AI in Oncology

Oncologists meet AI in three main places: detection, treatment planning and decision support. Detection is the most mature. In mammography, the Swedish MASAI randomized trial, first reported in 2023, found that AI-supported screening detected more cancers than standard double reading while cutting radiologist screen-reading workload by roughly 44 percent. In pathology, the FDA authorized Paige Prostate in 2021 as the first AI product to assist pathologists reading prostate biopsies. Research models can also estimate features such as microsatellite instability directly from routine stained slides, and MIT and Massachusetts General Hospital researchers built Sybil, which predicts future lung cancer risk from a single low-dose CT scan. Treatment planning is quietly widespread. Auto-contouring tools draft outlines of tumors and nearby organs for radiotherapy, a task that can take clinicians hours by hand. Clinicians still review and edit every contour. Treatment selection is where caution is most needed. IBM's Watson for Oncology was marketed to suggest therapies but faced reports of recommendations that disagreed with experts and some that were unsafe, and IBM later sold its Watson Health business. The lesson was that training on a small set of hypothetical cases and one institution's preferences does not produce reliable general advice. Tumor boards, where surgeons, oncologists, radiologists and pathologists discuss cases together, increasingly use software that assembles data and flags matching clinical trials or guideline pathways. Large language models are being tested to summarize records for these meetings, but hallucinated or outdated facts are a serious risk. The main limit of current evidence is that many studies are retrospective, measure accuracy rather than patient outcomes, and come from a few well-resourced centers. Finding more small cancers is only helpful if it improves survival rather than increasing overdiagnosis.