Evaluating the Impact of Medical Image Reconstruction on Downstream AI Fairness and Performance

1. Problem Statement

AI-based image reconstruction models (e.g., for low-dose X-rays or accelerated MRI) are increasingly deployed in clinical workflows to enhance image quality from noisy or undersampled data. However, these models are traditionally evaluated using pixel-level metrics (e.g., PSNR, SSIM), which fail to capture their impact on downstream clinical tasks (diagnosis, segmentation) and algorithmic fairness.

The core research questions are:

1. How does AI-based reconstruction affect the diagnostic accuracy of downstream models when image quality (PSNR) degrades?
2. Does the reconstruction process introduce or amplify demographic biases (fairness) in downstream predictions?
3. Can bias mitigation strategies applied at the reconstruction stage effectively reduce downstream disparities without compromising performance?

2. Methodology

The authors propose a scalable evaluation framework that integrates reconstruction and diagnostic models in a tandem pipeline (visualized in Figure 1).

Datasets & Tasks

• MRI (UCSF-PDGM): 501 pre-operative glioma exams. Tasks: Tumor segmentation and classification (Grade/Subtype).
• X-Ray (CheXpert): 224,316 radiographs. Task: Multi-label classification of 12 thoracic pathologies.
• Degradation Simulation:
  • MRI: Radial undersampling in k-space at acceleration factors 4, 8, and 16.
  • X-Ray: Poisson noise simulation at varying photon counts (100k, 10k, 3k) to mimic low-dose acquisition.

Models

• Reconstruction Models: Three distinct architectures trained from scratch on each dataset:
  1. U-Net: Standard deterministic model with MSE loss.
  2. GAN (Pix2Pix): Adversarial training approach.
  3. Diffusion (SDE): Stochastic Differential Equation-based generative model.
• Diagnostic Models:
  • Classification: ResNet50 (MRI) and DenseNet (X-Ray), trained on original, non-degraded images.
  • Segmentation: U-Net (MRI).

Evaluation Metrics

• Reconstruction Quality: PSNR (Peak Signal-to-Noise Ratio).
• Downstream Performance: AUROC (Classification) and Dice Score (Segmentation).
• Fairness Metrics:
  • Classification: Worst-case Equalized Odds (EODD) difference across subgroups (Sex, Age, Race).
  • Segmentation: Adapted Skewed-Error Ratio (SER) based on Dice scores.

Bias Mitigation Strategies

Two strategies, originally designed for classification, were adapted and applied exclusively during the fine-tuning of the reconstruction models (keeping diagnostic models frozen):

1. Sample Reweighting: Inverse subgroup frequency sampling to balance representation.
2. Differentiable Equalized Odds (EODD): A fairness constraint added to the reconstruction loss function, minimizing the covariance between sensitive attributes and classifier predictions.

3. Key Results

A. Stability of Downstream Performance

• Decoupling of Metrics: There is a significant disconnect between pixel-level quality and diagnostic utility. As noise increased (lowering PSNR by 10–26%), downstream diagnostic performance remained largely stable.
  • MRI Segmentation: Dice score variation < 3%.
  • X-Ray Classification: Mean AUROC fluctuation < 1%.
• Robustness: Diagnostic models proved resilient to reconstruction-induced noise, retaining necessary diagnostic features even when PSNR dropped substantially. However, subtle pathologies (e.g., lung lesions) showed slightly higher sensitivity to noise than obvious ones (e.g., cardiomegaly).

B. Impact on Fairness

• Bias Amplification: While aggregate performance was stable, reconstruction did alter fairness metrics.
  • Sex: Showed the most significant bias amplification, with a median increase in EODD of 24%.
  • Age: Showed mixed results (some decrease, some increase).
  • Race: High variance, but generally modest shifts due to small sample sizes in specific subgroups.
• Magnitude: The additional bias introduced by reconstruction was generally modest compared to the inherent bias already present in the diagnostic models. However, in specific cases, shifts corresponded to a ~5% difference in sensitivity/specificity between subgroups, which is clinically significant.
• Model Sensitivity: Generative models (GAN and Diffusion/SDE) introduced less additional bias than the standard U-Net, which is counterintuitive given their generative nature.

C. Effectiveness of Mitigation

• Reweighting & EODD: Both strategies reduced bias, particularly for sex-related disparities on the CheXpert dataset.
  • EODD constraints showed slightly better median bias reduction than reweighting.
• Trade-offs: Mitigation strategies incurred minimal performance penalties (AUROC/PSNR drops < 1% for X-ray). However, on the MRI dataset, EODD constraints caused larger drops in PSNR and some downstream performance, suggesting dataset-specific efficacy.
• Limitations: Mitigation was not universally effective; results varied by pathology and sensitive attribute.

4. Key Contributions

1. Evaluation Framework: Introduced a holistic pipeline to assess reconstruction models not just by image fidelity, but by their downstream diagnostic and fairness implications.
2. Metric Decoupling: Demonstrated that standard pixel-level metrics (PSNR) are poor proxies for clinical utility and fairness in reconstruction tasks.
3. Fairness Analysis: Provided empirical evidence that reconstruction can amplify demographic biases (specifically regarding patient sex), even when overall accuracy remains stable.
4. Mitigation Adaptation: Successfully adapted classification-based bias mitigation techniques to the reconstruction stage, showing that fairness can be improved without retraining the diagnostic model, though with variable efficacy.

5. Significance and Clinical Implications

• Workflow Integration: As AI reconstruction becomes standard in clinical practice (over 80 FDA-cleared devices), this study argues that fairness and downstream performance must be evaluated alongside image quality. Relying solely on PSNR is insufficient.
• Bias Monitoring: The findings suggest that while reconstruction models are generally robust, they can act as a vector for bias propagation. Active monitoring of demographic disparities is necessary, especially for patient sex.
• Development Guidelines: Developers should prioritize downstream task evaluation. If bias mitigation is required, it may be more effective to address it at the classifier stage rather than the reconstruction stage, though reconstruction-stage interventions can provide incremental benefits.
• Future Work: The authors call for multi-stage bias mitigation strategies and the application of this framework to more diverse datasets and clinical populations to ensure generalization.

In conclusion, the paper provides a critical "reality check" for the medical AI community: AI reconstruction is safe for diagnostic accuracy but requires careful scrutiny regarding fairness.