People First: A Paradigm Shift in Medical Imaging AI

High Accuracy Doesn't Mean High Adoption: The Deployment Dilemma of Medical AI

In recent years, data-centric medical artificial intelligence has made remarkable progress, with diagnostic systems repeatedly setting new accuracy records. Yet a puzzling phenomenon is emerging — these AI systems that perform brilliantly in the lab are seeing far lower adoption rates than expected in real clinical settings.

A newly published research paper on arXiv (arXiv:2604.26991v1) offers an in-depth analysis of this contradiction and presents a key argument: the development of medical imaging AI needs to shift from a "data-centric" approach to a "people-centred" one.

Core Issues: Fairness Gaps and Workflow Disconnect

The research team identifies two major obstacles currently facing medical AI:

First, performance disparities are creating a fairness crisis. Existing diagnostic systems exhibit significant performance variations across different patient populations. For instance, diagnostic accuracy may drop noticeably for certain racial, gender, or age groups. Such bias is not only an ethical concern but also a direct barrier to regulatory approval. Medical device regulatory agencies worldwide are placing increasing emphasis on the equitable performance of AI systems across diverse populations, making performance disparities a major roadblock to market authorization.

Second, clinical workflow integration remains inadequate. Many AI systems are designed in isolation from real clinical scenarios. Researchers have focused excessively on data curation and performance metric optimization while neglecting physicians' practical needs, continuity within diagnostic workflows, and human-machine collaboration efficiency. An AI tool boasting 99% accuracy but unable to integrate with existing PACS systems or that disrupts a radiologist's reading rhythm will see its clinical value severely diminished.

Core Principles of the "People-Centred" Paradigm

The "People-Centred Medical Image Analysis" framework advocated in the paper does not reject the value of data-driven methods. Instead, it proposes a more comprehensive optimization framework:

• Multidimensional fairness evaluation: Rather than relying solely on overall accuracy as the core metric, the approach systematically assesses performance disparities across groups defined by different demographic characteristics, ensuring a balanced distribution of diagnostic capability.
• Clinically needs-driven design: Incorporating the needs of physicians, patients, and healthcare system administrators throughout the entire system design lifecycle — from problem definition to deployment and maintenance.
• Regulation-friendly development processes: Considering regulatory compliance requirements early in the R&D phase to reduce the risk of approval delays caused by fairness issues.

Industry Implications and Future Outlook

This study arrives at a critical juncture as global medical AI regulation is tightening. The U.S. FDA, the EU MDR, and China's NMPA are all intensifying their scrutiny of AI-based medical devices, with fairness and clinical utility becoming key evaluation criteria.

From a technology trend perspective, the "people-centred" philosophy is driving progress in several important directions: first, the construction of multi-center, multi-ethnic clinical validation datasets; second, the deeper application of explainable AI in medical imaging; and third, the standardization of human-machine collaborative diagnostic models.

For medical AI practitioners, this paper sends a clear signal: merely pursuing SOTA (state-of-the-art) performance is far from sufficient. The true technological breakthrough lies in making AI systems serve every patient equitably and integrate seamlessly into clinical practice. Only by bridging the last mile from the lab to the bedside can medical imaging AI truly fulfill its promise of transforming healthcare.