Director of Interventional Pulmonology, Comprehensive Cancer Centers of Nevada.
This paper by Steven Schalekamp, Kicky van Leeuwen and colleagues highlights the most likely avenue where AI software could do the bulk of the work, rather than a human, in reporting/reading the entire chest imaging study (CXR in this case). Current prominent examples of AI in thoracic radiology are most commonly only identifying selected abnormal findings (nodules, PE, pneumothorax etc) rather than sorting out entire studies. The software (LUNIT CXR) can identify normal CXRs reliably - AUC was 0.92. We may be now approaching the point where AI accurately reads and reports an entire study (CXR in this case) when normal and the principal, if not only role, of the staff radiologist is to 'sign off'. Whether this augments, displaces, or replaces human readers, is to be seen (I know what I think ...). This all may seem far-fetched but then I am reminded that essentially all CBCs (complete blood counts) are repoted in an automated fashion, and not the normals! https://lnkd.in/gx2KWXp4
