Beyond Benchmark Scores: Auditing Medical Vision-Language Models for Chest X-Ray Tuberculosis Screening

Mushir Akhtar; M. Tanveer; Mohd. Arshad

cs.CVSep 18, 2026

Beyond Benchmark Scores: Auditing Medical Vision-Language Models for Chest X-Ray Tuberculosis Screening

Authors: Mushir Akhtar, M. Tanveer, Mohd. Arshad

Organizations: Department of Mathematics, Indian Institute of Technology Indore, India

Abstract

A medical model's benchmark score does not establish that the same conclusion holds under a different evaluation. This study tests whether claims about model ranking, score reliability and screening performance survive changes in cohort, prompt, negative spectrum, specified prevalence and operating threshold. We audit three medical vision-language models (BioMedCLIP, CheXficient, and MedSigLIP) and a general-domain OpenCLIP comparator on 12,200 chest radiograph records from four datasets (Montgomery, Shenzhen, TBX11K, and VinDr-CXR). Five fixed prompt families yield 244,000 model--image--prompt scores. No model leads every cohort and reliability criterion. Prompt-family changes alter AUROC in 21 of 48 multiplicity-controlled comparisons. Replacing healthy controls with sick non-tuberculosis controls reduces AUROC by 0.075--0.306 across all four models. On VinDr-CXR, the three medical models distinguish tuberculosis from no-finding controls substantially better than from pneumonia or lung tumor; their AUROC point estimates for both named diseases fall below 0.5. CheXficient has documented VinDr-CXR pretraining exposure, which limits the interpretation of its results. Thresholds chosen for 95% sensitivity on TBX11K training retain that constraint by point estimate in only four of sixteen target evaluations. A five-seed supervised source model reaches 0.999 AUROC on TBX11K validation but 0.629 on each of two external cohorts. Conservative exclusion of perceptual-overlap candidates narrows this gap without closing it. These retrospective, single-task results show that discrimination, score reliability and threshold retention support different portability claims. Evidence for chest X-ray tuberculosis screening should identify the complete evaluation specification rather than attribute clinical portability to a checkpoint alone.

Figures & tables

Cohort	Non-TB composition	TB	Non-TB	TB (%)	Role
Montgomery	Normal	58	80	42.0	External comparison
Shenzhen	Normal	336	326	50.8	External comparison
TBX11K validation	800 healthy; 800 sick	200	1,600	11.1	Primary benchmark
TBX11K training	3,000 healthy; 3,000 sick	600	6,000	9.1	Source and replication
VinDr-CXR test	No finding; other findings	164	2,836	5.5	Named-spectrum extension

Table 1: Cohort composition and experimental role.

Model	Checkpoint and representation	Pretraining context and exposure qualification
OpenCLIP	ViT-B/32; laion2b_s34b_b79k ; $224\times 224$ input	General image–text training on LAION-2B. No documented direct dataset exposure in the examined documentation; web-image overlap remains unresolved [ 18 ] .
BioMedCLIP	microsoft/BiomedCLIP- PubMedBERT_256-vit_base_ patch16_224 ; $224\times 224$ input	Biomedical figure–caption training on PMC-15M. Scientific-figure reuse prevents certification of image-level independence [ 3 ] .
CheXficient	StanfordAIMI/CheXficient ; DINOv2-base and BioClinicalBERT; checkpoint image processor	CXR-specific training. VinDr-CXR is a documented training source; TBX11K is described as unseen in the model report. The VinDr experiment is a within-dataset stress test, not independent external validation [ 19 ] .
MedSigLIP	google/medsiglip-448 ; medical SigLIP image and text encoders; $448\times 448$ input	Medical training across specialties, including public and non-public sources. No documented direct exposure to the evaluated datasets; absence of exposure is not established [ 20 , 21 ] .

Table 2: Vision-language checkpoints and the interpretation of their training provenance.

Family	TB text	Non-TB text
Clinical	a chest x-ray showing pulmonary tuberculosis a chest radiograph with tuberculosis a chest x-ray with findings compatible with tuberculosis a frontal chest radiograph showing active tuberculosis	a normal chest x-ray without tuberculosis a chest radiograph with no evidence of tuberculosis a chest x-ray without pulmonary tuberculosis a frontal chest radiograph without signs of tuberculosis
Photo	this is a photo of a chest X-ray showing pulmonary tuberculosis	this is a photo of a chest X-ray without pulmonary tuberculosis
Report	chest radiograph findings are compatible with pulmonary tuberculosis	no radiographic evidence of pulmonary tuberculosis
Presence	pulmonary tuberculosis is present on this chest radiograph	pulmonary tuberculosis is absent on this chest radiograph
Normal	a chest radiograph showing pulmonary tuberculosis	a normal chest radiograph

Table 3: Exact text supplied to each frozen text encoder.

Panel of Figure 1 — Figure 1: Primary clinical-prompt performance with 95% image-level bootstrap intervals. Each model retains its checkpoint, processor and prompt rule across cohorts. AUPRC changes with class balance as well as score ordering. Intervals use ordinary image resampling in the original cohorts and within-label resampling in VinDr-CXR. ∗ The exposure marker applies to CheXficient on VinDr-CXR only: that cohort is a documented pretraining source.

Cohort	Model	AUROC [95% CI]	AUPRC	Brier	ECE 15	AURC
Montgomery	CheXficient	0.969 [0.925, 0.998]	0.976	0.133	0.183	0.081
Montgomery	MedSigLIP	0.976 [0.945, 0.998]	0.979	0.101	0.181	0.039
Montgomery	BioMedCLIP	0.890 [0.826, 0.947]	0.905	0.173	0.200	0.108
Montgomery	OpenCLIP	0.610 [0.515, 0.706]	0.522	0.293	0.232	0.501
Shenzhen	CheXficient	0.950 [0.932, 0.965]	0.959	0.197	0.247	0.120
Shenzhen	MedSigLIP	0.907 [0.884, 0.928]	0.921	0.157	0.148	0.104

Table 4: Primary clinical-prompt results across four evaluation cohorts.

Cohort	Model	Se at 95% Sp	Sp at 95% Se	Risk at 80%	Risk at 90%
Montgomery	CheXficient	0.948	0.938	0.117	0.152
Montgomery	MedSigLIP	0.914	0.787	0.072	0.112
Montgomery	BioMedCLIP	0.741	0.200	0.153	0.184
Montgomery	OpenCLIP	0.138	0.075	0.577	0.568
Shenzhen	CheXficient	0.839	0.644	0.208	0.247
Shenzhen	MedSigLIP	0.670	0.521	0.181	0.206

Table 5: Within-cohort operating points and selective prediction under the clinical prompt.

Panel of Figure 2 — Figure 2: AUROC changes relative to the clinical prompt family. Each cell holds the checkpoint, images and labels fixed. Color scales are shared across panels. ∗ In (a)–(c), the paired DeLong test survives Benjamini–Hochberg correction across 48 comparisons. Panel (d) does not use that test family. † CheXficient has documented VinDr-CXR exposure. Photo, Report, Presence and Normal refer to the exact families in Table 3 .

Cohort	Model	Clinical	Photo	Report	Presence	Normal
Montgomery	CheXficient	0.969	0.978	0.972	0.909	0.947
Montgomery	MedSigLIP	0.976	0.976	0.976	0.940	0.974
Montgomery	BioMedCLIP	0.890	0.873	0.903	0.893	0.903
Montgomery	OpenCLIP	0.610	0.543	0.494	0.514	0.459
Shenzhen	CheXficient	0.950	0.945	0.925	0.927	0.939
Shenzhen	MedSigLIP	0.907	0.892	0.874	0.890	0.878

Table 6: Complete AUROC matrix for the five prompt families.

Cohort	Model	$\Delta$ AUROC [95% CI]	$\Delta$ Brier [95% CI]
Montgomery	CheXficient	0.005 [0.000, 0.014]	-0.017 [-0.024, -0.012]
Montgomery	MedSigLIP	0.003 [-0.002, 0.009]	0.014 [0.011, 0.017]
Montgomery	BioMedCLIP	0.010 [-0.009, 0.032]	0.002 [-0.003, 0.007]
Montgomery	OpenCLIP	-0.076 [-0.152, 0.002]	-0.028 [-0.032, -0.024]
Shenzhen	CheXficient	-0.007 [-0.011, -0.003]	-0.047 [-0.054, -0.041]
Shenzhen	MedSigLIP	-0.014 [-0.018, -0.011]	0.016 [0.015, 0.018]

Table 7: Equal-weight averaging of all five prompt-family scores relative to the clinical prompt.

Component	AUROC (%)	Brier (%)	ECE 15 (%)
Model	76.6	37.7	10.3
Cohort	10.9	3.0	7.0
Prompt	0.8	5.8	5.6
Model $\times$ cohort	5.8	5.9	17.8
Model $\times$ prompt	1.9	24.4	30.0
Cohort $\times$ prompt	0.2	11.9	14.8

Table 8: Descriptive decomposition of the original balanced model–cohort–prompt matrix.

Panel of Figure 3 — Figure 3: Discrimination changes when the negative group changes. Lines connect the same model; error bars show marginal 95% intervals. Panel (a) uses 800 controls per group. Panel (b) uses 2,051 no-finding, 210 pneumonia and 73 lung-tumor controls. The positive sample remains fixed within each panel. The dotted reference marks AUROC 0.5. ∗ CheXficient has documented VinDr-CXR pretraining exposure. Paired uncertainty for the negative-group difference is distinct from these marginal intervals.

Model	Healthy	Sick non-TB	AUROC loss [95% CI]	Holm $p$
CheXficient	0.865	0.725	0.140 [0.117, 0.163]	0.0004
MedSigLIP	0.953	0.647	0.306 [0.271, 0.344]	0.0004
BioMedCLIP	0.872	0.611	0.261 [0.231, 0.292]	0.0004
OpenCLIP	0.602	0.527	0.075 [0.048, 0.103]	0.0004

Table 9: Paired negative-spectrum inference on TBX11K validation.

Model	No finding [95% CI]	Pneumonia [95% CI]	Lung tumor [95% CI]
CheXficient ∗	0.945 [0.924, 0.964]	0.466 [0.404, 0.524]	0.372 [0.302, 0.444]
MedSigLIP	0.918 [0.893, 0.942]	0.331 [0.277, 0.388]	0.380 [0.311, 0.451]
BioMedCLIP	0.780 [0.741, 0.819]	0.334 [0.275, 0.392]	0.306 [0.237, 0.380]
OpenCLIP	0.525 [0.481, 0.571]	0.535 [0.474, 0.590]	0.550 [0.468, 0.628]

Table 10: VinDr-CXR AUROC by named negative group under the clinical prompt.

Model	Pneumonia minus no finding	Lung tumor minus no finding
CheXficient ∗	-0.479 [-0.533, -0.426]	-0.573 [-0.640, -0.505]
MedSigLIP	-0.586 [-0.634, -0.535]	-0.537 [-0.606, -0.470]
BioMedCLIP	-0.446 [-0.492, -0.402]	-0.474 [-0.537, -0.408]
OpenCLIP	0.010 [-0.033, 0.051]	0.024 [-0.048, 0.099]

Table 11: Paired VinDr-CXR AUROC differences relative to no-finding controls.

Panel of Figure 4 — Figure 4: Raw-score reliability for CheXficient (coral circles, solid line) and MedSigLIP (olive squares, dashed line). Ten equal-width bins display mean score against observed TB fraction; larger markers indicate more images within a model–cohort panel. The diagonal denotes agreement. Empty bins are omitted. These plots do not use the 15-bin partition of ECE 15 and do not apply prevalence reweighting.

Panel of Figure 5 — Figure 5: Brier score under specified class weights. Lines connect evaluated prevalences; shaded regions show pointwise 95% within-label bootstrap intervals. The vertical scale is shared. The VinDr panel has a different prevalence grid, explicitly marked on its horizontal axis. Scores and conditional samples remain unchanged. Brier score includes discrimination and calibration effects; a lower value does not by itself establish calibration. ∗ Documented VinDr-CXR exposure for CheXficient.

Cohort	Model	Prevalences (%)	First	Second	Third
Montgomery	CheXficient	10/20/50	0.032	0.064	0.159
Montgomery	MedSigLIP	10/20/50	0.039	0.058	0.116
Montgomery	BioMedCLIP	10/20/50	0.042	0.083	0.205
Montgomery	OpenCLIP	10/20/50	0.390	0.359	0.269
Shenzhen	CheXficient	10/20/50	0.042	0.080	0.194
Shenzhen	MedSigLIP	10/20/50	0.071	0.092	0.156

Table 12: Brier scores standardized to specified TB prevalences.

Cohort	Model	$\pi$	ECE 15 [95% CI]	Log loss [95% CI]
Montgomery	CheXficient	0.50	0.221 [0.180, 0.264]	0.517 [0.372, 0.681]
Montgomery	MedSigLIP	0.50	0.198 [0.157, 0.243]	0.376 [0.313, 0.446]
Montgomery	BioMedCLIP	0.50	0.241 [0.190, 0.291]	0.801 [0.583, 1.049]
Montgomery	OpenCLIP	0.50	0.153 [0.148, 0.193]	0.732 [0.720, 0.745]
Shenzhen	CheXficient	0.50	0.243 [0.226, 0.261]	0.615 [0.555, 0.679]
Shenzhen	MedSigLIP	0.50	0.146 [0.127, 0.171]	0.476 [0.446, 0.508]

Table 13: Additional standardized score-reliability endpoints at each block’s primary prevalence.

Model	Source constraint	Threshold $\tau$	Source Se	Source Sp
CheXficient	95% sensitivity	0.032724	0.950	0.172
CheXficient	95% specificity	0.685756	0.392	0.950
MedSigLIP	95% sensitivity	0.225860	0.950	0.478
MedSigLIP	95% specificity	0.830078	0.272	0.950
BioMedCLIP	95% sensitivity	0.017594	0.950	0.279
BioMedCLIP	95% specificity	0.921753	0.212	0.950

Table 14: Operating thresholds estimated from all 6,600 TBX11K training images.

Panel of Figure 6 — Figure 6: Transport of a threshold chosen for 95% sensitivity on TBX11K training. The dashed line in (a) marks the source constraint. Error bars show 95% intervals. Original-target intervals incorporate source-threshold re-estimation; VinDr intervals condition on the frozen source threshold and quantify target sampling uncertainty only. ∗ CheXficient’s VinDr estimate has documented dataset exposure. Retention counts use point estimates, not confidence bounds.

Target	Model	Sensitivity [95% CI]	Specificity [95% CI]	Retains
Montgomery	CheXficient	0.966 [0.914, 1.000]	0.863 [0.787, 0.963]	Yes
Montgomery	MedSigLIP	0.914 [0.845, 0.983]	0.938 [0.863, 1.000]	No
Montgomery	BioMedCLIP	0.879 [0.776, 0.948]	0.738 [0.613, 0.838]	No
Montgomery	OpenCLIP	0.931 [0.828, 1.000]	0.075 [0.025, 0.188]	No
Shenzhen	CheXficient	0.932 [0.884, 0.958]	0.764 [0.690, 0.853]	No
Shenzhen	MedSigLIP	0.872 [0.830, 0.926]	0.745 [0.641, 0.816]	No

Table 15: Transport of the TBX11K-training threshold chosen for 95% sensitivity.

Target	Model	Threshold	Sensitivity [95% CI]	Specificity
Montgomery	BioMedCLIP	0.00694	0.914 [0.828, 0.983]	0.438
TBX11K val.	BioMedCLIP	0.00694	0.985 [0.965, 1.000]	0.095
Montgomery	CheXficient	0.02426	0.966 [0.914, 1.000]	0.812
TBX11K val.	CheXficient	0.02426	0.955 [0.910, 0.985]	0.072
Montgomery	MedSigLIP	0.17374	0.948 [0.897, 1.000]	0.800
TBX11K val.	MedSigLIP	0.17374	0.975 [0.930, 0.995]	0.311

Table 16: Sensitivity-threshold transport with Shenzhen as the alternative source.

Target	Model	Sensitivity	Specificity	Retains specificity
Montgomery	CheXficient	0.379	1.000	Yes
Montgomery	MedSigLIP	0.241	1.000	Yes
Montgomery	BioMedCLIP	0.293	1.000	Yes
Montgomery	OpenCLIP	0.000	0.988	Yes
Shenzhen	CheXficient	0.369	1.000	Yes
Shenzhen	MedSigLIP	0.190	1.000	Yes

Table 17: Secondary transport of the TBX11K-training threshold chosen for 95% specificity.

Panel of Figure 7 — Figure 7: Supervised transfer and overlap-candidate sensitivity. Panel (a) shows each training seed and the mean-score ensemble. Panel (b) distinguishes seed-paired AUROC changes from the ensemble-level paired image-bootstrap interval. Positive changes favor retraining after exclusion. The ensemble interval conditions on two fitted ensembles and does not incorporate repeated training uncertainty.

Cohort	Seed AUROC mean $\pm$ SD	Ensemble AUROC	AUPRC	Brier	AURC
TBX11K val.	0.997 $\pm$ 0.001	0.999	0.993	0.008	0.000
Shenzhen	0.630 $\pm$ 0.075	0.629	0.691	0.402	0.346
Montgomery	0.615 $\pm$ 0.035	0.629	0.583	0.269	0.323

Table 18: Supervised ResNet-50 source training and unchanged external evaluation.

Cohort	Original	Excluded	Retained	TB retained	Non-TB retained
Montgomery	138	1	137	58	79
Shenzhen	662	109	553	284	269
TBX11K train	6600	166	6434	589	5845
TBX11K val.	1800	37	1763	197	1566

Table 19: Conservative removal of all images implicated by the perceptual candidate screen.

Cohort	Original retained	Retrained	Ensemble $\Delta$ [95% CI]	Seed $\Delta$ mean $\pm$ SD
TBX11K val.	0.9989	0.9991	0.0002 [-0.0003, 0.0008]	-0.001 $\pm$ 0.001
Shenzhen	0.6329	0.7335	0.1006 [0.0732, 0.1282]	0.075 $\pm$ 0.130
Montgomery	0.6266	0.6779	0.0513 [0.0024, 0.1041]	0.033 $\pm$ 0.077

Table 20: Retraining after candidate exclusion, evaluated on identical retained target images.

Claim component	Question to state explicitly	Evidence from this audit
Implementation	Which weights, processor, score map and text define the classifier?	Prompt-family effects occur on identical images; a model name alone is incomplete.
Population and endpoint	Which setting and reference labels support the result?	Cohort ranking changes, and dataset TB labels do not share a verified common microbiological standard.
Negative spectrum	Are controls healthy, broadly abnormal or a named competing disease?	AUROC falls for all models on sick controls; named alternatives expose larger medical-model failures.
Score interpretation	Which prevalence and reliability measure support the claim?	Class weighting changes Brier comparisons without changing score ordering.
Operating rule	Does an unchanged source threshold retain sensitivity and acceptable specificity?	Only four of sixteen sensitivity constraints survive by point estimate; retention can coexist with near-zero specificity.
Uncertainty and provenance	What is resampled, what remains fixed, and what exposure is known?	Image-level intervals, seed variability and documented pretraining exposure support different inferences.

Table 21: Information needed to interpret a portability claim in this task.

Medical VLMs

Medical vision-language models connect clinical images with text for diagnosis, reporting, question answering, and image grounding.

See all 475 papers

Jun 16, 2026

Vision-language models for chest radiography do not always need the image

Vision-language models that answer questions about chest radiographs are evaluated by their accuracy on labels derived from radiology reports. High benchmark accuracy is often interpreted as evidence that the model uses the image. A model that answers from the finding named in the question can score as well as a model that uses the radiograph. Keeping the question fixed, we audit eight open-weight systems by swapping in another patient's radiograph with the same or the opposite label, occluding the radiologist-marked region or an equal region elsewhere, and removing the radiograph or replacing it with noise or a photograph. On 2,548 yes-or-no questions from MIMIC-CXR, one multimodal model answers Yes regardless of the image, another multimodal model changes its answers without following the label, and four systems use the image but keep about half of their correct answers when the radiograph is swapped for an opposite-label radiograph. A medical model that receives only the question text scores 55.3% on the pooled questions, higher than two multimodal systems. It scores 91.8% where every finding is present, and answering Yes to every question scores 100% there. Where the image is necessary, the best multimodal system exceeds this model by 10.4% in balanced accuracy. The categories are unchanged on CheXpert. Confidence is not higher when a correct answer depends on the marked region. In a reader study with three radiologists, the two radiologists who read a balanced set of 200 cases score 86.0% and 82.0%, and the systems score 50.0% to 73.0%. Accuracy does not establish image use, but an intervention on the image can test it.

Mahshad Lotfinia, Sebastian Ziegelmayer, Lisa Adams +4

Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany · Department of Diagnostic and Interventional Radiology, TUM University Clinic, School of Medicine and Health, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany · Lab for AI in Medicine, RWTH Aachen University, Aachen, Germany +1

VLM Evaluation Medical VQA Shortcut Learning

Ask PDF

Aug 1, 2026

Auditing Medical Vision-Language Models on Chest Radiographs: Estimating Reference Agreement Across Institutions

Vision-language models return structured chest-radiograph findings through interfaces exposing no confidence score, so a receiving institution cannot read off how far to trust an individual judgment. Whether agreement with an institution's reference standard transfers across sites, findings, prediction directions and question formats is largely unmeasured. We evaluated three generative vision-language models on three institutional chest-radiograph corpora and six findings under two elicitation protocols, comprising more than 345,000 finding-level predictions, and estimated finding-by-direction reference agreement at a receiving institution from a small budget of local labels. Estimation strategies were then stress-tested under repeated strict institution-held-out evaluation. Under evaluation excluding the receiving institution from development entirely, adaptive selection among the seven estimators that design admits did not improve on simple fixed alternatives: it achieved a mean Brier score of 0.1083, against 0.0853 for always using a Beta-Binomial empirical-Bayes estimator and 0.0855 for a target-only logistic model. Those two differ by 0.0003, less than this family's own sensitivity to a change of solver version, and each leads in about half the settings, so no default can be recommended. Their advantage over estimators pooling across institutions was concentrated at one site and not confirmatory once clustered by institution, and a plug-in empirical-Bayes posterior-predictive count interval at a nominal 95% level covered 87.0%, less at the hardest institution. Reference agreement therefore has to be re-evaluated per site and per interface; these results concern agreement with institutional labels, not clinical correctness.

Pengyang Yu, Yiou Wang, Zhongping Dong +3

School of Computer Science, University College Dublin, Dublin, Ireland · Department of Medical Imaging, The Third Affiliated Hospital of Southern Medical University, Guangzhou, China · Dublin City University, Dublin, Ireland

Chest X-Ray Classification Medical VLMs Language Model Auditing

Ask PDF

Sep 24, 2025

Revisiting Performance Claims for Chest X-Ray Models Using Clinical Context

Public datasets of Chest X-Rays (CXRs) have long been a popular benchmark for developing machine learning (ML) computer vision models in healthcare. However, the reported strong average-case performance of these models do not necessarily reflect their actual utility when used in heterogeneous clinical settings, potentially masking weaker performance in medically significant scenarios. In this work we use clinical context to provide a more holistic evaluation of models for CXR diagnosis. In particular, we use discharge summaries, recorded prior to each CXR, to derive a ``pre-CXR'' probability of each CXR label, as a proxy for existing contextual knowledge available to clinicians when interpreting CXRs. We use this measure to probe model performance along two dimensions: First, using a stratified analysis, we show that models tend to have lower performance (as measured by AUROC and other metrics) among individuals with higher pre-CXR probability. Second, by controlling for pre-CXR probability via matching and re-weighting, we demonstrate that performance degrades when the correlation is broken between prior context and the current CXR label, suggesting that model performance is highly sensitive to the underlying distribution of clinical context. Specifically, cases with high pre-test probabilities present a fundamentally more difficult visual classification task, highlighting a gap in clinical utility when models are applied to high-risk cohorts.

Andrew Wang, Jiashuo Zhang, Michael Oberst

Brown University · Johns Hopkins University

Chest X-Ray Classification Distribution Shift Healthcare Medical Image Benchmarks

Ask PDF

Abstract

Medical VLMs

Explore similar work

Vision-language models for chest radiography do not always need the image

Auditing Medical Vision-Language Models on Chest Radiographs: Estimating Reference Agreement Across Institutions

Revisiting Performance Claims for Chest X-Ray Models Using Clinical Context