People-Centred Medical Image Analysis
By: Maria Petrova, Alexei Ivanov, Svetlana Kuzmina, Dmitry Smirnov
Published: 2026-04-28
View on arXiv →Abstract
This paper presents PecMan, a human-AI framework for medical image analysis that optimizes fairness, diagnostic accuracy, and workflow effectiveness. It addresses limited clinical adoption of AI by ensuring fair performance across diverse patient populations and seamless workflow integration through a dynamic gating mechanism.
Impact
transformative
Topics
6
💡 Simple Explanation
Artificial intelligence in medicine is often built by computer scientists focused on getting the highest accuracy scores. However, in real hospitals, doctors often ignore these AI tools because they are hard to use, don't explain their reasoning, or disrupt daily routines. This paper proposes a 'people-centred' approach: building AI alongside doctors and patients from day one. By prioritizing user trust, fairness, and clear explanations over marginal accuracy gains, the resulting AI becomes much more useful and widely adopted in actual clinical settings.
🎯 Problem Statement
Despite significant advances in AI for medical image analysis, clinical translation remains poor. Models that achieve state-of-the-art performance in academic benchmarks often fail in real-world deployment. This is caused by a disconnect between algorithm developers and end-users, leading to issues with algorithmic bias, lack of explainability, poor workflow integration, and ultimately, a lack of clinical trust.
🔬 Methodology
The methodology introduces a People-Centred Medical Image Analysis (PCMIA) framework. It involves three core pillars: 1) Participatory Design, where clinicians co-define the target metrics and workflow integration; 2) Explainability and Uncertainty Quantification, providing doctors with visual (saliency maps) and textual justifications alongside confidence scores; and 3) Fairness Auditing, ensuring the model's performance does not degrade across different demographic groups. The framework was evaluated through an A/B test in a simulated clinical environment, comparing a standard high-AUC model against a PCMIA-optimized model.
📊 Results
Models developed using the PCMIA framework demonstrated a 30% increase in clinician adoption and daily usage compared to baseline models. While there was a marginal 1-2% decrease in raw AUC on the test set, diagnostic accuracy in a human-AI collaborative setting improved by 15%. Additionally, fairness audits showed a 40% reduction in false-negative rate disparity across different ethnic groups, indicating a more robust and equitable clinical tool.
✨ Key Takeaways
The key insight is that AI in healthcare is a socio-technical system, not merely a math problem. Optimizing for purely mathematical metrics often sacrifices clinical utility. By embedding explainability, fairness, and user-centric design into the core of AI development, we can bridge the translation gap, foster clinician trust, and ultimately improve patient care.
🔍 Critical Analysis
While the paper makes a compelling case for a paradigm shift toward people-centred AI, it lacks comprehensive cost-benefit analyses. Implementing participatory design and maintaining continuous feedback loops is resource-intensive. The authors provide strong qualitative evidence of increased trust but fall short on quantifying the exact financial and temporal overhead this approach introduces to standard MLOps pipelines.
💰 Practical Applications
- B2B SaaS platform for continuous clinical validation and alignment of existing AI models
- Consulting services to help legacy MedTech companies modernize their AI software interfaces
- Licensing proprietary human-centric UI components to medical imaging vendors