AI models claim EGFR from slides
Early thought leadership suggests foundation AI models can detect EGFR mutations directly from routine lung histology slides, promising faster and cheaper mutation calls than some molecular tests. The post argues these computational approaches—already discussed for 2025 deployments—could shift diagnostic workflows toward image‑driven molecular prediction, though governance and validation remain crucial (x.com).
A lung tumor slide is usually just a stained piece of tissue under a microscope, and the gene test is usually a separate lab step that can take more tissue, more money, and more time. In lung adenocarcinoma, doctors order that extra step because mutations in epidermal growth factor receptor can decide whether a patient gets a targeted pill instead of standard chemotherapy. (nature.com) (cap.org) A mutation is a spelling change in the tumor’s DNA, and epidermal growth factor receptor is one of the spellings lung cancer doctors care about most. If that spelling is present, drugs called tyrosine kinase inhibitors can be the first treatment, so missing it can send a patient down the wrong path. (nature.com) (iaslc.org) The surprise is that the ordinary slide may already contain faint visual clues to that DNA change. Artificial intelligence systems are being trained to read the slide the way a face-recognition system reads pixels, except here the goal is to guess a molecular alteration from tissue architecture, cell shape, and staining patterns. (aacrjournals.org) (nature.com) That idea has been around in research for a few years, but 2025 is when it started to look less like a lab demo and more like a workflow change. A Nature Medicine paper published on July 9, 2025 described a fine-tuned pathology foundation model trained on 8,461 digital lung adenocarcinoma slides to predict epidermal growth factor receptor status from routine hematoxylin and eosin images. (nature.com) A foundation model is a large general-purpose image model that gets adapted to a narrower task, the way a person who already knows thousands of faces can learn one family faster than someone starting from zero. In that July 2025 study, the model reached an area under the curve of 0.847 on internal testing, 0.870 on external testing, and 0.890 in a prospective silent trial. (nature.com) A silent trial means the model runs in the background while doctors keep using the normal process, so researchers can see whether the software works without letting it change care yet. In that trial, the authors reported the artificial-intelligence-assisted workflow could cut rapid molecular tests by up to 43% while maintaining the existing clinical standard. (nature.com) Another 2025 study pushed the same idea across a wider mix of hospitals and scanners. A Cancer Research Communications paper published on December 8, 2025 said its model was trained and tuned on more than 12,000 whole-slide images and reached an overall area under the receiver operating characteristic curve of 0.905 in one test set and 0.860 in an independent test set from 11 countries. (aacrjournals.org) That scanner detail matters because digital pathology is messy in the real world. The same paper reported 90.4% concordance across five of six commonly used slide scanners, which is the kind of result people look for when they ask whether a model learned cancer biology or just learned one hospital’s hardware quirks. (aacrjournals.org) None of this replaces molecular testing guidelines today. The College of American Pathologists, the International Association for the Study of Lung Cancer, and the Association for Molecular Pathology still frame molecular analysis as the standard for guiding targeted therapy decisions, and the College of American Pathologists says its current update is meant to strengthen or reaffirm most prior recommendations while adding newer genes. (iaslc.org) (cap.org) So the near-term use is closer to triage than replacement. If a slide-based model flags a likely epidermal growth factor receptor mutation within minutes of scanning, a lab can prioritize confirmatory testing, conserve scarce biopsy tissue, and get the right targeted drug discussion moving sooner for patients with lung adenocarcinoma. (nature.com) (aacrjournals.org) The catch is that a wrong call here is not a harmless software bug. These systems need site-by-site validation, rules for when humans overrule the model, and evidence that performance holds across biopsy samples, surgical resections, metastatic tumors, and different scanners before hospitals trust image-driven molecular prediction in routine care. (nature.com) (aacrjournals.org)
