Accuracy is not the finish line
The papers are full of CNNs hitting 90%+ on ChestX-ray, skin cancer, and diabetic retinopathy screening. Eight years on, fewer than 30% of AI imaging products have actually changed a clinical workflow.
Why?
Three bottlenecks that get overlooked in real deployments
1. False negatives are far more dangerous than false positives
Missing a lung nodule = missing an early lung cancer. Over-calling one = one extra CT.
Evaluate on averaged F1 or accuracy and weak recall disappears into the number. What the clinic cares about is sensitivity, not average accuracy.
2. If the model won’t explain itself, doctors won’t use it
A radiologist doesn’t need “92% malignant”. They need “these texture features I’ve marked support a malignant read”.
Interpretability isn’t a research topic here. It’s a hard gate on shipping.
3. Data drift degrades the model quietly
Scanner settings, imaging protocols, and patient populations all differ from hospital to hospital. Move a model trained at hospital A over to hospital B and a 10-point drop in accuracy is normal.
Three rules if you’re building one
- Make a clinical metric your north star, not a technical one
- Design interpretability in from day one, don’t bolt it on afterwards
- Monitor distribution drift continuously, with a defined trigger for retraining
The value of an AI imaging product isn’t beating the radiologist. It’s taking on the repetitive work nobody wants to do, and leaving the work that needs judgment to a person.