Patient motion tracking for non‐isocentric and non‐coplanar treatments via fixed frame‐of‐reference 3D camera

Abstract Purpose As C‐arm linac radiation therapy evolves toward faster, more efficient delivery, and more conformal dosimetry, treatments with increasingly complex couch motions are emerging. Monitoring the patient motion independently of the couch motion during non‐coplanar, non‐isocentric, or dynamic couch treatments is a key bottleneck to their clinical implementation. The goal of this study is to develop a prototype real‐time monitoring system for unconventional beam trajectories to ensure a safe and accurate treatment delivery. Methods An in‐house algorithm was developed for tracking using a couch‐mounted three‐dimensional (3D) depth camera. The accuracy of patient motion detection on the couch was tested on a 3D printed phantom created from the body surface contour exported from the treatment planning system. The technique was evaluated against a commercial optical surface monitoring system with known phantom displacements of 3, 5, and 7 mm in lateral, longitudinal, and vertical directions by placing a head phantom on a dynamic platform on the treatment couch. The stability of the monitoring system was evaluated during dynamic couch trajectories, at speeds between 10.6 and 65 cm/min. Results The proposed monitoring system agreed with the ceiling mounted optical surface monitoring system in longitudinal, lateral, and vertical directions within 0.5 mm. The uncertainty caused by couch vibration increased with couch speed but remained sub‐millimeter for speeds up to 32 cm/min. For couch speeds of 10.6, 32.2, and 65 cm/min, the uncertainty ranges were 0.27– 0.73 mm, 0.15–0.87 mm, and 0.28–1.29 mm, respectively. Conclusion By mounting a 3D camera in the same frame‐of‐reference as the patient and eliminating dead spots, this proof of concept demonstrates real‐time patient monitoring during couch motion. For treatments with non‐coplanar beams, multiple isocenters, or dynamic couch motion, this provides additional safety without additional radiation dose and avoids some of the complexity and limitations of room mounted systems.

With the development of digital LINACs, advanced radiation therapy techniques with non-coplanar and non-isocentric beams, as well as beams with dynamic couch motion are emerging. These methods such as station parameter optimized radiation therapy (SPORT), 4-Pi, trajectory optimization in radiotherapy using sectioning (TORUS), trajectory modulated arc therapy (TMAT), and HyperArc TM1-5 offer enhanced dosimetry and more efficient treatments. Patient position monitoring during these new techniques is a key bottleneck to clinical implementation. Specifically, non-coplanar and non-isocentric beam arrangements prohibit the use of gantry mounted on-board imaging systems. Although the ceiling mounted x-ray systems can track patient position, they are limited to isocentric treatments. 6,7 Optical surface monitor systems, such as OSMS, AlignRT, C-RAD, and humediQ, are some of the solutions for non-coplanar treatments. 8 However, there are a few known issues with those surface imaging systems: (a) the systems are calibrated to accurately monitor the patient around the isocenter and may not be accurate for nonisocentric treatments, (b) some systems require manual couch angle input, which complicates dynamic couch treatments, (c) increased uncertainty occurs when one or more cameras are blocked by the gantry (blind spots), 9 and (d) the inaccuracy of the optical system increases for couch rotations, due to misalignment effects during the calibration process. 10 This proposed couch-mounted system allows the tracking data to be unaffected by the movement of the couch, blind spots and allows easy calibration due to the fact that a threedimensional (3D) camera is in the same frame-of-reference as the patient.
To truly benefit from complex arc deliveries or dynamic couch movement, 11 an efficient and reliable monitoring system must be in place. With computer vision, it is feasible to simultaneously monitor patient position and ensure the treatment beam delivery. The goal of this study is to develop a real-time patient position monitoring system for advanced unconventional beam trajectories to ensure a safe and accurate treatment delivery with the camera on the couch.
By putting the depth sensor in the same frame-of-reference as the patient, that is, camera on the couch, all the issues mentioned above become resolvable.

| MATERIALS AND METHODS
There are a range of commercially available depth cameras suitable for this purpose. In this work, a Kinect v2 depth camera (Microsoft, Redmond, WA, USA) was used for the patient position monitoring system. The camera uses infrared laser projectors and a monochrome CMOS sensor to measure the depth via the time of flight technique, that is, the time between sending and receiving IR pulses is converted to depth for each location on the 2D CMOS sensor. 12 The Kinect v2 camera has a maximum frame rate of 60 Hz and a depth range from 0.5 to 4.5 m. The firmware allows a desired depth range to be resolved into a maximum of 768 depth values. This gives voxel depth resolution from 0.68 to 5.8 mm. For our setup with 0.8 m depth range, the voxel depths are approximately 1 mm. The proposed monitoring system is being used with a relative concept, so the absolute isocenter position calibration is not required the absolute distance between the patient and the camera does not change during dynamic couch treatments, so the tracking accuracy will not degrade while the couch is moving. A check of the scaling and orientation of axes is needed for these relative measurements. The tracking system starts by storing the first 60 frames.
After this phase of the algorithm is complete, displacement readings will begin to output to the user. With this proposed system, the calibration is simplified because the 3D camera is in the same frameof-reference as the patient which makes it independent of couch movement.    Table 1 shows the 3D camera measured lateral, longitudinal, and vertical phantom displacements for planned shifts from 3 to 7 mm. All the Kinect measured shifts were within 0.4 mm of the OSMS system.

3.B | Uncertainty with dynamic couch movement
The tracking delta was measured at different couch speeds while the phantom is stationary on the couch (

| DISCUSSION
The overarching goal of this study is to develop a real-time monitoring technique for treatments with unconventional beam trajectories to ensure safe and accurate delivery. A couch-mounted depth camera offers fixed frame-of-reference tracking of patient motion independent of couch and gantry positions. This is projected to become yet more significant as gantry-based linac techniques strive to improve dosimetry by including complex couch motion.
For the accuracy of the software, the lateral displacement was more difficult to track due to the fact that the depth sensor does not intrinsically calculate lateral displacement. The software approximates lateral displacement using the movement of the pixels being Non-coplanar and non-isocentric treatments provide promising dosimetric results, however, without intra-fractional monitoring, it is hard to ensure the patient position throughout treatment. During treatments, the couch will be rotated for non-coplanar beams or moved away from the isocenter to treat extended volumes. Even for currently available surface imaging techniques such as OSMS and C-RAD, there is reduced accuracy for couch rotations. These inaccuracies are caused by the misalignment of the calibration plate during the calibration process. 10 By mounting a 3D camera in the same frame-of-reference as the patient provides a novel and feasible way to monitor patient position during the treatment. It not only eliminates the complicated calibration process but also the blind spots caused by gantry and the on-board imager systems. The continuous patient position monitoring afforded by a couch mounted camera can provide confidence that the planned dose is accurately delivered during the whole treatment.
For future work, it is not only critical to know if the patient position deviates from the plan, but also to send the correcting shift back to the treatment console. Through collaboration with vendors, it is feasible that the program will send out the delta to correct the patient position to the treatment console in real time via the motion management interface (Fig. 2). Collision of the machine with the patient during treatment, however, is still unsolved for these nonisocentric, and non-coplanar treatments. Most studies focus on collision prediction of the treatment plan, but not real-time monitoring. 13,14 Since the camera is on the couch, with the combination of the 3D computer-aided design of the linac, it is feasible to build a real-time collision avoidance systemif we know the location of patient relative to the machine, the collision can be avoided during the treatment.

| CONCLUSION
An affordable system for monitoring advanced non-coplanar, nonisocentric, and dynamic couch treatment strategies is demonstrated.
A motion tracking software with a camera mounted to the treatment table was designed and evaluated. By putting a depth sensor in the patients' frame-of-reference, dead spots can be eliminated. With this system, real-time surface monitoring during complex treatments with dynamic couch motion is feasible.

CONF LICT OF I NTERESTS
Authors declare no financial or other relationships, which may lead to a conflict of interest.