AI in oncology is no longer a futuristic experiment. Regulators are already approving algorithms for oncology decision support. The FDA's Oncology Center of Excellence launched its Oncology AI Program in 2023, as AI moved deeper into cancer drug development and clinical decision-making.
Yet there's an odd asymmetry in the public conversation. Much of it is still speculative: Can AI understand cancer? Can an algorithm really help an oncologist? Can machines be trusted with decisions involving life and death?
Meanwhile, the regulatory system has been quietly moving ahead — building expertise, evaluating emerging methodologies, and advancing the science needed to assess AI in oncology. That doesn't prove every algorithm works. But it tells us something more fundamental: AI in oncology has moved from a technological possibility to a regulatory and scientific problem. That's a far more interesting stage of development.
Part of the scepticism around medical AI comes from framing its task incorrectly. An algorithm doesn't need to "understand cancer" the way a human does. It needs to identify clinically meaningful patterns, quantify them reproducibly, and demonstrate that its predictions perform well enough for their intended use.
A pathologist recognises hundreds of features in a tissue sample. An oncologist integrates pathology, stage, biomarkers, imaging, history, and guidelines. An algorithm can examine far more measurable features across much larger datasets and surface statistical relationships that are hard to perceive consistently at human scale.
So the useful question isn't whether a machine thinks like an oncologist. It's whether computation can reveal information that improves an oncologist's judgment — a far more testable proposition.
This is no longer theoretical. In 2025, the FDA granted De Novo authorisation to ArteraAI Prostate, which analyses digitised prostate biopsy images to estimate long-term risks like distant metastasis and prostate-cancer-specific mortality. In 2026, it cleared ArteraAI Breast, which combines deep-learning analysis of pathology images with clinical variables to estimate distant-metastasis risk in a defined population of breast-cancer patients.
The significance isn't just that two companies received regulatory decisions — it's that algorithms now occupy a recognisable position inside the medical decision architecture, designed to answer questions like: How likely is this cancer to recur? How aggressive is it likely to be? Which patients may need more intensive treatment?
These aren't administrative questions — they can shape clinical choices. That means the standard for these systems can't be the same as for a consumer AI app. Performance, validation, intended use, generalisability, and patient safety all become central. That's precisely why regulation matters.
An even more consequential question follows: what treatment is this tumour likely to respond to? Prognosis tells us something about the likely course of disease. Treatment-response prediction asks something different — whether a specific intervention is likely to work for a specific patient.
This is where computational pathology companies like Valar Labs get interesting. Its Pancreas ChemoPredict test predicts how advanced pancreatic cancer patients will respond to first-line chemotherapy. Its Vesta platform extracts information from routine pathology images, including features tied to bladder cancer risk and treatment response — and in 2026, the FDA granted Breakthrough Device Designation to Vesta Bladder Risk Stratify Dx.
That designation isn't proof of clinical effectiveness. But it shows the direction of travel: regulators are engaging with algorithms aimed at serious diseases and unmet needs. Knowing a patient is high-risk is useful. Knowing why, which treatment is most likely to change that trajectory, and what toxicity might accompany it would be considerably more useful — and that distinction may become one of the most important in oncology AI.
Cancer treatment is fundamentally a problem of uncertainty. Two patients with the same diagnostic label can have very different biological trajectories — different tumours, molecular profiles, treatment histories, toxicity tolerance, even preferences. Yet clinical decision-making often has to compress that complexity into a small number of choices.
AI offers another possibility: instead of reducing complexity too early, computational systems could help preserve and interrogate it — analysing pathology, radiology, genomics, history, pharmacology, and outcomes together. The goal isn't more data; we already have plenty. It's converting existing data into patient-specific information that changes a decision. That's a much higher bar — and the right one.
The next stage may be larger than individual diagnostic algorithms. Imagine a clinical system that can ask, in sequence: What is the patient's disease trajectory? What's driving it? Which therapies are biologically and clinically plausible? Which is most likely to produce benefit? What's the probability of toxicity? What evidence supports the recommendation? What should the tumour board reconsider?
That starts to look less like a diagnostic tool and more like an oncology intelligence layer — sitting above specialised models rather than replacing them: one for pathology, one for radiology, one for molecular profiling, pharmacology, prognosis. The hard problem then isn't building another model. It's orchestrating the models around the patient — a distinction that could matter enormously for precision oncology.
Scepticism about medical AI is healthy — oncology needs more of it, not less. But it should be aimed at the right question. Instead of "Can AI replace an oncologist?" we should ask: "Can a validated computational system provide clinically useful information an oncologist cannot reliably derive at the same scale?"
That question can be tested, regulated, and compared against existing practice — and ultimately judged by what matters most: whether it improves decisions and outcomes for patients.
The FDA's growing engagement with AI in oncology doesn't mean the future has arrived. It means the infrastructure for that future is already being built. Perhaps the most important development in oncology AI won't be an algorithm that replaces the oncologist — it will be an ecosystem of algorithms that lets the oncologist see the patient with greater resolution. A less dramatic claim. Possibly the more consequential one.
Dr. Kanuru's AI-Assisted Precision Guidance consultation maps your specific case against the latest research — exactly the kind of patient-level intelligence this article describes. Always reviewed by Dr. Kanuru before it reaches you.