Optical tracking systems have become state-of-the-art in radiation therapy. Nevertheless, they still cannot cope with the properties of soft tissue. We are developing a non-invasive method to automatically detect the cranial bone's surface using only infrared laser light. Here, we present simulations to describe the influence of increased incident angles of the laser ray on the forehead's surface caused by replacing the original robotic setup by galvanometric scanning mirrors. We show that the usable area on the forehead (angles below 45°) will decrease, but that successful scanning is still possible when the scanning distance is increased from 300 to 500 mm. These results are confirmed when comparing optical scans acquired with the robotic setup and the scanning mirror assembly. Additionally, we could determine the needed depth of field for three subjects (based on high-resolution MRI scans). At a distance of 500 mm, it ranges from 37 to 86 mm.
To improve the sampling rate of the optical setup, either the robot can be moved faster or the laser beam has to be moved by some other means. For tracking in real time, a grid resolution of 20 by 20 points ten times per second - 4k points per second - is necessary to reliably track objects. Because this is impossible with a robot, we decided to add galvanometric scanning mirrors to our initial setup, which is described in more detail in [9]. The setup is aligned such that the optical axis of the high dynamic range camera (Andor Zyla) is parallel to the laser beam. A beamsplitter is used to overlay both axes. In addition, the optical axis of the camera is centered with respect to the scanning mirrors. Therefore, it is possible to look along the laser beam (inBeam or beam’s eye view) during the whole scanning process. Our simulations showed that this configuration prevents disturbances produced by an angle between the laser- and camera axes. Figure 1a shows a typical image of one point on the forehead captured with this setup. Similar to our simulations, we extract the optical features and estimate the skin thickness for every point on the forehead. In contrast to the robot setup, the angle between the surface normal of a point on the forehead and the laser beam increases significantly. We know from our simulations that reliable estimation of soft tissue thickness is possible for angles up to 45°. Therefore, the area AF, i.e. the number of points on the forehead where the incident angles are smaller than 45° (Fig. 2 marked in black), is decreased. Note that it is not possible to refocus the sample because the optical setup is fixed at a given distance. Consequently, the depth of field of the high dynamic range camera has to be higher than the vertical expansion of the forehead. Adapting the size of circles of confusion to the pixel resolution (1 pix = 6.5 µm) and the used lenses (Computar M7528-MP, f = 75 mm), the depth of field can be determined. However, it is necessary to simulate the scanning process to estimate the expected angles and the area AF to realize reliable measurements. For this purpose, we rotated forehead pointclouds extracted from MRI scans (resolution of 0.15 × 0.15 × 1 mm³) such that the x-y-plane was parallel to the face. Next, the central point on the forehead in x-y-direction was detected and its z-value was offset by the measuring distance d, the distance between our optical setup and the forehead. For our robot measurements, we had kept d at 300 mm [9]. Afterwards we determined the surface normals for each point of the forehead. Then we determined the vector between the laser source Q (Fig. 1b-c, black arrow in upper left corner) and the points of the forehead (laser vector, Fig. 1b, black lines). Finally, we computed the angle between the laser vectors and the surface normals. Additionally, we determined the needed depth of field (nDOF) for the given measuring distance d by subtracting the longest and the shortest laser vectors (which is valid due to our inBeamconfiguration).
For comparison, the same simulation was performed for the robot setup. In contrast to the scanning mirror setup, the xand y-components of Q were the same as the values of the point on the forehead. Additionally, d was kept constant, i.e. the z-value of Q was the z-value of the point offset by d. The results are shown in Fig. 1c. 3
We expected that the angular distribution changes when we use a galvanometer. This results in a smaller area AF which is usable for skin thickness estimations to improve optical tracking. As a consequence of these results we have to adapt our scanning grid resolution to the smaller area AF. The area AF is maximized when the measuring distance d is maximized. At the same time the nDOF decreases.
To realize a robust optical tracking system which can handle the problem of soft tissue artifacts, we must increase the measuring distance up to 600 mm. This would be the best tradeoff between resolution of the camera and depth of field. Further, it is now possible to scan non-rigid objects much faster without using a robot and performing several calibration steps as it was necessary in [9]. This will substantially improve the system’s accuracy Nevertheless, real measurements show strong angle dependencies for the galvanometric setup. Figure 3a shows an optical scan of a forehead using the robot setup. Figure 3b shows the same forehead scanned using a prototypic galvanometric setup. As shown by our simulations [8] we assume that the blur is mainly caused by higher incident angles in that area.
We demonstrated a theoretical estimation for fast optical scans to improve optical tracking. We showed results of simulations as they could influence the acquired data. The nDOF decreases for higher measuring distances d whereby the number of points in the area AF is maximized. As a consequence of these results, d should be maximized as long as it is possible to reliably extract optical features from the high dynamic range camera images. At the same time focus problems are handled. How far these angular influences are relevant for the results after post-processing the acquired data has to be evaluated in future work.
This work was supported by Varian Medical Systems Inc. (Palo Alto, CA, USA) and by the Graduate School for Computing in Medicine and Life Sciences, funded by Germany’s Excellence Initiative [DFG GSC 235/1]. 6 7