Development of an automated region of interest selection method for 3D surface monitoring of head motion
Manually selected ROIs (solid surface) demonstrating (a) vendor recommended and (b) suboptimal ROIs for accurate image registration. 3D image (a) aims to include only surface areas stable over time and that rigidly move with the head. 3D image (b) is a much larger area including areas that may locally move depending on facial expressions or mandible motions.
(a) 3D surface image system (AlignRT) installed in a treatment room. AlignRT consists of three camera units (“pods”); one pod at each lateral side of the couch (1 and 3 in the figure) and one at the center above the foot of the couch (2 in the figure). (b) Each pod contains stereoscopic cameras, a projector to emit red light with pseudorandom speckle pattern, a texture camera to add texture on a reconstructed surface, and a flash camera for calibration.
Signed distance, d(P0, π) between a vertex point (P0) and a triangle plane (π) as explained in Eq. (1). d(P0, π) is defined as the distance of P0 projected onto n, the normal vector of π, to V0, one of three vertex point pertained in π.
(a) Manually selected initial ROI (reference image, entire face and neck) on the full image of volunteer 5 who kept the head stationary except for naturally blinking the eyes. Unselected stray points on the thorax were mistakenly included in the initial ROI by the system. (b) Real-time image (darker area) overlapped with the full image showing actual surface area captured in real-time delta mode. (c) Final ROI that the algorithm selected on the reference image. Patches represent the surface areas excluded by the algorithm.
The automatically selected ROI on an image of volunteer 5 for each selection criterion. Patches in (a) through (d) represent surface areas excluded from the final ROI based on criteria 1, 2, 3, and 4 (listed in Sec. II), respectively.
Final ROI for volunteers 3 [(a) and (b)] and 4 [(c) and (d)] when instructed to remain still except for local blinking of eyes. The surface area around the eyes was excluded from the final ROI, showing that the algorithm detected eye motion. All data were taken with 4 mm spatial resolution.
Final ROIs for volunteers 1 and 3 when instructed to open/close their mouths several times are shown in (a) and (b) (6 mm spatial resolution) and (c) and (d) (4 mm spatial resolution), respectively. The mandible area was excluded mostly based on the distance derivative criteria (criteria 3 and 4).
Final ROIs for volunteers 2, 4, and 5 when translated or rotated from the initial position by a large amount are shown in (a), (b), and (c), respectively. The algorithm either excluded most of the facial areas (a) and (b) or selected small ROIs with irregular patterns (c).
Final ROIs for images taken at (a) 40° (volunteer 4), (b) 55° (volunteer 3), and (c) 75° (volunteer 4) couch angles. The surface areas obscured by the gantry were excluded from the ROI. Note that (a) the ROI on the images captured at couch angle 40° was similar to that at couch angle 0°.
Translational motions (bottom) calculated using image registration for three different ROI types (top) on prior recorded 3D image series. Large suboptimal (top left), automatically selected (top middle), and vendor recommended (top right) ROIs correspond to blue, red, and green points, respectively. (a) Volunteer 3 remained still except for blinking of the eyes and (b) volunteer 1 remained still except for opening/closing of the mouth. Shaded areas correspond to where mouth motions were detected through visual inspection of image frames.
Final ROIs for volunteer 4 when instructed to open/close his mouth several times. The algorithm excluded most of the cheeks and sides with irregular patterns, similar to final ROIs shown in Fig. 8.
Mean ± standard deviation of translational motions for 11 data sets when the volunteers were instructed to remain still while either only blinking their eyes or opening/closing their mouth (motion types 1 and 2 in Sec. II C).
Article metrics loading...
Full text loading...