DeLaS: Deployment-Oriented Multi-Layered Safeguards for Medical Vision–Language Models

Selection → Screening → Decision → Retrospection

Medical vision–language models (VLMs) are approaching clinical use, but hallucinations in the form of plausible yet visually unsupported answers remain a major barrier to safe deployment. DeLaS incorporates visual grounding into reliability assessment across the deployment pathway, connecting model selection, hallucination screening, confidence-guided decisions, and retrospective monitoring.

The paper evaluates DeLaS across three medical VLMs, eight public benchmarks, and an independent breast-ultrasound cohort. DeLaS identified the best-performing VLM on all nine benchmarks and improved screening, calibration, and subgroup reliability estimation. In a multi-clinician reader study, DeLaS-guided assistance reduced incorrect-response propagation to 8.25%, compared with 15.50–32.50% under the comparator strategies.

DeLaS-Scr examines how a response representation changes under a mild, controlled visual intervention. Weakly grounded responses tend to exhibit larger shifts:

$$ \Delta = \left|\mathbf{h}_{\mathrm{ori}}-\mathbf{h}_{\mathrm{int}}\right|_2, $$

where $\mathbf{h}{\mathrm{ori}}$ and $\mathbf{h}{\mathrm{int}}$ are response representations from the original and intervened images.

DeLaS-Scr uses two separate probes to map the original response representation and the intervention-induced shift to decision scores, then combines them to estimate hallucination risk:

$$ p_{\mathrm{hall}} =\sigma\left(\mathbf{w}_m^\top[z_h;z_{\Delta}]+b_m\right). $$

Here, $z_h$ and $z_{\Delta}$ are the decision scores from two $\ell_2$-regularized logistic regression probes applied to the original representation and the shift, respectively. The scores are concatenated and combined using learned weights $\mathbf{w}m$ and bias $b_m$, followed by the sigmoid function $\sigma$. Higher $p{\mathrm{hall}}$ indicates greater hallucination risk.

DeLaS-Cal adjusts a response-level reliability score using three complementary visual-grounding cues: Vision Attention Share (VAS), Vision Attention Concentration (VAC), and Jacobian Norm (JN). Together, they characterize the amount, distribution, and sensitivity of visual reliance.

These cues form a sample-specific gate that modulates the original score $s$:

$$ g=\sigma\left(\mathbf{w}_c^\top[\mathrm{VAS},\mathrm{VAC},\mathrm{JN}]+b_c\right), \qquad p_{\mathrm{correct}}=s\cdot g^{1/T}+\beta. $$

Here, $\sigma$ is the sigmoid function, $T$ controls the gating effect, and $\beta$ provides an offset correction.

The resulting probability of response correctness serves as a shared reliability signal for model selection, response routing, and subgroup monitoring. Model selection and retrospective accuracy estimation use this signal without requiring ground-truth answers in the target setting. Likewise, DeLaS-Scr and DeLaS-Cal require no ground-truth labels at deployment once fitted on labeled development data.

repository/
├── configs/                         Workflow configuration
├── docs/                            Framework overview and assets
├── scripts/                         End-to-end workflow
├── src/layered_guardrails/
│   ├── layer0_inference/            Inference and feature preparation
│   ├── layer1_selection/            Model selection
│   ├── layer2_screening/            DeLaS-Scr and screening baselines
│   ├── layer3_calibration/          DeLaS-Cal and confidence routing
│   ├── layer4_retrospection/        Subgroup reliability assessment
│   └── evaluation/                  Cross-layer evaluation
├── tests/                           Workflow checks
├── CITATION.cff
└── pyproject.toml

Layer 0 prepares the inputs and features used by the four safeguard layers.

Install the package from the repository root:

python -m pip install -e .

For a quick start, run DeLaS with MedGemma on the VQA-RAD dataset. Configure the runtime settings and required credentials in your environment, then execute:

scripts/run_pipeline.sh configs/medgemma_vqa_rad.yaml

To evaluate another VLM–VQA pair, create a configuration for your model and dataset, adapting the model interface and data loading as needed. Then run the same workflow with your configuration file:

scripts/run_pipeline.sh path/to/your_config.yaml

The workflow saves its outputs under outputs/delas/, including hallucination scores, calibrated confidence, routing decisions, model comparisons, and subgroup reliability summaries.

With reference labels, evaluation reports include hallucination discrimination (AUROC), calibration (ECE and ACE), and correctness among accepted responses (safe rate).

The manuscript describing DeLaS is currently under review. If you use this code, please cite the software repository below. The paper citation will be added when publicly available.

Liu L, et al. (2026). DeLaS: Deployment-Oriented Multi-Layered Safeguards for Medical Vision–Language Models [Software]. GitHub. https://github.com/fdu-farm/DeLaS

@misc{liu2026delas,
  author = {Liu, Lei and others},
  title  = {{DeLaS}: Deployment-Oriented Multi-Layered Safeguards
            for Medical Vision-Language Models},
  year   = {2026},
  url    = {https://github.com/fdu-farm/DeLaS}
}