Discussion of AI in screening tends to jump straight to the diagnostic question: can a model tell a positive case from a negative one? It is the most interesting question and the most heavily published one. It is also, in a real programme, rarely the first problem you hit.
The first problem is that a meaningful share of images collected in field conditions are not adequate to assess at all. They may be out of focus, poorly lit, obstructed, or not framed on the anatomy of interest. Any downstream analysis performed on those images is not just unreliable — it is confidently unreliable, which is worse.
Selection before classification
Research published in Diagnostics in 2020, using ensemble deep-learning methods on smartphone-acquired cervical images including data from the SEVIA programme, addressed precisely this: identifying which images in a dataset are adequate for assessment.
Framed as a product decision rather than a research result, an adequacy model earns its place for reasons that have little to do with diagnostic accuracy.
- It can run at the point of capture, while the patient is still present and a retake costs almost nothing.
- Its failure mode is mild — asking for a second photograph — compared with a missed or false finding.
- It improves every downstream step, including human review, by removing noise from the queue.
- It produces an objective, trackable quality signal for supervision and training.
A model that asks for a better photograph is doing quality assurance. It is not making a clinical decision, and it should not be described as though it were.
Ensembles and the cost of being wrong
Ensemble approaches — combining several models rather than relying on one — are attractive here because the relevant property is not peak accuracy but stability. A single model can be confidently wrong on inputs unlike its training distribution, and field-collected images are full of those. Agreement across differently-trained models is a usable proxy for "this input is unusual, treat the output with caution".
That matters more in deployment than a marginal gain on a benchmark. A screening programme can absorb a model that occasionally says "I am not sure, please review". It cannot easily absorb one that is silently wrong at a low but steady rate.
Where this leaves clinical AI
RADIX, our medical-imaging platform, is in development and explores AI-assisted analysis across X-ray, CT, MRI and ultrasound. The framing we hold to is the same one that came out of the cervical-imaging work: the system assists, and interpretation stays with clinicians.
The cited research contributes evidence on image adequacy and quality improvement. It is not evidence for unrelated clinical-AI claims, and should not be cited as such.
