BIONEWS
Cancer Care's AI Tools Work. The Rules to Trust Them Don't.
AI assisted CT reads, tumor board documentation tools, and pathology models that pull molecular signals from routine slides are already running in cancer clinics.
AI assisted CT reads, tumor board documentation tools, and pathology models that pull molecular signals from routine slides are already running in cancer clinics.
Oncology AI has stopped being a punch line. In 2026, AI-assisted CT is flagging nodules for earlier review, specialized language models are drafting tumor-board documentation, and AI-inferred transcriptomics is pulling recurrence-risk scores out of routine pathology slides. The clinical tools work well enough to be deployed. What is missing is the institutional scaffolding that would tell an oncologist when to trust them.
The contrast is sharpest in imaging. A peer-reviewed trial published in Lancet Oncology, the PANORAMA study00567-4/abstract), tested AI plus radiologists against radiologists alone for pancreatic cancer detection on standard-of-care CT scans. The framing matters: the trial evaluates augmentation, not autonomous detection. The radiologist remains the decision-maker. The AI is a second reader with a particular eye for the subtle density changes that mark early pancreatic tumors, the kind of lesion that easily reads as normal tissue on a busy day.
The same pattern shows up across the rest of the field. The American Association for Cancer Research recently highlighted a multimodal AI model that improves recurrence risk stratification in early breast cancer, layering imaging features with clinical and pathology data to refine who needs more aggressive therapy. At the San Antonio Breast Cancer Symposium in December, researchers presented TRINITY AI, a system that infers gene-expression-style recurrence scores directly from the standard H&E-stained slides that every pathology lab already produces. A 2025 review in Diagnostic and Interventional Imaging frames digital pathology as the next AI frontier alongside imaging, with foundation models trained on millions of slides that smaller labs can adapt rather than build from scratch.
In trial workflows, the friction is the slowest part of drug development. Reading a tumor's response to therapy under RECIST criteria, the standard measurement rules for solid-tumor scans, is a manual, radiologist-hour-intensive process. The Friends of Cancer Research ai.RECIST initiative is trying to standardize AI-assisted response evaluation, so a model's reading of "this lesion shrank by 30 percent" means the same thing across sponsors, regulators, and trial sites. If that scaffolding works, oncology could compress the readout cycle that currently stretches many solid-tumor trials by months.
Then there is the throughput signal. OpenEvidence, a clinical-AI platform, said this year that verified physicians ran 1 million consultations with its system in a single day. That figure is a company-reported adoption number, not a clinical-outcomes statistic, so the right reading is that doctors are willing to consult an AI in their workflow, not that the AI is making better decisions. Even so, it marks a volume at which oncology AI is no longer a curiosity.
What has not kept pace is the validation, bias, and explainability layer that decides whether any of these tools become durable clinical infrastructure. The National Cancer Institute, as summarized in a state-of-the-field roundtable in CancerNetwork, describes AI as an "unprecedented opportunity" for cancer care. That same piece, written by a roundtable that included Matthew Matasar of Rutgers Cancer Institute, frames the field's recent arc as a move away from what Matasar described there as a "hallucination-prone punch line," an AI output that sounds confident but is factually wrong, toward something increasingly serious. The same source flags the unaddressed problem: models trained on patient data that underrepresents some groups will work worse for those groups, and clinicians still cannot reliably inspect why a recommendation was made.
The practical question for 2026 is no longer whether AI can read a CT or summarize a tumor board. It is whether a community oncology practice in Ohio has the same confidence in the tool's output as an academic center running the original validation. Until benchmarks, bias audits, and explainability standards catch up to the deployments, the field is shipping faster than the safety harness is being built. The watch items are concrete: post-publication data from PANORAMA-style trials, FDA guidance on AI updates that change model behavior after clearance, and whether ai.RECIST or a successor becomes the default way sponsors measure response.