Community Discussion · Policy

Medical AI Deployment: Don't Blindly Trust That 99% Sensitivity

PM YuanJul 102026/07/10 87 views

Lately working on AI imaging products at United Imaging Healthcare, what struck me most is this: 99% sensitivity in the lab might drop straight to 70% in clinical settings. Why?

Policy

2 replies

Ctrl + Enter to reply

Tang WenyuanJul 24(edited)

[quote="yuan_wenxuan, post:1, topic:272"]

Recently working on AI imaging products at United Imaging Healthcare, the most profound realization is this: 99% sensitivity in the lab can drop straight to 70% in clinical settings. Why?

[/quote]

This phenomenon is super common in the medical imaging circle. I know a few teams doing lung nodule detection who later found that the key issue was the noise distribution in clinical data being completely different from the lab, such as differences in imaging protocols across device manufacturers. Have you guys tried adding augmentation strategies that simulate clinical environments during the data collection phase?

Teacher ShenJul 17(edited)

[quote="yuan_wenxuan, post:1, topic:272"]

Recently working on AI imaging products at United Imaging, what struck me most was: 99% sensitivity in the lab might drop straight to 70% in clinical settings. Why?

[/quote]

This phenomenon is perfect for explaining "overfitting" and "data distribution shift" in teaching. When my innovation group worked on lung nodule detection, students also found that accuracy dropped when moving from test sets to hospitals. The students' feedback was, "It doesn't recognize them once the background changes."