RobMed LLM Notes

04-research-thrusts

6 min read

Research Thrusts: From Discovery to Deployment

Four interconnected thrusts addressing measurement, causality, mitigation, and safe deployment of phrasing-robust medical VLMs

← Key Concepts | Timeline →

Thrust 1: Measuring Linguistic Brittleness Through VSF Med

Objective

Establish comprehensive measurement framework and dataset to quantify flip-with-stable-focus (FSF) and error-faithfulness gap (EFG) phenomena in medical VLMs.

Key Components

VSF Med Dataset

  • Scale: 2,000+ radiological questions from MIMIC-CXR
  • Paraphrases: 8-10 validated variants per question (16,847 total)
  • Categories:
    • Binary finding detection (40%, 800 questions)
    • Localization queries (20%, 400 questions)
    • Severity assessment (15%, 300 questions)
    • Differential diagnosis (15%, 300 questions)
    • Temporal comparison (10%, 200 questions)

Paraphrase Generation Protocol

Three-stage validation process:

  1. Linguistic Generation: Systematic variation across 5 dimensions
    • Lexical: Medical synonym mapping (23.1%)
    • Syntactic: Parse tree transformation (19.1%)
    • Pragmatic: Confidence modulation (15.9%)
    • Negation: Polarity inversion (14.6%)
    • Scope: Modifier reattachment (13.6%)
  2. Clinical Filtering: Radiology resident review (15% rejection)
  3. Semantic Validation: Board-certified radiologist (8% additional rejection)

ROI Annotation Framework

1,500 images with three annotation types:

  • Primary ROIs: Direct pathology evidence (Dice = 0.78 ± 0.12)
  • Contextual ROIs: Interpretive context (Dice = 0.64 ± 0.18)
  • Negative ROIs: Exclusion regions

Key Findings

Overall Metrics

ModelAccuracyFlip RateFSF IndexEFG (Del)AAC
MedGemma-4b-it0.72412.7%0.6830.3420.214
LLaVA-Rad0.68915.6%0.7140.3870.186
Human Radiologist0.8922.3%N/AN/AN/A

Phenomenon-Specific Flip Rates

  • Negation patterns: >22% (highest risk)
  • Scope ambiguities: 18-20%
  • Syntactic variations: 12-15%
  • Lexical substitutions: 8-10%

Attention Stability Paradox

  • Average SSIM across paraphrases: 0.876 ± 0.082
  • No significant difference between flipping (0.871) and consistent (0.879) pairs
  • Confirms visual processing stability despite answer changes

Clinical Risk Stratification

High-risk applications (require human oversight):

  • Emergency triage with negation-heavy queries
  • Automated reporting systems
  • Teaching/training scenarios

Moderate-risk (with safeguards):

  • Workflow prioritization with mandatory review
  • Quality assurance with ensemble voting
  • Research applications with aggregate statistics

Lower-risk (current deployment suitable):

  • Administrative routing by modality
  • Training data curation
  • Retrospective analysis with validation

Thrust 2: Causal Analysis of Failure Mechanisms

Objective

Identify computational pathways through which linguistic variation propagates to cause FSF and EFG, enabling targeted interventions.

Methodological Approach

Layer-wise Representation Analysis

Track similarity degradation through model layers:

LayerSim_l = cos(h_l^(p1), h_l^(p2))

Key findings:

  • MedGemma: Sharp drop at layers 12-16 (cross-attention)
  • LLaVA-Rad: Earlier divergence at layers 8-12
  • Vision encoder maintains >0.92 similarity throughout

Cross-Attention Interchange Experiments

Systematic component swapping to isolate failure sources:

  1. Swap queries while keeping K,V fixed
  2. Swap key-value pairs while keeping Q fixed
  3. Full cross-attention replacement

Results:

  • K,V swapping reduces flips by 38-43%
  • Query formation (text encoding) drives failures
  • Vision-language alignment remains stable

Gradient-Based Token Importance

I_token = ||∂L_flip/∂e_token||_2

Findings:

  • Negation tokens: 2.8× average importance
  • Medical terminology: 1.6× average
  • Image patches: Uniform (SD < 0.12)

Causal Pathway Mapping

graph TD
    A[Paraphrased Question] --> B[Text Encoder]
    B --> C[Query Formation]
    C --> D[Cross-Attention Layer 12-16]
    D --> E[Answer Divergence]
    
    F[Original Image] --> G[Vision Encoder]
    G --> H[Stable Features]
    H --> D
    
    style D fill:#f96,stroke:#333,stroke-width:2px
    style E fill:#fbb,stroke:#333,stroke-width:2px

Mediation Analysis Results

  • Total effect: 15.6% flip rate
  • Direct effect (bypassing attention): 9.2%
  • Mediated effect: 6.4% (41% of total)
  • Conclusion: Cross-attention mediates but doesn’t fully explain FSF

Thrust 3: Parameter-Efficient Mitigation

Objective

Develop targeted interventions that reduce FSF to <5% while maintaining diagnostic accuracy using minimal computational resources.

Theoretical Framework

Design Space of Interventions

  1. Architectural Modifications

    • Attention regularization
    • Cross-modal fusion redesign
    • Linguistic encoding stabilization

  2. Training Objectives

    • Paraphrase consistency loss
    • Contrastive learning on variants
    • ROI-aligned supervision

  3. Inference Strategies

    • Ensemble voting
    • Calibrated abstention
    • Linguistic preprocessing

Practical Implementation

LoRA-Based Targeted Adaptation

Focus on causally-identified components:

# Target layers 12-16 for MedGemma
lora_config = LoraConfig(
    r=16,  # Low rank
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],  # Query and value projections
    layers=[12, 13, 14, 15, 16]  # Causal hotspots
)

Multi-Objective Training

L_total = λ₁L_task + λ₂L_consistency + λ₃L_attention

Where:

  • L_task: Standard VQA loss
  • L_consistency: KL divergence between paraphrase outputs
  • L_attention: Attention stability regularization

Expected Outcomes

  • FSF reduction: 12-18% → <5%
  • Parameters modified: <1% (50M of 4B)
  • Training time: 8-12 epochs on 8× A100s
  • Accuracy delta: ±2% on diagnostic tasks

Thrust 4: Safe Clinical Deployment

Objective

Integrate robust models into clinical workflow with selective prediction and calibrated abstention achieving >99% sensitivity for critical findings.

Selective Conformal Triage Framework

Uncertainty Quantification

Multiple uncertainty sources:

  1. Paraphrase Disagreement: Ensemble variance
  2. Model Uncertainty: MC dropout, temperature scaling
  3. Attention Instability: Cross-paraphrase SSIM variance

Triage Decision Logic

def triage_decision(image, question):
    # Get predictions across paraphrases
    predictions = [model(image, p) for p in paraphrases]
    
    # Check consensus
    if variance(predictions) > τ_consensus:
        return "defer", "paraphrase_disagreement"
    
    # Check confidence calibration
    confidence = calibrated_confidence(predictions)
    if confidence < τ_confidence:
        return "defer", "low_confidence"
    
    # Check for critical findings
    if any_critical_finding(predictions):
        return "radiologist_review", "critical_finding"
    
    # Safe to auto-clear if normal with high confidence
    if all_normal(predictions) and confidence > 0.95:
        return "auto_clear", "normal"
    
    return "radiologist_review", "needs_review"

Deployment Metrics

Safety Guarantees

  • Critical finding sensitivity: >99%
  • False negative rate (normal cases): <0.1%
  • Calibration error (ECE): <0.05

Efficiency Gains

  • Auto-clearance rate: 30-40% of normal cases
  • Radiologist time savings: 25-30%
  • Average turnaround reduction: 2-3 hours

Integration Requirements

Technical Infrastructure

  • PACS integration for seamless workflow
  • Real-time inference (<2 seconds per case)
  • Audit trail for all decisions
  • Fallback mechanisms for system failures

Human Factors

  • Clear uncertainty communication
  • Explanation interfaces showing attention
  • Training for radiologists on system limitations
  • Feedback loops for continuous improvement

Cross-Thrust Integration

Data Flow

Thrust 1 (VSF Med) → Thrust 2 (Causal Analysis) → Thrust 3 (Mitigation) → Thrust 4 (Deployment)
     ↓                        ↓                          ↓                        ↓
  Dataset               Failure Pathways           Robust Models           Clinical System

Iterative Refinement

  • Thrust 4 deployment reveals new failure modes → Thrust 1 measurement
  • Thrust 2 analysis guides Thrust 3 intervention targets
  • Thrust 3 models enable Thrust 4 safety guarantees

Open Science Contributions

Each thrust produces reusable artifacts:

  1. VSF Med dataset (Thrust 1)
  2. Causal analysis toolkit (Thrust 2)
  3. Robust model checkpoints (Thrust 3)
  4. Deployment framework (Thrust 4)

Timeline Integration

Phase 1 (Sep-Dec 2025): Foundation

  • Complete VSF Med construction
  • Initial causal analysis
  • Baseline measurements

Phase 2 (Jan-Apr 2026): Development

  • Full causal investigation
  • Mitigation implementation
  • Model training and validation

Phase 3 (May-Aug 2026): Deployment

  • Clinical evaluation
  • Safety validation
  • Thesis completion

Success Criteria

Quantitative Targets

  • FSF rate: <5% (from 12-18%)
  • EFG elimination: Faithfulness independent of correctness
  • Deployment metrics: >99% sensitivity, 30% auto-clearance

Qualitative Goals

  • Radiologist trust and adoption
  • Regulatory pathway clarity
  • Open science impact

This integrated approach transforms the discovery of FSF and EFG from concerning observations into actionable insights, enabling safe deployment of medical VLMs in clinical practice.

Graph View

Backlinks