To explore the impact of AI on medical imaging diagnostics and the verifiability of AI applications in oncology.
Approach:
AI Application Tiers: Dr. Kann outlined three tiers of AI applications: tier one identifies features visible to humans, tier two identifies features poorly seen by humans, and tier three identifies features invisible to humans.
Verifiability Challenges: He discussed the challenges of verifiability across tiers, noting that tier one applications are easily verifiable, while tier two and three applications face significant challenges in validation and monitoring.
Opportunities in Tier One Applications: Dr. Kann highlighted the potential of tier one applications to uncover previously unmeasured features, such as using AI to assess body composition in pediatric brain tumor patients.
Key Findings:
Tier one AI applications are primarily automation tools and are more widely accepted due to their verifiability.
Tier two applications can detect features that humans see poorly, but require local validation.
Tier three applications, which have the highest actionability, currently lack FDA clearance in radiology and face significant challenges in verifiability.
Interpretation:
The verifiability of AI applications decreases with higher tiers, impacting their acceptance and implementation in clinical settings.
Limitations:
Higher-tier applications are subject to data set drift, which can affect their predictive accuracy over time.
The complexity of biological understanding can complicate the training and stability of AI models.
Conclusion:
AI has the potential to enhance diagnostic capabilities in oncology, but careful consideration of verifiability and model limitations is essential.
Automated lesion measurements increased agreement on treatment response, although expert-generated proposals were accepted more often and required fewer adjustments.
