02-key-concepts

Key Concepts: FSF and EFG in Medical VLMs

Understanding the critical failure modes that threaten safe clinical deployment of medical vision-language models

Terminology note: In my LaTeX proposal and newer documents, I refer to these phenomena as Phrasing-Sensitive Failure (formerly FSF) and the Misleading Explanation Effect (formerly EFG).

← Proposal | Timeline →

---

Flip-with-Stable-Focus (FSF)

Definition

FSF occurs when a medical VLM provides contradictory answers to semantically equivalent questions while maintaining stable visual attention patterns. This represents a dangerous disconnect between linguistic processing and visual grounding.

Mathematical Formulation

FSF Index = |{(q_i, q_j) : f(x, q_i) ≠ f(x, q_j) ∧ SSIM(A_i, A_j) > τ}| / |{(q_i, q_j) : f(x, q_i) ≠ f(x, q_j)}|

Where:

• f(x, q) is the model’s answer to question q about image x
• A_i, A_j are attention maps for questions q_i, q_j
• τ = 0.85 (empirically calibrated threshold)
• SSIM measures structural similarity between attention maps

Clinical Example

Original: "Is there evidence of pneumothorax?"
Model: "No" (with attention on lung periphery)

Paraphrase: "Can you see any collapsed lung?"
Model: "Yes" (with nearly identical attention pattern)

The radiologist sees consistent visual focus but receives contradictory answers—a silent failure that could impact patient care.

Empirical Findings

• Prevalence: 12-18% of paraphrase pairs in medical VLMs
• Attention Stability: 68-71% of flips occur with SSIM > 0.85
• Linguistic Triggers: Negations (>22% flip rate), scope ambiguities (18-20%)
• Model Comparison: MedGemma (12.7% FR) vs LLaVA-Rad (15.6% FR)

Error-Faithfulness Gap (EFG)

Definition

EFG describes the counterintuitive phenomenon where standard faithfulness metrics (deletion/insertion AUC) show higher scores for incorrect predictions than correct ones. This means explanations appear most “faithful” precisely when the model is wrong.

Mathematical Formulation

EFG_metric = E[metric | ŷ ≠ y*] - E[metric | ŷ = y*]

Where:

• metric ∈ {Deletion AUC, Insertion AUC, Sufficiency, Necessity}
• ŷ is the model prediction
• y* is the ground truth

Key Metrics Affected

Metric	Correct Predictions	Incorrect Predictions	Gap (EFG)
Deletion AUC	0.412	0.754	+0.342***
Insertion AUC	0.623	0.921	+0.298***
Sufficiency	0.534	0.687	+0.153**
Necessity	0.298	0.476	+0.178***

Clinical Implications

When a model incorrectly identifies pathology:

1. Saliency maps highlight regions more precisely
2. Deletion of highlighted regions causes larger performance drops
3. Clinicians see “stronger” visual evidence for wrong answers
4. Trust calibration becomes inverted—highest confidence when least reliable

The Coupled Failure Mode

FSF and EFG create a particularly dangerous combination:

Scenario: Missed Critical Finding

1. Query: “Is there evidence of pulmonary embolism?”
2. FSF Effect: Model answers inconsistently across phrasings
3. Stable Attention: Focus remains on pulmonary arteries (correct region)
4. EFG Effect: When wrong, explanation metrics are paradoxically stronger
5. Clinical Risk: Radiologist trusts the “well-explained” incorrect answer

Why This Matters

• Silent Failures: Errors hidden behind convincing explanations
• Trust Erosion: Inconsistency discovered only through accidental rephrasing
• Safety Compromise: Critical findings missed due to linguistic variation
• Deployment Barrier: Current metrics inadequate for clinical validation

Mechanistic Understanding

Layer-wise Analysis

Representation similarity between paraphrases degrades sharply at:

• MedGemma: Layers 12-16 (cross-attention layers)
• LLaVA-Rad: Layers 8-12 (earlier due to shallower architecture)

Causal Evidence

Cross-attention interchange experiments show:

• Swapping key-value pairs reduces flip rate by 38-43%
• Query formation (text encoding) drives most failures
• Visual features remain stable throughout

Token Importance

Gradient analysis reveals:

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

Mitigation Strategies

Parameter-Efficient Targeting

Based on causal analysis, we focus interventions on:

1. LoRA Adapters: Language attention blocks (layers 12-16)
2. Consistency Loss: KL divergence between paraphrase outputs
3. Selective Training: <1% of parameters modified

Expected Improvements

• FSF reduction: 12-18% → <5%
• Maintained diagnostic accuracy (±2%)
• Convergence in 8-12 epochs on limited compute

Deployment Safeguards

1. Paraphrase Ensembling: Vote across 3+ phrasings
2. Inverted Confidence: Lower deletion AUC → higher trust
3. Selective Abstention: Defer when paraphrases disagree
4. Linguistic Preprocessing: Standardize to avoid negations

Open Questions

1. Cross-Modal Universality: Do FSF/EFG occur in CT, MRI, ultrasound?
2. Multilingual Manifestation: How do translations affect these phenomena?
3. Temporal Consistency: Do longitudinal comparisons show similar patterns?
4. Human Factors: How do radiologists adapt to these failure modes?
5. Regulatory Implications: How should FDA evaluate linguistic robustness?

Resources

• VSF Med Dataset: HuggingFace
• Evaluation Toolkit: GitHub
• Technical Paper: ArXiv
• Clinical Guidelines: [Link to deployment recommendations]