Evaluating Cross-Signal Coherence in Mammography Vision-Language Models
Authors: Rebeca Nam, Chase M. Fensore, Christian A. Fauria-Robinson, Judy Gichoya, Joyce C. Ho
Organizations: Yonsei University, 50 Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea. · Department of Computer Science, Emory University, 400 Dowman Drive, Atlanta, 30322, GA, USA. · Department of Radiology and Imaging Sciences, Emory University, 100 Woodruff Circle, Atlanta, 30322, GA, USA.
As vision-language models (VLMs) move toward interpreting medical images directly, their outputs increasingly include not just a label but a stated confidence and a written justification. Yet the field still predominantly evaluates VLMs using just accuracy. We evaluate the coherence between answer, confidence, and explanation across four VLMs on breast density and BI-RADS categorization, using two mammography datasets. We perform controlled perturbations of the question, multiple choice option orders, and the images and assess changes in the three outputs. Our results suggest confidence does not decrease even as the mammography image is changed. Shuffling the options also changes most VLMs' answers. Notably, there are cases where a model's own stated reasoning contradicts its final answer, a distinction accuracy alone cannot capture but one that governs whether an AI-generated answer can be trusted.
Figures & tables
Figure 1: Our proposed evaluation methodology with six conditions: the original image and five perturbations.
Figure 2: Prompt used to jointly elicit an explicit answer, per-option confidence scores, and reasoning in a single response.
Model
Output acc. %
NOTA sel. %
RSNA – BI-RADS (6 true classes; majority 59.4%)
MedGemma-4B
42.4
0.0
Lingshu-7B
12.0
0.0
Qwen2-VL-7B
2.9
62.8
InternVL3-8B
7.7
13.7
RSNA – Density (ACR) (4 true classes; majority 42.6%)
Table 1: Per-task output accuracy and “None of the above” selection rate, with the majority-class prior. The answer is the model’s explicitly stated single-best choice.
Figure 3: Answer × reasoning coherence on RSNA. Each item is classified by whether the explicit answer y^=y and the reasoning answer y^e=y : A (both correct), B (answer correct, reasoning wrong), C (answer wrong, reasoning correct; the disconnect ), D (both wrong).
Figure 4: Two models on the same RSNA mammogram (ground truth BI-RADS 0). Each model’s stated answer agrees with the category implied by its explanation, yet the correct answer (InternVL, BI-RADS 0) rests on a self-contradictory rationale—it calls the exam “normal” (a BI-RADS 1 finding) while assigning BI-RADS 0—whereas the incorrect answer (MedGemma, BI-RADS 1) follows validly from the same normal reading.
Figure 5: Mean confidence on the condition-correct answer, cy (solid) the original category on real-image conditions, NOTA on the visual perturbations vs. on the explicitly chosen answer, cy^ (dashed), across the six RSNA conditions.
Model
Real %
Visual %
Δ (pp)
BI-RADS
MedGemma-4B
0.0
0.0
+0.0
Lingshu-7B
0.0
20.6
+20.6
Qwen2-VL-7B
62.8
91.7
+28.9
InternVL3-8B
13.7
91.0
+77.3
Density (ACR)
Table 2: RSNA “None of the above” selection rate on the real image vs. its average over the three visual perturbations ( unrelated / blank / no image ), where NOTA is the correct response, with Δ= Visual − Real.
Figure 6: Perturbed-Options condition effect on RSNA. Bars = share of items whose stated answer changes when the options are reordered, split by effect on accuracy.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Hugging Face checkpoint
Domain
Params
MedGemma-4B
google/medgemma-4b-it
Medical
4B
Lingshu-7B
lingshu-medical-mllm/Lingshu-7B
Medical
7B
Qwen2-VL-7B
Qwen/Qwen2-VL-7B-Instruct
General
7B
InternVL3-8B
OpenGVLab/InternVL3-8B
General
8B
Appendix
Table 3: Model configurations. All models use the default (latest) Hugging Face checkpoint revision, bfloat16 precision (no quantization), greedy decoding (temperature 0 ), and a 384-token generation budget.
Cohort
Task
N
Majority-class prior
RSNA
BI-RADS
5,106
59.4%
RSNA
Density
5,821
42.6%
EMBED
BI-RADS
14,303
62.1%
Appendix
Table 4: Cohort and task composition. Majority-class prior is the accuracy of always predicting the most frequent label.
Model
Task
Parsed %
Answer acc. %
Reasoning acc. %
Δ (pp)
MedGemma-4B
BI-RADS
99.8
42.4
35.8
−6.7
Density
94.1
33.8
33.5
−0.3
Lingshu-7B
BI-RADS
97.6
12.3
16.6
+4.3
Density
92.2
48.4
48.3
+0.0
Qwen2-VL-7B
BI-RADS
66.3
3.7
7.1
+3.4
Density
75.9
32.5
32.9
+0.5
Appendix
Table 5: Explicit answer vs. reasoning accuracy on the parsable fraction of RSNA results. Both accuracies are computed only on items whose explanation yields a readable category, so the comparison within each row is paired. Parse rate varies across model × task cells, from 66% (Qwen2-VL-7B, BI-RADS) to ∼100% (MedGemma-4B, BI-RADS); low-parse cells rest on a correspondingly smaller subset. The gap Δ=(reasoning−answer) is small and mixed in sign ( −6.7 to +4.3 ): the free-text rationale is no more accurate than the explicitly stated answer.
Figure 7: Example explanations for the four reasoning-parser outcomes: BI-RADS label, ACR-density label, “None of the above,” and unparsable. Highlighted text is the phrase the parser matches.
Comparison
Agreement %
κ
Matcher vs. human
80.8
0.777
Judge vs. human
82.5
0.798
Matcher vs. judge
78.3
0.753
Appendix
Table 6: Reasoning-extraction agreement on 120 hand-labeled RSNA explanations.
Model
Task
Agreement %
MedGemma-4B
BI-RADS
93.5
Density
74.5
Lingshu-7B
BI-RADS
58.0
Density
66.7
Qwen2-VL-7B
BI-RADS
84.9
Density
59.1
Appendix
Table 7: Matcher vs. LLM-judge agreement at scale ( 1,000 explanations per task, judge blinded). Overall ≈70.6% .
Figure 8: NOTA-selection rate (left) and mean confidence when NOTA is chosen (right), for BI-RADS (top) and Density/ACR (bottom), across all six conditions. Shaded band marks the three conditions where NOTA is the correct response.
Figure 9: Effect of rewording the question, on items whose wording actually changed. Each bar is the share of items whose answer changed, split by effect on accuracy. The cross-hatched portions are the share that changed to a NOTA. Numbers above the bars give the net accuracy change (percentage points).
Model
Task
Pert.
Orig.%
Pert.%
C → W
W → C
Net Δ
MedGemma-4B
BI-RADS
Question
46.0
45.6
88
75
−0.3
BI-RADS
Options
42.5
17.7
1685
431
−24.8 *
Density
Question
31.5
26.0
544
303
−5.4 *
Density
Options
33.9
21.7
1406
697
−12.2 *
Lingshu-7B
BI-RADS
Question
10.5
10.0
52
36
−0.4
BI-RADS
Options
14.3
5.5
433
68
−8.9 *
Appendix
Table 8: McNemar test results for the two elicitation perturbations (reworded question, shuffled options), per model and task, on the explicitly stated answer. Density refers to ACR breast density. C → W and W → C are item counts; Net Δ is the percentage-point change in accuracy. *: p<0.05 .
Model
Parsed %
Answer acc. %
Reasoning acc. %
Δ (pp)
MedGemma-4B
96.3
36.8
20.9
−15.9
Lingshu-7B
99.8
10.3
57.6
+47.3
Qwen2-VL-7B
91.4
15.4
2.3
−13.1
InternVL3-8B
82.4
4.8
4.7
−0.1
Appendix
Table 9: Explicit answer vs. reasoning accuracy on the parsable fraction of EMBED BI-RADS, both computed on items with parseable reasoning (paired). Parsed % is relative to the ∼14.3 k item set; Δ= reasoning − answer.
Figure 11: Mean confidence on the condition-correct answer cy (solid; original category on real images, NOTA on the visual perturbations) vs. the explicitly chosen answer cy^ (dashed), across the six EMBED (BI-RADS) conditions.
Figure 12: NOTA-selection rate on the real image vs. junk images (EMBED, BI-RADS). Evidence-driven NOTA should rise under junk; most models do not.
Figure 13: Perturbation sensitivity (EMBED, BI-RADS): question wording (left) vs. option shuffling + fillers (right). Option shuffling flips most answers without raising accuracy.
Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.
Vision-language models (VLMs) could support medical-image interpretation, but their use requires deciding which interpretations to accept and which images to refer to a clinician. We test whether thresholding stated confidence can retain useful coverage while limiting errors among accepted answers, and how this depends on the population. Eight open-weight VLMs, including two reasoning models, assess 3,888 brain MRI slices from post-treatment glioma patients (BraTS 2024) and healthy volunteers (IXI) for tumor or gross abnormality. BraTS labels are mask-derived; IXI slices are treated as negatives. On BraTS, no threshold selects a nonempty set with at most 5% empirical error for any model, even when chosen retrospectively. This also holds for four models evaluated on a large-lesion, distant-negative subset. At a 10% error budget, Gemma-4-12B retains 24.4% of slices and every other model at most 2.5%. To investigate this limited selection ability, we separate confidence differences between predicted answers from differences within each answer. For four models with available outcome-level summaries, pooled mean confidence differs more between answers than between correct and incorrect responses within each answer. A covariance decomposition explains how this pattern can produce opposite confidence-correctness associations across populations. At a 20% error budget, accepting every "tumor" answer covers more slices than any confidence threshold for two models. These findings caution against using unvalidated stated confidence as a standalone deferral signal and support population-specific validation before deployment. Code is available at https://github.com/amir-sbg/VLM-Answer-or-Defer.
Amir Sabbaghziarani, Parsa Ghazvinian, Sepehr Salem +1
Joint Georgia Institute of Technology / Georgia State University / Emory University Center for Translational Research in Data Science and Neuroimaging · Georgia State University
Vision-language models (VLMs), including medical specialists, are increasingly proposed for medical imaging, yet their stated confidence is rarely evaluated separately from correctness. We use brain MRI as a controlled, high-stakes testbed for a broader failure mode in frontier multimodal systems: models can appear competent while lacking reliable self-knowledge. We present an automatically graded behavioral audit and pilot study of six instruction-tuned VLMs (five general-purpose and one medical specialist) on 4,102 images (4,032 axial/coronal/sagittal MRI slices from 250 subjects plus 70 non-brain/noise controls), with labels derived from public metadata and released expert segmentation masks rather than new human annotation. Across models, answer coverage is near-complete, but verbalized-confidence calibration is poor: ECE ranges from 0.27 to 0.40, mean confidence on incorrect answers ranges from 0.82 to 0.97, and 33-46% of answered items are high-confidence errors. The most accurate model is also the most confident on its errors, while a base/specialist family contrast suggests that medical adaptation improves tumor-presence detection without improving confidence reliability. Open-ended diagnostics further show that hallucination and abstention vary separately from multiple-choice accuracy. These findings argue that medical-image VLM evaluation should report verbalized-confidence reliability, confident error, hallucination, and abstention alongside accuracy.
Amir Sabbaghziarani, Mohammadsajad Abavisani, Sergey Plis
Attention and saliency heatmaps are widely used to explain medical Vision-Language Model (VLM) outputs on chest X-rays, yet whether they truly highlight the image evidence driving predictions has not been causally tested. We audit faithfulness via overlap with radiologist bounding boxes on PadChest (n=637), attribution mass within radiologist masks on CheXlocalize (n=643), and 16x16 patch-occlusion maps that record which regions, when hidden, change the answer. We study three MedGemma-4B variants, cross-family probes on LLaVA-RAD and Qwen3-VL-8B-Instruct, and the specialist CheXagent-2-3b, with two CXR-trained classifiers (DenseNet121, ResNet50) as positive controls. A heatmap is faithful only if the model uses the image and attention concentrates on regions whose occlusion alters the prediction. No evaluated VLM meets both criteria. MedGemma and Qwen3-VL use the image, but attention anti-correlates with patch-occlusion importance (rho < 0 with 95% bootstrap CIs below zero). LLaVA-RAD's attention correlates positively, but the model is almost text-only (99.1% text-only agreement, near-zero causal mass), so correlation ties two near-zero signals. Attention also misses annotated anatomy: overlap with true regions never beats shifted or random controls, and no method places more than 22% of its mass inside radiologist masks. The two CXR classifiers pass all metrics, indicating the failure is specific to VLM heatmaps, not the evaluation. These heatmaps are visually reassuring but not faithful; clinical explanations require controlled localization metrics and causal perturbation, not visual inspection alone.
Binesh Sadanandan, Vahid Behzadan
SAIL Lab, University of New Haven, West Haven, CT, USA