French MedTech company Raidium, a pioneer in precision radiology, has unveiled the first promptable foundation model capable of measuring solid tumors in 3D and in real time with expert-level accuracy.
It claims that OncoPilot provides clinicians, in seconds, with the volumetric biomarkers needed to accelerate the development of new treatments and enhance clinical decision-making.
The intelligent assistant transforms image analysis into a fast, interactive, and reproducible process. With OncoPilot, says Raidium, it sets a new milestone toward faster and more consistent oncology care. The solution was officially unveiled at ASCO 2025 in Chicago this week.
Raidium provided the following information:
While most medical AI tools operate as peripheral modules, OncoPilot was designed as a natively integrated copilot within Raidium’s proprietary PACS viewer. No export, no context switching: the AI operates directly within the familiar reading interface, maintaining seamless workflow continuity. With a simple click or manual selection, users can prompt OncoPilot to generate an editable 3D mask of the target lesion — ready for direct use.
Each manual correction contributes to the AI’s ongoing learning, continuously improving segmentation quality based on real-world usage. This in situ learning capability represents a significant advantage over static, pre-trained models. Future versions will automatically localise target lesions, calculate complex biomarkers (e.g., total tumor burden, growth kinetics), and draft structured reports — with final validation always in the hands of the radiologist.
A true imaging copilot
This collaborative approach keeps the radiologist at the center of the loop, ensuring clinical control, full traceability of edits, and continuous learning from human interaction. OncoPilot embodies an augmented radiology paradigm, where AI does not replace the physician but takes over repetitive tasks while providing advanced analytics — fluidly, without friction or cognitive overload.
Toward faster, more personalized oncology care
Traditional tumor measurements following RECIST 1.1 are labor-intensive, poorly reproducible, and typically restricted to two dimensions. By transforming a simple click into an editable mask, OncoPilot simplifies radiologist workflows and opens the door to volumetric biomarkers essential for personalised medicine.
- Clinical value for patients: With more precise 3D analysis, OncoPilot enables earlier and more personalised therapeutic decisions, such as adjusting chemotherapy regimens or dosing strategies based on actual tumor volume rather than mere diameters. Over time, patients will benefit from novel imaging biomarkers that make precision medicine viable for routine follow-up.
- Value for clinical research: Next-generation therapies — including immunotherapies, antibody-drug conjugates (ADCs), and targeted radioligands — often induce complex, nonlinear responses. OncoPilot is designed to extract advanced biomarkers such as total tumor burden, ultra-early response signatures, and comprehensive whole-body longitudinal follow-up. These metrics support faster go/no-go decisions, improve adaptive trial design, and help identify surrogate endpoints that can shorten time-to-market.
A scalable platform for clinical trials
Beyond clinical use, OncoPilot serves as a biomarker factory for cloud-based clinical trials, enabling pharmaceutical companies to standardise, accelerate, and de-risk imaging analysis at scale.
With OncoPilot, Raidium sets a new standard for precision oncologic imaging. Regulatory processes are underway, and the company is actively working to accelerate clinical deployment while exploring partnerships to expand into other imaging modalities such as PET-CT and MRI.
Key Features of OncoPilot:
- Real-time 3D, multi-organ measurements
- Radiologist-level precision: mean DICE score of 0.79 after editing
- Speed: 17.2 seconds per measurement — 17% faster than manual 2D reading
- Reduced inter-reader variability: standard deviation improved from 2.4 mm to 1.7 mm
- 7 organs covered — versus 1 to 2 for most AI solutions