How to evaluate oncology AI
Oncology AI covers very different products: radiation contouring, cancer detection on images or slides, clinical-trial matching, and treatment decision support or outcome prediction. Each needs its own evidence.
Find what the tool does below. The questions in the core guide apply to all of them.
Last reviewed November 2026 · Independent and vendor-neutral
What does the tool you’re evaluating do?
Jump to the matching checklist. Some answers lead to a different guide.
Radiation contouring and planning
Applies if the tool draws organs at risk or target volumes, or automates treatment planning.
Questions to ask
- How much editing time does it save dosimetrists and physicians, and what share of contours are accepted with minor or no edits?
- Which disease sites, scanners and planning systems was it validated on? Does it work with yours?
- How are updates to the model handled, and are contours re-checked after an update?
Test it locally
- Run it on 20 to 30 of your own cases per disease site and measure edit time and geometric agreement with your approved contours.
Red flags
- Validated on only one planning system
- No site-by-site results
Clinical-trial matching
Applies if the tool screens patient records against trial eligibility criteria.
Questions to ask
- Compared with coordinator review, how many suggested matches are truly eligible (precision), and how many eligible patients does it miss?
- Does it read unstructured notes and pathology reports, or only structured data?
- What is the effect on screen-to-enroll ratio and coordinator hours?
Measure
- Screened, eligible and enrolled patients per trial, before and after go-live
Treatment decision support and outcome prediction
Applies if the tool suggests treatment options or predicts survival, toxicity or response.
Questions to ask
- Is there prospective evidence that it changes decisions or outcomes, not only retrospective accuracy?
- How well do its recommendations agree with current guidelines (for example NCCN), and how quickly is it updated when guidelines change?
- For predictions: is it calibrated in your population, and how was the outcome defined?
Red flags
- Survival predictions without calibration data
- No process for guideline updates
The predictive-tools guide covers calibration, thresholds and silent testing in more detail. Read that guide →
Browse oncology tools
Every listing shows regulatory status, published evidence and named health-system users where public.