Can AI Make Diagnosis Better?

What improves, and why

Diagnosis is, in large part, pattern recognition under time pressure — exactly the kind of task deep learning does well. In screening mammography, AI-assisted reading has raised the cancer-detection rate and, at the same time, reduced recalls: more true cancers found, fewer women called back for nothing.^{Source 1} A large randomised trial found AI-supported single reading was a safe alternative to standard double reading, easing a workload that keeps rising faster than the specialist supply.^{Source 2}

The important word is assisted. The evidence is strong for AI as a second set of eyes and a triage layer; it is far weaker for AI reading alone. Even the best benchmarks place fully autonomous interpretation years away, on ethical and safety grounds as much as technical ones.^{Source 6}

Where it falls short

• Dataset shift. Accuracy measured on the training population can fall 15–30% when the model meets a different scanner, protocol or demographic. A model validated in Stockholm is not validated in Sydney.^{Source 5}

• The CAD precedent. Earlier computer-aided detection looked excellent retrospectively, then degraded accuracy in routine practice. The field has made this exact mistake before.^{Source 4}

• Narrow, not general. Strong performance on one lesion type or modality says little about the next. In head-to-head benchmarking, overall radiologist accuracy still exceeded AI.^{Source 3}

Where governance decides the outcome

A better diagnosis is only better if it holds up after go-live. That makes diagnosis a governance problem as much as a modelling one: external validation on local data before deployment, continuous monitoring for drift, and a clear line of accountability for an AI-influenced decision. Automation bias — the tendency to defer to a confident machine — is a documented failure mode, and the countermeasure is process, not a better model. Better is the headline dimension here; Safer is the one that determines whether the headline lasts.

How to think about adoption

Adopt AI as an assistive layer with the clinician firmly in the loop, not as a replacement reader. Insist on validation against your own population before trusting a vendor's benchmark, instrument for drift from day one, and measure recall and detection — not headline accuracy — as the outcomes that matter. Treat every published figure, including the ones above, as a hypothesis to be re-tested in your setting.