💻 computer science

Stride-Net: Fairness-Aware Disentangled Representation Learning for Chest X-Ray Diagnosis

Stride-Net is a fairness-aware framework for chest X-ray diagnosis that utilizes patch-level learnable masking, adversarial confusion loss, and semantic alignment via Group Optimal Transport to learn disease-discriminative representations that are invariant to sensitive demographic attributes.

Original authors: Darakshan Rashid, Raza Imam, Dwarikanath Mahapatra, Brejesh Lall

Published 2026-02-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Darakshan Rashid, Raza Imam, Dwarikanath Mahapatra, Brejesh Lall

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Biased Detective"

Imagine you are a detective trying to solve a mystery (diagnosing a disease) by looking at crime scene photos (Chest X-rays).

Most AI models act like detectives who have a bad habit: they don't just look at the fingerprints or the broken window; they also notice the color of the victim's shoes or the type of furniture in the room. If, in their training, they saw that people wearing red shoes were often involved in certain crimes, they might start "predicting" a crime just because they see red shoes.

In medical terms, an AI might notice that certain demographic groups (based on race or gender) have specific anatomical traits or were scanned on different types of machines. Instead of looking at the lungs, the AI starts "cheating" by using these demographic clues to make a guess. This is dangerous because it leads to unfairness: the AI might be great at diagnosing one group but terrible at another, simply because it’s looking at the "wrong clues."

The Solution: Stride-Net (The "Precision Investigator")

The researchers created Stride-Net. Think of Stride-Net not as a detective who looks at the whole messy room, but as a specialized investigator with a high-tech magnifying glass.

Here is how Stride-Net works, using three clever "tools":

1. The Smart Magnifying Glass (Stride-Based Masking)

Instead of looking at the entire X-ray at once, Stride-Net uses a "learnable mask." Imagine the detective has a piece of cardboard with tiny holes cut in it. They slide this cardboard over the photo to focus only on the specific spots where the evidence (the disease) is likely to be—like the lung tissue—while physically blocking out the "background noise" (like the patient's bone structure or skin markers that might reveal their race or gender).

2. The Medical Dictionary (Semantic Alignment via BioBERT)

To make sure the detective doesn't get distracted, Stride-Net gives them a "Medical Dictionary." Every time the AI finds a visual pattern, it has to "check it" against a professional medical definition (using a tool called BioBERT).

The Analogy: It’s like telling the detective, "Don't just tell me you see something suspicious; you must prove that what you see matches the official medical definition of 'Pneumonia'." This prevents the AI from using "shortcuts" and forces it to stick to real medical science.

3. The "Forgetful" Assistant (Adversarial Disentanglement)

Stride-Net uses a clever trick called "adversarial training." Imagine the detective has an assistant whose only job is to try and guess the patient's race or gender based on the detective's notes.
The detective's goal is to write notes that are so focused on the disease that the assistant cannot guess the patient's identity. If the assistant can guess the race, the detective is told, "You're still looking at the wrong clues! Try again!" This forces the AI to "unlearn" or "disentangle" the demographic information from the medical information.

The Result: Better Accuracy, Better Fairness

Usually, when you try to make something "fair," it gets a little worse at its main job (this is called the "accuracy-fairness trade-off"). It’s like telling a student, "You can only use a dictionary during the test to be fair to everyone," which might lower their score.

Stride-Net broke this rule.

By being more precise about what it looks at, it actually became more accurate AND more fair at the same time. It performed better across different races and genders, especially in the critical "No Finding" category (making sure people who are healthy aren't misdiagnosed).

Summary in a Nutshell

Stride-Net is like a doctor who ignores everything about a patient except the actual medical evidence in their lungs, ensuring that every patient gets a fair and accurate diagnosis, regardless of who they are.

Technical Summary: Stride-Net

Stride-Net: Fairness-Aware Disentangled Representation Learning for Chest X-Ray Diagnosis

1. Problem Statement

Deep learning models for chest X-ray (CXR) classification often exhibit significant performance disparities across demographic subgroups (e.g., race and gender). These disparities are not merely statistical; they pose clinical risks, particularly for "gatekeeper" labels like "No Finding," where errors can lead to delayed diagnoses or unnecessary interventions.

The authors identify three core issues in existing approaches:

Shortcut Learning: Models exploit spurious correlations between sensitive demographic attributes (encoded in anatomy or acquisition protocols) and disease labels to boost average accuracy.
Ineffective Debiasing: Existing fairness methods often treat fairness as a post-hoc constraint, leading to a "leveling down" effect where overall diagnostic utility is sacrificed to achieve parity.
Entangled Representations: Standard models fail to separate clinically relevant features from demographic information within the latent space.

2. Methodology

The proposed Stride-Net framework aims to learn representations that are simultaneously disease-discriminative and demographically invariant. It operates at the patch level to localize clinical signals and uses three primary mechanisms:

Learnable Stride-Based Masking: Instead of using the entire image, Stride-Net employs a learnable mask ( $M_\theta$ ) that computes relevance scores between image patch embeddings and disease label embeddings. This selects only the most "label-aligned" patches, forcing the model to focus on localized clinical evidence (e.g., lung fields) rather than global demographic proxies.
Group-Optimal Transport (GOT) Alignment: To anchor visual features in clinical reality, the model aligns patch embeddings with semantic embeddings of disease labels generated by BioBERT. The GOT loss enforces two things: (1) close alignment between patches and their corresponding disease semantics, and (2) group-wise consistency to ensure that the representation of a disease is similar across different demographic groups.
Adversarial Disentanglement: To actively suppress sensitive information, the framework uses an adversarial training setup. A sensitive attribute classifier attempts to predict race or gender from the latent representation, while a gradient reversal layer (adversarial confusion loss) trains the feature extractor to make these predictions as difficult as possible, effectively "scrubbing" demographic information from the latent space.

Total Objective Function:
$L_{total} = L_{classification} + \alpha L_{GOT} + \beta L_{sensitive\_supervision} - \gamma L_{confusion}$

3. Key Contributions

Novel Architecture: A framework that integrates patch-level selection, semantic alignment via GOT, and adversarial debiasing.
Semantic Grounding: The use of BioBERT-based label embeddings to prevent shortcut learning by tying visual features to medical concepts.
Intersectional Focus: A methodology designed to address not just individual attributes (race) but intersectional identities (race-gender), which are often the most vulnerable to bias.

4. Results

The model was evaluated on the MIMIC-CXR and CheXpert benchmarks using two metrics: Predictive Quality Disparity (PQD) and Equality of Opportunity Measure (EOM).

Superior Accuracy-Fairness Trade-off: Unlike prior methods (like UBAIA or CheXclusion) that often sacrifice accuracy to gain fairness, Stride-Net consistently improved both. On MIMIC-CXR, it improved average accuracy by +1.6% and EOM by +4.0% over the best baseline.
Robustness to Intersectionality: Stride-Net showed significant gains in intersectional race-gender subgroups, where compounding biases typically cause standard models to fail.
Ablation Validation: Ablation studies confirmed that the GOT module stabilizes subgroup performance and that the stride-based masking is essential for preventing the model from relying on background artifacts.

5. Significance

Stride-Net represents a shift from post-hoc fairness constraints to intrinsic fairness through representation learning. By forcing the model to "look" only at clinically relevant patches and "align" those patches with medical semantics, the authors provide a blueprint for building medical AI that is both highly accurate and ethically equitable. This approach is highly transferable to other medical imaging modalities where demographic information is subtly embedded in the visual data.