Proton therapy is similar to conventional radiotherapy in that it is used to destroy tumours, but in principle has a finite range that spares surrounding healthy tissue. Such precision, however, can be thwarted by a patient's changing anatomy during a treatment plan. To avoid missing the tumour, radiotherapists will sometimes expand the target of radiation over the course of treatment, but studies have shown this method to be insufficient for proton therapy, as it can still end up delivering too little radiation to the tumour and too much radiation to healthy tissue, worsening side effects. The problem is more acute for lung cancer patients, as the contrast in tissue density in the thorax is much greater.

Researchers have been developing methods to accurately estimate how the target should evolve in proton therapy, and have so far taken into account respiratory motion, as well as setup and range errors in the proton therapy itself. Until now, however, a method to estimate a patient's anatomical variations has been lacking.

To create such a method, Jan-Jakob Sonke of the Netherlands Cancer Institute in Amsterdam and colleagues collected scans acquired for treatment planning and daily image guidance for a large cohort of patients. For each scan they used deformable image registration to measure, for each voxel, the displacement between treatment planning and treatment delivery.

In the next step, the researchers transferred the displacements to a reference anatomy, and averaged over the first treatment week to characterize the systematic variations. Finally, they used a method known as principle component analysis to extract independent "modes" of variations that are commonly found in the population.

"Similar methods have been applied [to] other parts of the body, but for the thorax, only the respiratory motions have been studied," said lead author Yenny Szeto of the Netherlands Cancer Institute. "With this, we can make statistical predictions of the systematic variations for a new patient to be treated for lung cancer radiation therapy… The planned dose will better predict the actual delivered dose."

Sonke, Szeto and colleagues are now planning to integrate their model into proton treatment plans. "[Our] model generates plausible instances of the geometric changes," says Szeto. "We can convert these changes into probability distributions of range uncertainties, which will be included in the probabilistic plan optimization."

Related stories

• Shrinkage predictions boost lung tumour dose
• Lung proton therapy demands robust plans
• Motion mitigation for scanned protons
• Motion analysis enhances predictive model