Contrast enhancement for portal images by combination of subtraction and reprojection processes for Compton scattering

Abstract For patient setup of the IGRT technique, various imaging systems are currently available. MV portal imaging is performed in identical geometry with the treatment beam so that the portal image provides accurate geometric information. However, MV imaging suffers from poor image contrast due to larger Compton scatter photons. In this work, an original image processing algorithm is proposed to improve and enhance the image contrast without increasing the imaging dose. Scatter estimation was performed in detail by MC simulation based on patient CT data. In the image processing, scatter photons were eliminated and then they were reprojected as primary photons on the assumption that Compton interaction did not take place. To improve the processing efficiency, the dose spread function within the EPID was investigated and implemented on the developed code. Portal images with and without the proposed image processing were evaluated by the image contrast profile. By the subtraction process, the image contrast was improved but the EPID signal was weakened because 15.2% of the signal was eliminated due to the contribution of scatter photons. Hence, these scatter photons were reprojected in the reprojection process. As a result, the tumor, bronchi, mediastinal space and ribs were observed more clearly than in the original image. It was clarified that image processing with the dose spread functions provides stronger contrast enhancement while maintaining a sufficient signal‐to‐noise ratio. This work shows the feasibility of improving and enhancing the contrast of portal images.

of kV imaging and treatment beam must be adjusted carefully for correct patient repositioning. Hence, quality assurance (QA) is more complex than that for MV imaging. 3,4 MV portal and MV-CBCT imaging can be performed in identical geometry with the treatment beam so that accurate geometric information can be provided. 5,6 However, MV imaging suffers from poor image contrast due to the lower difference of X-ray attenuation and larger Compton scattering compared with kV imaging. 7,8 Scatter correction methods that comprise scatter estimation and compensation have been reported. 9,10 The beam-scatter-kernel (BSK) superposition approach is the most promising in the scatter estimation method with respect to the computational efficiency. 11 The BSK is generally obtained using water rather than heterogeneous mediums and consequently it causes over-or underestimation of scatter photons.
In this work, the scatter estimation was performed in detail by Monte Carlo (MC) simulation based on patient CT data. Additionally, an original image processing algorithm was proposed for the scatter compensation. In this process, scatter photons were eliminated and reprojected as primary photons on the electronic portal imaging device (EPID). By the combination of the MC simulation and the proposed image processing, improvement and enhancement of the image contrast were attempted without increasing the imaging dose.
To assess the feasibility of the image processing, portal images with and without the scatter compensation were compared.

2.A | The proposed image-processing algorithm
The original portal image is generated by primary and scatter photons that occur on the EPID. The signal P o at the pixel coordinate (x, y) is the sum of the signals by primary photons P p (x, y) and scatter photons P s (x, y) as follows: To improve the image contrast, scatter photons must be eliminated. It has been reported that the signal of the EPID P is proportional to absorbed dose D to the scintillator. 12 However, the entire EPID signal is weakened by the subtraction process. To enhance the image contrast without increasing the imaging dose, we propose a means of reusing the scatter photons that were eliminated by the subtraction process. Eliminated scatter photons are made to reproject from scattering points to the EPID as primary photons on the assumption that Compton interaction did not take place. In consideration of the energy difference between the reprojecting photon hm r and the scatter photon hm s , the signal by the reprojecting photon DP r is estimated by the ratio of absorbed dose by the reprojecting photon DD(hm r ) to that by the scatter photon DD(hm s ).
where DD(hm r ) and DD(hm s ) are the absorbed dose to the scintillator by a photon with energy hm r and hm s respectively. They can be estimated by MC simulation for each photon. DP s is the signal by a scatter photon that is calculated as follows: Then, the signal P r (x, y) by n reprojecting photons can be calculated by the summation of DP r (x, y), Finally, the signal of the contrast enhanced image P c (x, y) is obtained by the sum of P p (x, y) and P r (x, y).
where w is the weight factor for adjustment of the contrast enhancement.
2.B | Simulation of absorbed dose to Gd 2 O 2 S:Tb by a photon The EPID is mainly composed of a Cu plate, Gd 2 O 2 S:Tb scintillator and a-Si photodiodes. In the simulation, equally spaced radial bins with Dr = 0.392 mm (1/2 of pixel width) were arranged, and the absorbed dose to Gd 2 O 2 S:Tb by a photon from the EPID surface was simulated using the DOSRZnrc code.
the Cu plate. It is estimated that 99.5% of the total signal is generated within the scintillator. 13 Electron trajectories are complicated within the EPID. To that end, the absorbed dose to Gd 2 O 2 S:Tb by a photon from the EPID surface was simulated using the DOSRZnrc code, 14 and the dose spread functions of photon energy hm and the radial distance from pencil beam r, DD(hm, r) were obtained. In the simulation, equally spaced radial bins with Dr = 0.392 mm (1/2 of pixel width) were arranged. The EPID consists of not only the main three layers but also low-density materials, such as air, paper and foamed body. In order to consider the spread of low-energy particles within low-density materials, the cut-off energies of photons and electrons were set to 10 and 521 keV. and 1400 mm, respectively. The field size was set to 40 cm 9 30 cm at the EPID, which has 512 9 384 pixels, the pixel size was 0.784 mm 9 0.784 mm and the signals were recorded as a 16-bit integer.

2.D | Photon sampling and image processing
3D-CT data of the thorax phantom was modeled in EGS5 to investigate photon trajectories in detail. 15 Figure 2 shows a simplified diagram of the photon sampling. The simulation geometry, e.g., SAD, SDD, and field size, was the same as the MV portal imaging described in 2.C. A 6 MV beam was reproduced according to the energy spectrum. When the photon reached the EPID surface, the coordinates (x, y) and energy hm of primary and scatter photons were sampled. Additionally, if it was a scatter photon, the coordinates (x, y) and energy hm r of the reprojecting photon were sampled on the assumption that the Compton interaction does not take place.
To calculate the absorbed dose D for each pixel, the deposit energy was sampled within r away from the incident point (x, y) according to the dose spread function DD(hm, r). Thus, D s (x, y) and D p (x, y) for each pixel were calculated by accumulation of the dose spread by scatter and primary photons, respectively. In the subtraction process, P s (x, y) and P p (x, y) were obtained according to eqs. (2) and (3). In the reprojection process, signals by a scatter photon DP s were calculated using DD(hm, r) and the signal by the reprojecting photon DP r was calculated by eq. (4) but DD(hm r )/DD(hm s ) was replaced with DD(hm r , r)/DD(hm s , r) in consideration of the dose spread. The image processing code was developed using the Qt Simplified diagram of the photon sampling for the image processing. The simulation geometry was the same as the MV portal imaging. 3D-CT data of the thorax phantom was modeled as a patient. A 6 MV beam was reproduced and photon trajectories were investigated in detail. When the photon reached the EPID surface, the coordinates (x, y) and energy of primary and scatter photons, hm p and hm s , were sampled. If it was a scatter photon, the coordinates (x, y) and energy hm r of the reprojecting photon were sampled on the assumption that the Compton interaction does not take place.
where P ref is the mean signal of the reference region that is indicated as a square in Fig. 3(a). The reference region was selected as the homogeneous background in the portal image. C(x, y) was evaluated along the line profile shown in Fig. 3        . Two images were displayed with the same window width, and gray levels were adjusted to be the same at the coordinates (x = 257, y = 26) where the spinous process was observed.
of the interaction was Compton scattering for the 6 MV X-ray beam.
By the MC simulation, it was calcified that 15.2% of the EPID signal was generated by scatter photons. Therefore, it is confirmed that image processing against scatter photons is required for MV imaging.  Figure 14 shows the SNR as a function of the weight factor w in the reference region of the P c image. The SNR increased until w = 1.0, then SNR decreased with increase in w. The SNR is improved by increasing the number history but it takes a longer processing time. Hence, to observe the thorax structures clearly while reducing the processing time, the optimal weight factor and sufficient number history are 1.0 and 1.5 9 10 10 respectively.
For other treatment site, the optimal parameters, i.e. the weight factor and the number history, could be different. As an example,

| CONCLUSION
Original image processing was proposed to improve and enhance the contrast of portal images. In the image processing, a combination of the subtraction and reprojection processes was performed using the photon sampling data. To improve the processing efficiency, the dose spread functions within the EPID were investigated and implemented on the developed code. In the contrast enhanced image, the structures were observed more clearly than in the original portal image. Consequently, this work demonstrates the feasibility of improving and enhancing the contrast of portal images.

ACKNOWLEDG MENT
This work was supported by a grant-in-aid (GC215, 2016) for research on priority areas of Tokyo Metropolitan University.

CONFLI CT OF INTEREST
The authors declare no conflict of interest.  Two images were displayed with the same window width, and gray levels were adjusted to be the same at the coordinates (x = 256, y = 20) between vertebral bodies. The contrast enhanced image with sufficient SNR was obtained when the number history was 1.0 9 10 10 .