Reconstruction of patient-specific confounders in AI-based radiologic image interpretation using generative...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Radiology

Background:

• Automated diagnostic assistance systems, particularly those using artificial intelligence (AI), require reliable detection of misleading patterns to ensure user trust.
• Current methods for visualizing confounding factors in AI diagnostic systems are insufficient.
• Identifying and mitigating biases in AI models is critical for clinical adoption.

Purpose of the Study:

• To introduce DiffChest, a novel self-conditioned diffusion model designed to detect and visualize confounding factors in chest radiographs.
• To enhance the reliability and interpretability of AI-driven diagnostic assistance systems.
• To improve diagnostic accuracy by addressing potential model biases.

Main Methods:

• Developed DiffChest, a self-conditioned diffusion model trained on a large dataset of 515,704 chest radiographs from 194,956 patients.
• Utilized patient-specific explanations and visualization techniques to identify confounding factors.
• Validated confounder detection capabilities using inter-reader agreement (Fleiss' kappa values ≥ 0.8) and assessed accuracy across various prevalence rates (10%-100%).

Main Results:

• DiffChest demonstrated high inter-reader agreement (Fleiss' kappa ≥ 0.8) in identifying treatment-related confounders.
• Confounding factors were accurately detected across a wide prevalence range.
• The model achieved excellent diagnostic accuracy for 11 common chest conditions, including pleural effusion and heart insufficiency.

Conclusions:

• Diffusion models show significant potential for robust medical image classification.
• DiffChest provides valuable insights into confounding factors, enhancing AI model interpretability and reliability.
• The proposed approach can improve user trust and clinical utility of AI in medical diagnostics.