AI Research Portal

This paper introduces FairVLM, a unified framework designed to address demographic bias and prompt sensitivity in vision-language models (VLMs) for medical image segmentation. FairVLM integrates three components: Semantic-Retaining Counterfactual Prompting (SRCP) for generating diverse and clinically consistent prompts, Demographic-Aware Feature Normalization (DAFN) to mitigate latent representation bias across demographic groups, and a Fairness-Calibrated Loss (FCL) to penalize performance disparities and encourage prompt consistency. The authors claim that FairVLM significantly improves equity-scaled segmentation, reduces demographic disparity and relative performance gap, and maintains or boosts overall accuracy, while also demonstrating robustness to prompt changes and generalization to unseen datasets.

The deployment of VLMs in clinical settings faces two intertwined challenges:

Demographic Bias: Performance varies across different demographic groups, leading to disparities in segmentation accuracy. This is often exacerbated by underrepresentation of certain groups in training data.

Introduction of a prompt counterfactual generation strategy (SRCP) using LLMs to enhance robustness to linguistic variability while preserving clinical semantics.
Proposal of a lightweight demographic normalization module (DAFN) to reduce representation disparities across protected groups.
Design of a fairness-calibrated loss (FCL) that explicitly penalizes inter-group performance gaps and encourages equitable segmentation quality.
Extensive evaluations on the Harvard-FairSeg dataset and out-of-distribution settings demonstrating FairVLM's effectiveness in both fairness and accuracy.

FairVLM is a unified framework built upon existing VLM architectures (SAMed and LViT) that integrates three key components to address demographic bias and prompt sensitivity.

Semantic-Retaining Counterfactual Prompting (SRCP): This module generates multiple, semantically controlled prompt variations for a given image and its original prompt. It uses a language model (GPT-40) to create counterfactual prompts that are evaluated for diversity (Jaccard distance) and semantic consistency (cosine similarity) against the original prompt. The top-k (k=3) valid counterfactual prompts are selected to form an augmented set. This process exposes the model to linguistic variability, improving generalization.
Demographic-Aware Feature Normalization (DAFN): This module normalizes encoded visual and linguistic features to mitigate demographic biases. It computes group-specific mean and standard deviation statistics for features and uses an exponential moving average (EMA) to maintain stable statistics across mini-batches. These statistics are then used to normalize the features, making them invariant to demographic groups. DAFN operates post-encoding and requires no architectural changes.
Fairness-Calibrated Loss (FCL): This component is a regularization term designed to minimize performance disparities across groups while preserving task accuracy. It includes a fairness gap metric (Agap) and an entropy-based weighting scheme (πg) to adaptively emphasize underrepresented groups. Additionally, a counterfactual prompt regularization (CPR) loss is computed to minimize divergence between segmentation outputs from original and counterfactual prompts, promoting robustness. The total loss (Ltotal) is a combination of the base segmentation loss (Lbase), CPR loss (LCPR), and FCL (LFCL).

Architecture: FairVLM integrates modules into existing VLM baselines like SAMed and LViT. The DAFN module operates post-encoding.
Key Equations:
- SRCP validity criteria: Valid(p) = 1 if Dᵢ ∈ [€₁, €₂] ^ Sᵢ ≥ τ, else 0.
- SRCP weighted selection: Score(pᵢ) = λ · Dᵢ + (1 − λ) · Sᵢ, with λ=0.4.
- DAFN group-wise statistics: µg = (1/|Dg|) Σ_{zₓ ∈ Dg} zₓ, σg = √((1/|Dg|) Σ_{zₓ ∈ Dg} (zₓ - µg)²).
- DAFN EMA update: µg⁽ᵗ⁾ = αµg⁽ᵗ⁻¹⁾ + (1-α)µg, ôg⁽ᵗ⁾ = ασg⁽ᵗ⁻¹⁾ + (1-α)σg, with α = 0.3.
- DAFN average normalization parameters: µavg = (1/|G|) Σ_{g∈G} µg, σavg = (1/|G|) Σ_{g∈G} σg.
- DAFN normalization: z'ₓ = (zₓ - µavg) / σavg.
- Fairness gap: Agap = max_{g∈G} Diceg - min_{g∈G} Diceg.
- FCL: LFCL = Σ_{g∈G} πg log(Diceg / (Diceavg + ε) + Agap).
- Group importance weight: πg = -ng log(ng) / Σ_{g'∈G} -ng' log(ng').
- Counterfactual prompt regularization (CPR) loss: LCPR = Σᵢ=₁ᵏ (Dice(y, ŷpᵢ) + BCE(y, ŷpᵢ)).
- Total loss: Ltotal = Lbase + LCPR + LFCL.
Loss/Objective: The primary task loss is Lbase (Dice + BCE). The framework adds LCPR for prompt robustness and LFCL for fairness.
Optimization: AdamW optimizer with exponential learning rate decay. Learning rate: 0.001, weight decay: 1 × 10⁻⁴, momentum: 0.9.
Augmentation: Not explicitly detailed beyond prompt augmentation via SRCP.
Data Handling: Uses the Harvard-FairSeg dataset, with 10% of training data for validation.
Implementation Details: Batch size: 16. Early stopping patience: 30 epochs, minimum improvement: 0.05. Training up to 300 epochs.

Datasets: Harvard-FairSeg dataset. Out-of-distribution (OOD) tests included manual vs. GPT-generated prompts and cross-dataset generalization to MosMedData+ and QaTa-COV19.
Metrics: Dice, Intersection over Union (IoU), Equity-Scaled (ES)-Dice, ES-IoU, Standard Deviation (STD), Disparity Index (DI), Relative Performance Gap (RPG %).
Baselines: SAMed, LViT, SAMed+GroupDRO, SAMed+FEBS, LViT+GroupDRO, LViT+FEBS, and other fairness/robustness methods.
Numeric Results:
- FairVLM achieved the best overall performance on Harvard-FairSeg, improving ES-Dice for Cup from 86.91 to 88.53 (SAMed) and 87.21 to 88.58 (LViT). Rim ES-Dice improved from 85.44 to 86.91 (SAMed) and 85.77 to 88.76 (LViT).
- ES-IoU also improved: Cup from 77.93 to 78.89 (SAMed) and 78.86 to 79.77 (LViT); Rim from 71.73 to 72.62 (SAMed) and 72.95 to 74.28 (LViT).
- DI and RPG metrics showed significant reductions, indicating improved fairness. For example, DI Dice for Cup dropped from 1.17 to 0.60 (SAMed) and 1.52 to 0.51 (LViT).
- Prompt robustness: FairVLM showed negligible drops (<0.5%) in Dice and IoU across varied prompts, while baselines showed 2-3% declines.
- Cross-dataset generalization: FairVLM improved ES-Dice and ES-IoU by 1.5-2% on external datasets, with a slight drop in overall accuracy.
Runtime/Training Cost: Not stated in the paper.

No explicit paper-reported limitations were found.

Method evidence: "FairVLM integrates three key components: (1) Semantic-Retaining Counterfactual Prompting (SRCP), which generates clinically consistent and diverse prompt variations via large language models; (2) Demographic-Aware Feature Normalization (DAFN), a lightweight module that mitigates latent representation bias across demographic groups; and (3) a Fairness-Calibrated Loss (FCL) that explicitly penalizes performance disparities while encouraging prompt consistency." (Abstract)
Result evidence: "FairVLM achieves the best overall performance, enhancing segmentation accuracy while decreasing fairness gaps across all metrics." (Table 1 caption)
Limitation evidence: Not stated in the paper.

FairVLM: Enhancing Fairness and Prompt Sensitivity in Vision Language Models for Medical Image Segmentation

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Evidence

Additional Figures