Neuron Incidence Redistribution for Fairness in Medical Image Classification

Technical Summary: Neuron Incidence Redistribution for Fairness in Medical Image Classification

Problem Statement
Deep learning models for medical image classification often exhibit systematic performance disparities across demographic subgroups (e.g., age, gender, race). While existing fairness interventions typically target output-level metrics (such as sensitivity or false-positive rates) through pre-processing, adversarial training, or post-processing calibration, they frequently fail to explain the internal mechanisms driving these disparities. The authors identify a specific latent mechanism: feature entanglement (or polysemanticity) within the penultimate layer of transfer-learned models.

Empirical analysis reveals that in standard fine-tuned models (e.g., ResNet-18), the dominant activation channels are not purely disease-specific. Instead:

• The dominant channel under positive predictions is co-activated by both disease-positive samples and "privileged" demographic groups (e.g., males, older patients). This leads to inflated predicted probabilities and over-diagnosis for these groups.
• The dominant channel under negative predictions is preferentially co-activated by "disadvantaged" groups (e.g., females, younger patients), suppressing their scores and causing systematic under-diagnosis.

This entanglement means a single feature channel encodes both clinical morphology and demographic correlates simultaneously, directly causing True Positive Rate (TPR) and False Positive Rate (FPR) disparities.

Methodology: Neuron Incidence Redistribution (NIR)
To address this, the authors propose Neuron Incidence Redistribution (NIR), a lightweight, differentiable regularization method that operates without requiring demographic labels during training.

1. Neuron Incidence Definition: For a mini-batch, the "incidence" of a specific neuron $j$ is defined as the predicted-probability-weighted mean activation:
   $$\phi_j = \frac{\sum_{i=1}^{B} \hat{p}_i z_{ij}}{\sum_{i=1}^{B} \hat{p}_i + \epsilon}$$
   where $\hat{p}_i$ is the predicted probability of the positive class and $z_{ij}$ is the activation of neuron $j$ for sample $i$. This weighting ensures the metric focuses on the representational structure of samples the model already considers positive.

2. Redistribution Loss: The method penalizes the variance of these incidence values across all neurons in the penultimate layer. If positive-class evidence is concentrated in a few neurons (high variance), the loss increases. The objective is to distribute the "incidence" of positive predictions uniformly across the entire layer:
   $$\mathcal{L}_{IR} = \frac{1}{d} \sum_{j=1}^{d} (\phi_j - \bar{\phi})^2$$

3. Training Objective: The total loss combines standard Binary Cross-Entropy (BCE) with the NIR regularizer:
   $$\mathcal{L}_{total} = \mathcal{L}_{BCE} + \lambda \mathcal{L}_{IR}$$
   The authors use a fixed $\lambda = 0.1$ across experiments, demonstrating that the method provides consistent improvements without dataset-specific hyperparameter tuning.

Key Contributions

1. Mechanistic Insight: An empirical analysis linking "dominant-channel entanglement" to asymmetric over- and under-diagnosis. The authors demonstrate that top positive-class neurons are preferentially co-activated by privileged demographics independent of the disease label.
2. NIR Algorithm: A differentiable, label-free regularizer that encourages distributed positive-class representations across the full penultimate layer, breaking the entanglement between disease signals and demographic correlates.
3. Empirical Validation: Substantial fairness improvements on the HAM10000 skin lesion dataset and generally positive (though mixed) results on the Harvard OCT-RNFL glaucoma dataset, achieved without sensitive attribute supervision.

Experimental Results
The method was evaluated on two datasets using a fine-tuned ResNet-18 architecture.

• HAM10000 (Skin Lesions):

  • Fairness: NIR reduced TPR disparity across age groups from 10.81% to 0.93% and across gender from 12.04% to 0.74% (reductions of >90%). FPR disparities also improved significantly.
  • Accuracy: The method yielded a marginal AUC improvement of 0.51 points (89.74% to 90.25%), indicating that distributed representations do not harm discriminative performance.

• Harvard OCT-RNFL (Glaucoma):

  • Fairness: NIR substantially reduced FPR disparity for race (15.68% to 10.66%) and age (12.69% to 1.80%). TPR disparities saw modest improvements for race and gender.
  • Accuracy: AUC decreased by 3.03 points (84.71% to 81.68). The authors attribute this to the smaller dataset size and class imbalance, where enforcing uniform incidence may partially suppress high-incidence neurons that carry genuine disease information entangled with demographics. This highlights an inherent accuracy–fairness tradeoff in constrained data regimes.

Significance and Claims
The paper claims that Neuron Incidence Redistribution offers a principled strategy for improving demographic fairness in medical AI by targeting the root cause of disparities: the polysemantic nature of dominant neurons in transfer-learned models.

• Practicality: Unlike many fairness interventions, NIR requires no demographic labels at training time, making it highly practical for clinical deployment where sensitive attributes are often unavailable or protected.
• Interpretability: The work connects abstract representation learning concepts (polysemanticity, neuron collapse) to observable demographic disparities, providing a mechanistic explanation for why models fail specific subgroups.
• Trade-off Awareness: The authors modestly acknowledge that while NIR is highly effective on larger datasets (HAM10000), it may incur an accuracy cost on smaller, imbalanced datasets (OCT-RNFL). They frame this not as a failure, but as an important consideration for the accuracy–fairness tradeoff, suggesting that forcing broader feature utilization can reduce reliance on the most discriminative but demographically entangled neurons.

The authors conclude that redistributing latent disease evidence across the full penultimate layer is an effective, label-free approach to mitigating subgroup disparities, though future work should explore combining this with explicit disentanglement or causal constraints to further decouple features.