Radiotherapy contour ranking can support quality assurance, but vision-language model (VLM) gains may reflect supervision, score construction, available inputs, or representation access. We evaluated frozen MedGemma 1.5 in a single-institution study, a 12-patient held-out HaN-Seg test, five-fold cross-fitting on 30 non-test HaN-Seg patients, and transfer without refitting to 120 SegRap2023 patients. On held-out HaN-Seg, a supervised readout of the final prompt state exceeded a supervised readout of the model's yes/no output probabilities by +0.3047 pooled average precision (AP; 95% patient-bootstrap interval [+0.2822, +0.3269]); the advantage persisted against a parameter-count-matched nonlinear control. A readout of projected visual tokens exceeded the final-state readout in cross-fitting and transfer (+0.0706 AP averaged equally across patients on SegRap; 95% interval [+0.0607, +0.0804]), while a structured panel-feature readout remained competitive with visual tokens. The anatomy-conditioned readout did not demonstrate an advantage over the native-feature comparator. These results motivate visual-token and structured-feature controls in medical-imaging VLM evaluations; outcomes were computational or synthetic, not clinical acceptability.
Figures & tables
Figure 1: Scoring interfaces. (A) Institutional paths compare reference-assisted T0/D1/D2 with native-input G0; (B) public paths compare output-space, final-state, visual-token, and structured-panel readouts. T0/P0 are unfitted; other heads are task-trained. P0/P1/P2/P3 use the 12-patient test; later controls use 30 non-test patients; separate source fits transfer to 120 SegRap patients.
Table 2
Interface or contrast
HaN-Seg cross-fit
SegRap zero-refit transfer
Full population (primary)
Panel-identifiable sensitivity
Absolute patient-macro AP
P1-MLP: nonlinear output-space
0.3630
0.3708
0.4074
P2: final prompt state
0.6339
0.5877
0.7266
P4: pooled projected visual tokens
0.7229
0.6583
0.8171
P3+-MLP: all-panel structured features
0.7896
0.6702
0.8290
Table 3: Patient-macro AP and paired contrasts in non-test HaN-Seg cross-fitting and zero-refit SegRap transfer.
Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.
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
We introduce MedGemma 1.5 4B, the latest model in the MedGemma collection. MedGemma 1.5 expands on MedGemma 1 by integrating additional capabilities: high-dimensional medical imaging (CT/MRI volumes and histopathology whole slide images), anatomical localization via bounding boxes, multi-timepoint chest X-ray analysis, and improved medical document understanding (lab reports, electronic health records). We detail the innovations required to enable these modalities within a single architecture, including new training data, long-context 3D volume slicing, and whole-slide pathology sampling. Compared to MedGemma 1 4B, MedGemma 1.5 4B demonstrates significant gains in these new areas, improving 3D MRI condition classification accuracy by 11% and 3D CT condition classification by 3% (absolute improvements). In whole slide pathology imaging, MedGemma 1.5 4B achieves a 47% macro F1 gain. Additionally, it improves anatomical localization with a 35% increase in Intersection over Union on chest X-rays and achieves a 4% macro accuracy for longitudinal (multi-timepoint) chest x-ray analysis. Beyond its improved multimodal performance over MedGemma 1, MedGemma 1.5 improves on text-based clinical knowledge and reasoning, improving by 5% on MedQA accuracy and 22% on EHRQA accuracy. It also achieves an average of 18% macro F1 on 4 different lab report information extraction datasets (EHR Datasets 2, 3, 4, and Mendeley Clinical Laboratory Test Reports). Taken together, MedGemma 1.5 serves as a robust, open resource for the community, designed as an improved foundation on which developers can create the next generation of medical AI systems. Resources and tutorials for building upon MedGemma 1.5 can be found at https://goo.gle/medgemma.
Andrew Sellergren, Chufan Gao, Fereshteh Mahvar +39
Vision--language models (VLMs) process images as visual tokens, yet their intermediate reasoning is often carried out in text, which can be suboptimal for visually grounded radiology tasks. Radiologists instead diagnose via sequential visual search; eye-tracking captures this process as time-ordered gaze trajectories that reveal how evidence is acquired over time. We use eye-gaze as supervision to guide VLM reasoning by introducing a small set of dedicated gaze tokens. These tokens are trained to predict gaze-selected image patch indices in temporal order, encouraging the model to follow human-like evidence acquisition and integration. Experiments on MIMIC-EYE and multiple external zero-shot benchmarks show consistent gains over baselines, achieving state-of-the-art in-domain performance and improved out-of-domain robustness. These results highlight temporally ordered gaze as an effective supervision signal for learning visually grounded medical reasoning.
Yiwei Li, Yifan Zhou, Huaqin Zhao +6
School of Computing, University of Georgia · Massachusetts General Hospital, Harvard Medical School · Department of Biomedical Engineering, New Jersey Institute of Technology