Postdeployment Monitoring and Surveillance Methods, Guidelines, and Possibilities for AI in Radiology

Area of Science:

• Radiology AI
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Radiology AI systems are transitioning from validation to routine clinical practice.
• Postdeployment monitoring and postmarket surveillance are essential components of the total product life cycle (TPLC) for AI.
• Human clinical oversight models for AI range from human-in-the-loop (HITL) to human-out-of-the-loop (HOOTL).

Purpose of the Study:

• To define human clinical oversight models in AI radiology.
• To propose the human-on-the-loop (HOTL) model as a pragmatic approach for high-stakes imaging.
• To present a proactive monitoring system for AI in radiology.

Main Methods:

• Defined a spectrum of human clinical oversight models (HITL, HIPL, HOTL, HOVL, HOOTL).
• Proposed a monitoring system using temporal stability of inputs/outputs and predictive divergence.
• Incorporated uncertainty quantification to prioritize reviews.

Main Results:

• The HOTL model balances verification workload, autonomy, and risk for AI oversight.
• The monitoring system utilizes data points not requiring immediate ground truth determination.
• A threshold-based alerting system with tiered escalation and root cause analysis is described.

Conclusions:

• A proactive roadmap for postdeployment monitoring of radiology AI is presented.
• The proposed system prioritizes patient safety while enabling responsible AI scaling.
• This approach supports the continuous evaluation and improvement of AI in clinical practice.