intensity profile and refuses if the difference exceeds 3% (threshold fixed in advance), verifying that defects are visible in the image the model receives; the authors explicitly do not propose the VLM as a coverage checker (a rule check suffices, AUC 0.98–0.99), and coverage is only a testbed.
What experiments did the paper conduct?
A fake header delivered after answering overturned most correct refusals (Haiku 91%, Nemotron 100%), while it was mostly resisted when delivered in the same prompt; a one-sentence warning can substantially mitigate this.
Experimental setup
- Models tested: Nemotron-3.5-VL-120B, Claude Haiku 5.5, Claude Sonnet 5.5, Gemma-4-31B, Llama-3.2-11B/90B-Vision, MedGemma-1.5-4B, Muse-Glimmer-30B, Qwen3-VL-8B, 9 in total.
- Data: 60 pairs of NLST CT used to construct 240 cases (three types of synthetic defects + clean pairs); 24 real mismatch pairs; the Gemma two-round arm and the Llama-90B native arm have only 15 patients.
- Metrics: SFR, FRR (patient-clustered bootstrap 95% CI, 2,000 resamples), Youden's J; order comparisons are within-model paired differences.
- Review method: rule-based parsing of the DECISION line, no LLM review; ground truth comes from scan geometry.
- Controls: pixel-only rule check; neutral "please confirm" and a bare claim with no values as control conditions.
- Repetition: Claude one sample per case; each arm's protocol was written before the first run, with prompts, parser, and analysis frozen.
Main results
Table 1: among defects the model detected from the image (refused), the share (%) subsequently admitted after being overturned by a fake "complete" header. FRR is the image-only false refusal rate; Tog.=header and image with the same prompt; After=header delivered after answering; +Warn=After plus a warning sentence; Bare=claim with no values; Tool=header delivered as a tool result; Hdr 1st=header first, image after (Claude only; the share of wrong answers caused by the image overturning the header).
Swipe horizontally to read the table
Model View FRR Tog. After +Warn Bare Tool Hdr 1st Nemotron A 17 43 100 1 44 61 – Nemotron N 4 97 100 0 30 99 – Haiku A 55 2 91 1 1 52 100 Haiku N 38 35 99 9 5 96 99 Sonnet A 33 0 21 0 0 71 98 Sonnet N 2 1 74 0 1 97 100 (According to Table 1; Gemma and Llama-90B are not included in this table because they have only 15 patients; the main text gives Gemma aligned 11% vs 64%.)
- The authors report paired differences for the aligned view: Haiku 2% [1, 4] vs 91% [86, 95] (+88 percentage points [83, 93]), Nemotron 43% vs 100% (+57 [53, 62]), Sonnet 0/178 vs 21% [16, 26] (Sec. 3).
- The authors state that with the same prompt, "refuse if either source objects": the fake "truncated" header made the three models refuse all 120 clean views; when the order was reversed, the image overturned 98–100% of the wrong answers caused by the fake header (Claude models).
- The authors state that this is not generalized sycophancy and that the behavior is consistent with "evidence replacement", but emphasize that what is measured is behavior, not mechanism; plain recency cannot fully explain it, because under the same prompt the same models behave conservatively.
Bidirectionality and control experiments
- Bidirectional overturning: a fake "truncated" header delivered after answering overturned all correct admissions of clean pairs (Nemotron 92/92, Haiku 64/64, Sonnet 99/99); a real header overturned almost all incorrect refusals (Nemotron 11/11, Haiku 56/56, Sonnet 20/21).
- Neutral follow-up: "Please confirm" overturned 0/1,072 correct refusals and 2/253 correct admissions.
- Bare claim and tool result: a claim with no values overturned only a minority of Claude detections (Table 1 Bare column 0–5%); the authors state that the header's values carry the effect; a header delivered as a tool result still overturned 52–99%.
Mitigation and real mismatches
- One warning sentence: overturning under the same prompt dropped to ≤2% (Nemotron, Haiku, Sonnet; Gemma 5–8%), at the cost of increased false refusals under a correct header (Nemotron 0%→12%); in the second round overturning dropped to ≤9%.
- Exception: the warning did not prevent Haiku from following a fake "truncated" header on clean pairs (95–96%).
- Real mismatches (24 pairs): among mismatches the model refused from the image, the fake header in the second round overturned far more than with the same prompt (aligned: Haiku 21/23 vs 5/23, Sonnet 10/21 vs 0/21, Nemotron 23/23 vs 19/23).
- Pixel-only check: missed 1 of 360 synthetic defect views and 0 of 48 real defect views (AUC 0.98–0.99); robustness checks changed the rates by ≤3.3 percentage points.
Other ablations and analyses
- Image-only detection ability varies: SFR ≤8% for Nemotron, Gemma, Haiku, Sonnet; Llama-3.2-11B and MedGemma-1.5 do not distinguish defects (SFR ≥96%); Muse-Glimmer-30B and Qwen3-VL-8B refuse almost all admissible pairs (Sec. 3).
- Restricting to the subset of defects admitted on clean pairs, the order pattern is the same (Sec. 3).
- The order analysis uses only the 4 models whose image-only decisions can distinguish defects (Nemotron, Gemma, Haiku, Sonnet) (Sec. 2).
What are some points that can be further explored?
The authors list limitations including single coronal plane, synthetic header, header-first tested only on Claude, and no GPT/Gemini; the experiments cover 9 VLMs and two types of cases.
Limitations noted by the authors and future directions
- Only one coronal plane per series is taken rather than the full volume (Sec. 4 Limitations).
- The z-shift in the native view can only be seen through the z labels; that part of the conflict is text-to-text (Sec. 4).
- The header is a synthetic one-line summary; the original fields were tested on only one model (Sec. 4).
- The header-first condition was tested only on Claude models (Sec. 4).
- GPT and Gemini models were not included (Sec. 4).
- The authors propose a validation protocol: on cases with known ground truth, separately test image-only, conflicting metadata in the same prompt, conflicting metadata after answering, and true metadata after a wrong answer plus a neutral follow-up, reporting SFR and FRR separately; a check that passes (1) but fails (2)(3) should not have access to metadata, or must be told that the metadata may be wrong; verify inputs outside the model where possible (Sec. 4, an author recommendation rather than a completed experiment).
Experimental coverage
- The models tested are 9 VLMs; the order analysis uses only the 4 of them whose image-only judgments could distinguish defects (Sec. 2).
- The data are 60 pairs of NLST synthetic defect cases (240 cases) plus 24 pairs of real mismatches; the Gemma two-turn arm and the Llama-90B native arm cover only 15 patients (Sec. 2).
- Delivery conditions covered: same prompt, second-turn user message after answering, tool result, header-first (Claude only), warning-sentence mitigation (Sec. 2).
- Statistics are patient-clustered bootstrap 95% CI (2,000 resamples); Claude is sampled once per case; the paper does not report the number of repeated samples for most models (Sec. 2).
- The paper states that decoding settings vary by provider and that comparisons are limited to within models (Sec. 2).
Summarize the main content of the paper
The same evidence delivered in different ways flips the VLM's CT admission decision: the later-arriving source wins, a single warning sentence can mitigate it, and the authors provide a validation protocol.
A study of VLM behavior in clinical input quality control: when CT images conflict with DICOM metadata, which one the model believes depends on how the evidence is delivered rather than on the evidence itself.
- Problem: a truncated CT series admitted as complete into lung cancer screening volume follow-up would bias measurements; image-only evaluation misses such failures, and in agentic pipelines metadata often arrives as a tool result after the visual judgment.
- Method: from 60 pairs of NLST CT, construct 240 defect cases with exact geometric ground truth (truncation, internal gap, z-shift), and have 9 VLMs make ADMIT/REFUSE decisions under conditions including image-only, image + fake header in the same prompt, and fake header delivered as a second turn after answering; plus 24 pairs of real mismatches.
- Main results: among defects the model refused from the image, a fake "complete" header rarely overturns when in the same prompt (Haiku aligned 2%), but overturns 91% (Haiku aligned) to 100% (Nemotron both views) when delivered after answering; when the order is reversed, the image overturns 98–100% of the errors caused by the fake header—the later source wins, and this holds in both directions (Table 1, Sec. 3).
- Control and mitigation: a neutral "please confirm" overturns 0/1,072 correct refusals, which the authors say is not generalized sycophancy; the header as a tool result still overturns 52–99%; a one-sentence warning to "trust the image on conflict" reduces second-turn overturning to ≤9%, but did not prevent Haiku from following a fake "truncated" header on clean pairs (95–96%).
- Authors' conclusion: VLMs that pass the image-only test may still abandon a detected defect when metadata arrives later, with the clinical risk that incomplete series enter longitudinal volume measurement unmarked; the authors recommend validating input checks with a four-condition protocol and verifying inputs outside the model where possible.