A new digital twin model can predict how a head and neck cancer patient's anatomy will shift weeks before treatment starts, letting doctors adjust proton therapy plans on the same day instead of a week later.
Researchers built a library of 302 anatomical changes recorded from 88 previously treated head and neck patients, then mapped that data onto each new patient's planning scan using a pretrained CT foundation model and deformable image registration. The process generated roughly 284 predicted CT scans per patient, each with its own set of tumor and organ contours. In a ten patient test, the team picked the predicted scan that best matched the actual treatment day anatomy, plus a less accurate one for comparison, and used both to build adaptive plans. After reoptimizing against real same day imaging, both plans recovered tumor coverage to about 98.2 to 98.3 percent, close to the 98.5 percent baseline set by physician approved replans.
Proton therapy for head and neck cancer runs six to seven weeks, long enough for tumors and healthy tissue to shift noticeably. Standard practice is to catch that drift with a scan, then spend about a week building a new plan. During that week, the patient is treated using an outdated map of their own anatomy. If forecasting tomorrow's anatomy today holds up in larger trials, it could compress that week down to a single clinic visit.
Ten patients is a feasibility study, not proof. One of the two plan variants showed a statistically significant drop in quality score versus baseline (p = 0.01). That's a reminder that forecasting is still an approximation, not a replacement for same day imaging.