Diagnostics and imaging

Algorithms that read images, signals and tests alongside the clinician. What they have become good at, where they fail, and what that means for a diagnosis you are given.

Algorithm-assisted diagnosis has moved from demonstrations on datasets to routine use in radiology, cardiology, dermatology and pathology. In most hospitals the systems do not diagnose: they flag, prioritise or measure — and the clinician still reads and signs the report.

That is why this section asks not “how good is the model” but “what does it change in the workflow”. An algorithm that detects an intracranial haemorrhage two minutes sooner helps nobody if the alert reaches a person who can do nothing with it at that hour.

For readers who do not work in the system: if a study was interpreted with software assistance, you can ask who read it and whether the algorithm contributed to the decision. Transparency is not a courtesy, it is part of informed consent.

• Detection, prioritisation and triage are three different functions, with three different ways of failing.
• A negative algorithm output does not cancel the clinician’s read: no flag is not the same as no disease.
• Ask on which population and which scanner it was validated, not only how large the reported performance is.
• Clinical benefit is measured in time to treatment and in outcomes, not only in accuracy on a test set.

Reading

• Siemens Healthineers obtains license for six new ARTIS systems in Canada — Siemens Healthineers has received a license from Health Canada for six new ARTIS systems, which can enhance medical interventions through advanced imaging technology.
• Recent Studies on Advanced Melanoma Treatment: RELATIVITY-047 — A recent study compared the efficacy of relatlimab plus nivolumab versus nivolumab alone in the treatment of advanced melanoma, showing promising long-term results.