Radiation therapy is an important treatment for pancreatic cancer, but the pancreas is a soft-tissue organ that moves and shifts with breathing and digestion. To deliver radiation precisely to the tumor, doctors typically implant small metal markers (fiducials) into or near the pancreas during a separate invasive procedure. These markers are then visible on X-ray during treatment and used to confirm the tumor's exact position before each radiation dose.
The fiducial implantation procedure is invasive and carries risks, including bleeding and infection. It requires specialist expertise, delays the start of treatment, and the metal markers can cause image artifacts that obscure tumor boundaries during planning. A way to localize the tumor precisely without implanted markers would be a significant clinical advance.
Researchers from Stanford University developed a deep learning algorithm that learns to localize the pancreatic tumor from standard 2D kV X-ray images taken during radiation treatment — the same images already routinely acquired but previously unused for soft-tissue localization. The network takes an X-ray image as input and outputs the predicted position of the planning target volume (PTV).
To generate enough training data for the algorithm, the team used each patient's planning CT scan and introduced simulated anatomical variations — deformations, rotations, and translations — to mimic the range of positions a patient's anatomy might take between and during treatment fractions. From these simulated scenarios, thousands of digitally reconstructed radiograph (DRR) images were created as training examples.
The algorithm is 'patient-specific' — trained for each individual patient using their own CT data — and was validated retrospectively on patients who had actually undergone pancreatic cancer radiotherapy.
The deep learning model predicted tumor positions with mean absolute differences (MADs) of less than 2.60 mm in all directions (anterior-posterior, lateral, and oblique). For comparison studies against patients with and without implanted fiducials, MADs were less than 2.49 mm — comparable to fiducial-based localization.
Lin's concordance correlation coefficients between predicted and actual positions exceeded 93% across all measurement axes, indicating excellent agreement. These results demonstrate that accurate tumor localization is achievable without fiducials, using only the X-ray images that are already taken during routine treatment sessions.
If this technology is adopted clinically, patients who currently require fiducial implantation before radiation therapy could avoid the procedure entirely. This would reduce procedural risk, eliminate treatment delays, and make radiation therapy more accessible for patients who are too frail for additional procedures.
The approach could also benefit patients on hypofractionated regimens, where each radiation dose is higher and accurate targeting is especially critical. The AI-based localization could be run in real-time during treatment sessions, enabling dynamic correction if the tumor shifts unexpectedly.
This study demonstrates that deep learning can localize pancreatic tumors on X-ray images with accuracy sufficient for clinical image-guided radiation therapy. The technique works without any implanted fiducial markers, which represents a meaningful reduction in treatment burden for patients.
Future work will need to validate the approach prospectively and explore real-time implementation during treatment delivery. If successful, this could become a standard component of pancreatic cancer radiotherapy workflows.