A priori patient-specific collision avoidance in radiotherapy using consumer grade depth cameras

In this study, we demonstrate and evaluate a low cost, fast, and accurate avoidance framework for radiotherapy treatments. Furthermore, we provide an implementation which is patient specific and can be implemented during the normal simulation process.

Four patients and a treatment unit were scanned with a set of consumer depth cameras to create a polygon mesh of each object. Using a fast polygon interference algorithm, the models were virtually collided to map out feasible treatment positions of the couch and gantry. The actual physical collision space was then mapped in the treatment room by moving the gantry and couch until a collision occurred with either the patient or hardware. The physical and virtual collision spaces were then compared to determine the accuracy of the system. To improve the collision predictions, a buffer geometry was added to the scanned gantry mesh and performance was assessed as a function of buffer thickness.

Each patient was optically scanned during simulation in less than 1 min. The average time to virtually map the collision space for 64, 800 gantry/couch states was 5.40 ± 2.88 s. The system had an average raw accuracy and negative prediction rate (NPR) across all patients of 97.3% ± 2.4% and 96.9% ± 2.2% respectively. Using a polygon buffer of 6 cm over the gantry geometry, the NPR was raised to unity for all patients, signifying the detection of all collision events. However, the average accuracy fell from 95.3% ± 3.1% to 91.5% ± 3.6% between the 3 and 6 cm buffer as more false positives were detected.

We successfully demonstrated a fast and low cost framework which can map an entire collision space a priori for a given patient during the time of simulation. All collisions can be avoided using polygon interference, but a polygon buffer may be required to account for geometric uncertainties of scanned objects.