AI expands; governance rises

A chest X-ray algorithm works like a pattern sorter for shadows on a scan: it compares millions of pixels against examples of pneumonia, fluid, nodules, and other findings that radiologists learn to spot over years of training. A Scientific Reports paper published April 10, 2026 tested two of these image models on the NIH ChestX-ray14 and CheXpert benchmark datasets instead of on a single hospital’s archive. (nature.com) The better-performing model in that paper, called MedViT, reached 93.34% accuracy and a 94.17% macro area under the receiver operating characteristic curve on NIH ChestXray14, then 79.22% accuracy and 75.11% macro area under the curve on CheXpert. The second model, a hybrid of a convolutional neural network and a vision transformer, scored lower on both datasets, which is a reminder that newer architecture names do not automatically win in clinic-like tests. (nature.com) Pathology is a different corner of medicine: instead of reading a chest image, a pathologist reads a digitized microscope slide that can be so large it behaves more like Google Maps than a photo. On April 7, 2026, MedStar Health said it would deploy PathAI’s AISight Dx platform and related algorithms across its laboratory network for more than 40 pathologists. (digitalhealthnews.com) That platform is not just a picture viewer. Digital Health News reported that AISight Dx has United States Food and Drug Administration clearance for primary diagnosis and bundles slide storage, image viewing, collaboration tools, and artificial intelligence integration into one cloud-native system. (digitalhealthnews.com) MedStar also said it would use one tool called ArtifactDetect to catch slide-preparation or scanning problems during workflow, which is the lab equivalent of spell-check catching an error before a report goes out. Another tool, TumorDetect, is still for research use only, which draws a bright line between software that can support routine care now and software that still needs more evidence. (digitalhealthnews.com) The same kind of software is moving into a much touchier job: deciding whether care gets paid for. KFF Health News reported that a federal pilot called Wasteful and Inappropriate Service Reduction, or WISeR, began on January 1, 2026 and tests artificial intelligence-assisted prior authorization in traditional Medicare across Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington through 2031. (kff.org) (kffhealthnews.org) Prior authorization means a doctor has to ask permission before a test, procedure, or treatment moves ahead, and that process has mostly belonged to private insurance and Medicare Advantage rather than traditional Medicare. KFF wrote that 69% of insured adults say prior authorization is a burden, and 34% say it is their single biggest burden beyond costs when getting care. (kff.org) That is where the governance fight starts. KFF Health News reported that class action lawsuits have accused insurers of using algorithms to wrongly deny treatment, and legal and policy critics are asking who is accountable when a model speeds up a bad decision instead of a good one. (kffhealthnews.org) So the story is splitting in two directions at once. In radiology and pathology, the software is getting better at seeing patterns inside images; in insurance and Medicare, the same push for speed and scale is colliding with questions about appeals, transparency, and whether a patient can challenge a machine-shaped “no” before the harm is done. (nature.com) (digitalhealthnews.com) (kffhealthnews.org)