Enhancement of megavoltage electronic portal images for markerless tumor tracking

Abstract Purpose The poor quality of megavoltage (MV) images from electronic portal imaging device (EPID) hinders visual verification of tumor targeting accuracy particularly during markerless tumor tracking. The aim of this study was to investigate the effect of a few representative image processing treatments on visual verification and detection capability of tumors under auto tracking. Methods Images of QC‐3 quality phantom, a single patient's setup image, and cine images of two‐lung cancer patients were acquired. Three image processing methods were individually employed to the same original images. For each deblurring, contrast enhancement, and denoising, a total variation deconvolution, contrast‐limited adaptive histogram equalization (CLAHE), and median filter were adopted, respectively. To study the effect of image enhancement on tumor auto‐detection, a tumor tracking algorithm was adopted in which the tumor position was determined as the minimum point of the mean of the sum of squared pixel differences (MSSD) between two images. The detectability and accuracy were compared. Results Deblurring of a quality phantom image yielded sharper edges, while the contrast‐enhanced image was more readable with improved structural differentiation. Meanwhile, the denoising operation resulted in noise reduction, however, at the cost of sharpness. Based on comparison of pixel value profiles, contrast enhancement outperformed others in image perception. During the tracking experiment, only contrast enhancement resulted in tumor detection in all images using our tracking algorithm. Deblurring failed to determine the target position in two frames out of a total of 75 images. For original and denoised set, target location was not determined for the same five images. Meanwhile, deblurred image showed increased detection accuracy compared with the original set. The denoised image resulted in decreased accuracy. In the case of contrast‐improved set, the tracking accuracy was nearly maintained as that of the original image. Conclusions Considering the effect of each processing on tumor tracking and the visual perception in a limited time, contrast enhancement would be the first consideration to visually verify the tracking accuracy of tumors on MV EPID without sacrificing tumor detectability and detection accuracy.

consideration to visually verify the tracking accuracy of tumors on MV EPID without sacrificing tumor detectability and detection accuracy. widely used as an on-line verification tool for treatment field in radiation therapy. [1][2][3][4] While the weight of the verification tool appears to be decreased after the emergence of kilovoltage (kV) imager mounted on a linear accelerator, EPID images still have their own advantages. For example, contrary to the fact that kV images are obtained from an x-ray source which is offset 90 o from the treatment beam (thus always questioning the accuracy of isocenter alignment), MV EPID images are produced from the treatment beams, thus eliminating the possibility of misalignment of targets. [5][6][7] Furthermore, EPID does not require additional dose when images are acquired during patient treatment.
One of the most useful applications of MV EPID is the markerless tumor tracking, in which the EPID is operated in cine mode and produces continuous portal images from the treatment beam. [8][9][10][11] The risk of pneumothorax from the marker implantation also prompted researchers to explore the possibility of markerless tumor tracking. 12 While most studies on tumor tracking have been focused on tracking algorithms, [8][9][10][11] the importance of consecutive image display on EPID cannot be neglected. The visual confirmation of the tracking accuracy during treatment greatly facilitates action against possible misalignment of beam aperture to the moving target.
However, electronic portal images are noisy, blurred, and show poor contrast in identifying patient's anatomy in detail. [13][14][15] Image enhancements have been used to obtain the optimal readability of EPID images, including contrast improvement, edge deblurring, and noise reduction. [16][17][18] A sequential application of these three image processing protocols was also proposed. 18 However, image processing during tumor tracking requires consideration of the time needed since the image should be processed without interfering with the next incoming image. Therefore, the most effective enhancement of EPID images is needed within a limited time. An effective algorithm showing several desirable improvements within a limited time is the best option. However, at first, comparative analysis of image processing results from each representative algorithm is an advantage.
Both unsharp masking for deblurring and contrast improvement have been the most widely used algorithms in medical imaging applications including noise reduction. In this study, therefore, these algorithms were individually applied on EPID portal images and the effect of each algorithm on image visualization was compared. In reality, contrast-limited adaptive histogram equalization (CLAHE) for contrast improvement and total variation (TV) deconvolution for unsharp masking were adopted. 16,[19][20][21][22][23] A median filter was employed for noise reduction. In addition, we were interested in determining whether the modified images affect the tumor auto-detection capability during tumor tracking. Therefore, by implementing a maskbased tracking algorithm, the tumor detection accuracy was also compared.

MV EPID aS-500 imager attached to a Varian 21 EX (Varian Medical
Systems, Palo Alto, CA, USA) was used with a 6 MV beam. Images with a resolution of 512 × 384 and pixel size of 0.784 mm were obtained. For image processing evaluation, the image of the QC-3 quality phantom (Standard Imaging, Middleton, WI, USA) placed at a distance of 35 cm from the EPID surface with the source to EPID distance of 140 cm was acquired. Furthermore, a single patient's setup image receiving whole brain treatment was processed to assess the clinical usefulness. To find out the effect of the enhanced images on the tumor tracking accuracy, two lung cancer patients' cine images were obtained during radiation treatments.
Human visual perception is known to be affected the most by contrast changes. 16,17,24 In this study, CLAHE algorithm was used for contrast enhancement. 16,21,23 The global histogram equalization treating the whole image occasionally yields an indiscernible result with a flat histogram. To remedy this drawback, adaptive histogram equalization (AHE) divides the image into subsections and applies histogram equalization to each subsection. However, it has a tendency to emphasize local histogram excessively and increases noise.  15 However, regularization method has been widely studied and established for the ill-defined problem, and the total variation is one of them. 19,20,22 Therefore, to obtain the real deblurred image preserving sharp edges, total variation deconvolution was adopted and the following was minimized, in which μ is a regularization parameter, βs are control parameters, and Ds are gradient operators. Additional details can be found in the reference. 22 The required 2D PSF was borrowed from the results of aS-1000 EPID notwithstanding different resolution of 1024 × 1024 with 0.4 mm pixel size, in which the Lorentzian function in normalized form 1 1þ x 2 þy 2 2 À Á 3=2 was suggested with a representative parameter value of λ = 0.5 for E = 6 MV. 15 Additionally, to examine the effect of reduced noise, a median filter for 3 × 3 pixels was also applied to original image set.
To assess the image quality of QC-3 phantom, three parameters of contrast value, signal-to-noise ratio (SNR), and blurriness were introduced. The contrast value was calculated as follows: Contrast ¼ IwÀIb Iw , where I w and I b were the average intensities within a 10 × 10 square pixels inside a white rectangular and a black rectangular, respectively, between two numbers '1' and '2' of the phantom image ( Fig. 1(a)). 26 The signal-to-noise ratio (SNR) was evaluated using the following formula: SNR ¼ Iw b , where I w was the same as above, and the σ b was the standard deviation inside the same black box also mentioned above. The blurriness was evaluated from the tilted line profile drawn in the Fig. 1(a). After fitting the measured profile data using a sigmoid-like Boltzman function, the width of 10-90% values between maximum and minimum was calculated.
Therefore, the smaller the width is, the less blurred the image is. The

| RESULTS
The original EPID image of QC-3 resolution phantom is shown in Fig. 1(a). TV regularized deblurred image and CLAHE processed one are shown in Fig. 1 As an example of clinical application, the lateral EPID setup image for the whole brain treatment was processed in Fig. 4. The central small white circle is the physical port film graticule. The structures including maxillary sinus and sphenoidal sinus are vague in the original image, which is a typical feature of MV images.
Deblurred result is displayed in Fig. 4 blurriness. However, the overall image quality after the deblurring procedure is similar to that of the original, and visual interpretation is slightly improved. CLAHE processed image shown in Fig. 4(c) has distinct features, thus providing easier identification of the anatomy.
The denoised image in Fig. 4(d) is soft and the object boundary is not clear. In Fig. 6, the longitudinal profiles through the arrow marked in the tracking algorithm and also on the mask shape in our case.
Therefore, the slightly higher detection power in CLAHE may not be definitive and needs further confirmation. The detection error was measured as the distance between the target and the manually selected feature-based ground truth ( Table 2). The overall accuracy is around 3 mm. However, deblurring improved the detection accuracy, and denoising resulted in worse error. CLAHE presents almost the same accuracy as that of original.

| DISCUSSION
In general, images obtained from devices are corrupted by noise and blurriness. Furthermore, images of MV EPID show poor contrast. For noticeable image improvement in EPID, it has already been suggested that a multistep image processing consisting of contrast enhancement, noise reduction and edge sharpening should be applied. 18 Diez et al. have improved the contrast of EPID images by introducing a combination of image manipulation algorithms, e.g., an inverse restoration filter and a local contrast enhancement. 28 Meanwhile, if the visual confirmation of the tumor tracking is straightforward, it is a great help for accurate beam delivery. However, for real-time application of EPID images in tumor tracking, the required processing time needs to be minimized and therefore, strategic approach is needed. Here, the 'real-time' refers to image processing completed and displayed without interfering with the next incoming image under cine mode. 10 The purpose of this study was to determine the enhancement resulting in substantially significant results for clinical application such as tumor tracking.
Deblurred images of the phantom and patients ( Fig. 1(b), Fig. 4(b) and Fig. 5(b)) display enhanced edges and demonstrate that our Meanwhile, CLAHE results shown in Fig. 1(c), Fig. 4(c) and Fig. 5(c) are captivating. The enhanced contrast makes these images very rich in structures and useful for clinical applications. The pixel value profile of CLAHE shown in Fig. 6 clearly demonstrates enhanced readability again. However, amplification of noise is also observed. Noise has been reported to be increased slightly during contrast improvement and deblurring, and incorporation of another filter was suggested to limit artifacts. 29,30 In our study, however, the increase of noise was neglected to simulate the minimal processing time, and noise influence on image interpretation was marginal. F I G . 6. Comparison of pixel value profiles of the original, deblurred, CLAHE, and denoised lung patient's images. These profiles were obtained along the arrow in Fig. 5(a). The superior-inferior (SI) and left-right (LR) tumor detection errors between manual and tracking in this study are larger than those of Anne et al., in which they adopted the same tracking algorithm as ours; however, the means with standard deviations in SI and LR were 1.0 ± 1.1 mm and 0.6 ± 0.6 mm, respectively. 8  ing. In this study, we did not thoroughly investigate the tracking algorithm itself, and its limitations are described in the reference. 27 However, we should mention one thing about the tracking accuracy in this study that the minimum position from MSSD for each frame was possibly influenced by the mask search region.
The detection accuracy was increased for deblurred set, and decreased for the denoised set. Yip et al., implemented a tracking algorithm (called STiL) based on matching of automatically detected multiple landmarks between the reference and the object image. 10,11 Their algorithm was reported to enhance the tracking accuracy compared with single template matching. They investigated the correlations between tumor tracking accuracy and blurriness, noise, and contrast by varying the number of frames to obtain the averages.
The results were that the accuracy increased by decreasing blurriness and increasing contrast. The relation with noise was not clear in the patient's images. These results are largely consistent with our findings except that noise was negatively correlated with detection accuracy in our study. The tracking algorithm in our study was based on pixel intensity and known to be suitable for low-contrast images. 27 Consistency in results from two different algorithms increased the reliability of our results.
We presumed that all the acquired images under cine mode should be processed with a better visibility. Toward this end, contrast modification might be the first option for high-frequency frame images. One of the recent studies utilized 12.87 frames per second for EPID-based tumor tracking, and also introduced a prediction process to compensate the delay time from image acquisition and tumor detection to adjustment of the treatment machine for accurate targeting. 11 It can be questionable whether all frames should be processed for visual perception. However, recognizing that visual verification of the tracking accuracy helps physicians/technicians intervene against possible irregularity, processing all frames is beneficial for prompt response. Of course, the automatic beam-off can be considered for the unexpected event.
Following this experiment, the effectiveness of total variation regularized deblurring was questioned for rapid image processing, which facilitates other applications of EPID image such as patient setup procedure since the setup accuracy can be affected by the detailed anatomy. However, repeated instant confirmation of tracking accuracy seems not to require the image detail. Therefore, unsharp masking, e.g., may be adequate even with its boundary shift mentioned above in the introduction. This problem of balance between image quality and effectiveness warrants further study. There are countless algo-

CONF LICT OF I NTEREST
No conflicts of interest.