Introduction
In radiotherapy, intra-fraction motion occurs during treatment and causes instant geometrical changes. In addition, it directly impacts the precision of dose delivery. Reference George, Keall and Kini1 Intra-fraction motion is the result of breathing or other physiological processes such as swallowing and the digestive system. Reference Langen and Jones2,Reference Watanabe, Isobe and Takisima3 The breast intra-fraction motion can be investigated during radiotherapy via portal imaging in cine mode. Reference Jones, Fitzgerald and Owen4–Reference Thomsen, Harrov and Fledelius8 Other methods include monitoring the motion of the breast’s surface using optical sensors or tracking the position of a radiopaque marker with fluoroscopic imaging. Reference Kinoshita, Shimizu and Taguchi9,Reference Price, Sharrock and Marchant10 The advantage of an electronic portal imaging device (EPID) over other methods is that it is a quick and easy technique to determine internal motions in the radiotherapy field. The technique acquired images from the region of interest without giving an additional dose to the patient during radiotherapy, and it mitigated concerns about the risk of infection due to the implantation of a marker. Reference Smith, Bloch and Harris7,Reference Kothary, Heit and Louie11
To quantify chest wall motion during irradiation, anatomical landmarks were manually marked on each cine EPID image. These landmarks included the central lung distance (CLD), central flash distance (CFD) and inferior central margin (ICM). Reference Fein, McGee and Schultheiss5 The range of motion measured based on changes in the position of each marker in all images. Reference Jones, Fitzgerald and Owen4–Reference Kron, Lee and Perera6,Reference Michalski, Atyeo and Cox12 Manual contouring of anatomical landmarks in EPID images suffers from several limitations. It is prone to interobserver variability, requiring substantial time and effort from clinicians, and introducing errors due to misidentification. Moreover, a nonautomated measurement system is not easily achievable for large datasets or low-contrast images. These drawbacks necessitate the development of automated motion analysis techniques based on computer vision and image processing algorithms to provide a more objective and efficient assessment of motion, enhancing patient positioning monitoring. Reference Mylonas, Booth and Nguyen13
Automatic motion estimation is a computer process in which two or more consecutive images are processed to determine motion vectors that describe the image transformation. One of the algorithms used for automatic motion estimation in consecutive images is the block-matching MATLAB algorithm. Reference Jain and Jain14,Reference Puri, Hang and Schilling15 While previous studies did not utilise block-matching algorithms to estimate motion in EPID images, this study presents a novel approach that employs this technique to automatically quantify intra-fraction motion in breast cancer patients.
Patients and methods
Ethical approval has been received by our Institutional Ethics Committee. Eighteen cases of breast cancer patients who were treated at the clinical department (Department of Radiation Oncology, Imam Reza Hospital, Mashhad, Iran) were selected for this study. All patients included in the study had undergone mastectomy surgery. The mean age of patients was 55·11 years (ranging from 43 to 69 years), with nine right and nine left breasts treated.
The three-dimensional treatment planning by computed tomography (CT) scan was used. The patients were positioned supine on the breast board and breathed normally during the CT scan and treatment. Patients were treated with six megavolts on an Elekta Precise linear accelerator equipped with a multileaf collimator and megavoltage EPID.
During the medial tangential treatment field delivery, EPID images were acquired in cine mode. Each image was acquired after receiving 30 monitor units of radiation during treatment. Approximately three to six images were obtained in each treatment step, depending upon the monitor unit. These images were saved in Joint Photographic Experts Group format and transferred to a personal computer and MATLAB software for processing and calculations. The pixel size and the scale of the EPID images were 1024 × 1024 pixels and 0·25 mm/pixel, respectively.
Calculation
A three-dimensional array was created in the first step to store and process each sequence of EPID images, followed by smoothing with a 10 × 10 pixels median filter for noise reduction. Reference Wang and Zhang16 The filter size was chosen according to the noise of images and the edge detection process. Patients’ skin surface and chest wall were segmented separately in all images using a Canny edge detector. Reference Canny17
For automated estimation of vector motion in cine images, the block-matching algorithm was applied to images. Reference Barjatya18,Reference Chen and Jou19 In this algorithm, each image of the sequence was divided into macroblocks. The range of some common block sizes is 4, 8, 16, 32 and 64 pixels. Reference Lu and Liou20 The choice of block size is a trade-off between accuracy and computational complexity. Smaller blocks provide more accuracy but require more processing power. Larger blocks offer faster processing but may introduce more motion estimation errors. Reference Yaakob, Aryanfar and Halin21 In this study, the image size and the macroblock size were 1024 × 1024 and 16 × 16 pixels, respectively. To track motion, each block was compared to its corresponding block and neighbours in the next image. This comparison estimates the direction and distance each block has moved, creating ‘motion vectors’ for the entire image. Mean absolute difference (MAD) and mean square error (MSE) are two standard metrics for measuring the similarity of the blocks of images. Reference Barjatya18
In this study, four algorithms are as follows: (a) exhaustive search with MAD criterion, (b) exhaustive search with MSE criterion, (c) three-step search with MAD criterion and (d) three-step search with MSE criterion were applied to all images. Reference Chung and Chang22
The intra-fraction motion of the skin surface and the chest wall equalled the average of the estimated motion vectors. Furthermore, the total intra-fraction motion of each patient was equal to the mean displacement of the skin surface and the chest wall.
Validation of calculations
Standard images with specified motion vectors were used to determine the accuracy of these four methods. An EPID image was initially selected as the primary standard, which was substituted by a geometric translation algorithm with a specified motion vector. This translated image was considered the next image of our sequence. The calculation process was conducted on these two images to estimate motion vectors. The error of estimation was calculated by the difference of the applied displacement vector from the average size of the calculated vectors:
$$\sigma = \left| {{d_{tr}} - \overline {{d_{es}}} } \right|$$
In the formula,
$\sigma $
,
${d_{tr}}$
and
${d_{es}}\;$
are the error of estimation, translation displacement and mean estimation displacements. The average of
$\sigma $
was calculated as the mean error of these four methods. The method with the lowest error was selected for the intra-fraction motion estimation of patients.
Statistical analysis
The correlation between the intra-fraction motion of the surface skin and chest wall was analysed by the Pearson correlation test. Descriptive statistics were calculated for individual patient motion and its correlation with age using Pearson correlation. Independent t-tests compared motion based on the side of the breast receiving radiation. Reference Sharma23 All statistical analyses were performed at the significance level of 95%. Version 20 of Statistical Package for the Social Sciences software was used for these analyses.
Results
Table 1 summarises the clinical characteristics of eighteen breast cancer patients, including their tumour side, age and mean movement.
Patients’ clinical features and mean intra-fraction motion estimation
A total of 168 images were obtained. The main EPID image was greyscale and noisy. In the first step, the skin surface and the chest wall were detected separately (Figure 1).
A sample of edge detection step. (a) Main EPID image obtained from a patient; (b) Result of the skin surface detection; (c) Result of the chest wall detection.
Figure 2 shows the motion vectors for the skin surface and chest wall. The motion vectors are represented as arrows, with the length and direction of the arrow indicating the magnitude and direction of the motion.
A sample of motion vector estimation step. (a) Result of the skin surface motion vector; (b) Result of the chest wall motion vector.
The mean error of estimation was computed to validate four methods of this algorithm (Table 2). The least error was obtained from the exhaustive search with the MSE criterion method (σ = 0·07 ± 0·04 mm). This method was employed to estimate the displacement vectors and statistical descriptions of parameters.
Mean error of four motion estimation algorithms
The mean intra-fraction displacements of the skin surface and the chest wall were 1·04 ± 0·24 and 1·10 ± 0·27 mm (Table 3). The maximum movement was calculated to be 3·5 mm. An analysis using Pearson’s correlation coefficient reveals a statistically significant linear relationship between the mean displacement of the skin surface and the chest wall (r = 0·61, p = 0·04) (Figure 3).
Mean and standard deviation of intra-fraction displacement of the skin surface and the chest wall, resulting from four methods of block-matching algorithms in the tangential radiation field (mm)
The linear relationship between the displacement of the skin surface and the chest wall.
The mean of the total intra-fraction motion of each patient was 1·04 ± 0·24 mm. Pearson’s correlation coefficient showed no correlation between the intra-fraction motion of patients and age (p = 0·17). The independent t-test revealed no relationship between the intra-fraction motions of patients and the side of the breast receiving radiation (p = 0·48).
Discussion
In the present study, we intended to quantify the mean chest motion during irradiation using an automated algorithm. The mean displacement of intra-fraction motion of the breast cancer patient’s skin surface and chest wall during radiotherapy was estimated to be 1·04 ± 0·24 and 1·10 ± 0·27 mm, respectively. We exploited the block-matching algorithm to precisely determine the mean movement in our participants. Unlike previous continuous portal imaging studies, Reference Jones, Fitzgerald and Owen4–Reference Thomsen, Harrov and Fledelius8 the current study not only automatically contoured the chest wall and skin surface for all frames but also calculated the displacement for the entire edge instead of focusing on a single point-based metric like CLD. In the study by Krone et al., Reference Kron, Lee and Perera6 a single observer manually measured CLD, CFD and ICM for 20 breast patients. To ensure the accuracy of these measurements, two patients’ images were evaluated by a separate observer. The average changes of these distances were reported as 1·06 ± 0·19, 0·98 ± 0·16 and 1·27 ± 0·45 mm, respectively. Jones et al. Reference Jones, Fitzgerald and Owen4 measured the average CLD and CFD displacements of 2·5 and 3·5 mm, respectively. Two independent radiation therapists conducted these measurements. A limitation of this study was the inability to visualise the CLD in one of the ten breast cancer patients. The patient’s chest wall was not clearly visible on the imaging scans, which prevented the CLD measurement.
Similar to the previously conducted investigations, the intra-fraction motion in our study was small (approximately 1 mm). Reference Fein, McGee and Schultheiss5,Reference Smith, Bloch and Harris7,Reference Thomsen, Harrov and Fledelius8 The reports of the American Association of Physicists in Medicine Task Group recommend using respiratory motion management techniques in breast cancer patients with tumour displacement of greater than 5 mm. Reference Keall, Mageras and Balter24 In our study, the maximum motion was less than 3·5 mm; therefore, the intra-fraction motions were negligible in conformal radiotherapy of breast cancer patients.
Although many studies reveal intra-fraction motion on portal images, few studies adopted automatic methods to estimate the quantity of these motions. In a study, Smith et al. Reference Smith, Bloch and Harris7 used software to apply a histogram equalisation and edge enhancement by the Sobel operator on all images. A physician manually contoured the lung, heart and breast in the first frame, and the software automatically determined the contours in subsequent frames. To track patient movement during radiotherapy, the CLD was calculated for each frame of the image in eight patients. These patients had a maximum CLD motion of 2·5 mm during their treatment course. Thomsen et al. Reference Thomsen, Harrov and Fledelius8 evaluated the residual positioning errors in breast cancer treatment after adjusting based on two orthogonal setup images taken during normal breathing. Through a ‘semi-automatic registration’ programme designed in MATLAB software, they determined the position of the chest wall on all EPID images. The registration process involved intra-fraction chest wall motion, with an average value of 2·0 ± 0·7 mm. The semi-automatic registration programme relies on user interaction to define landmarks or reference points, while the block-matching algorithm automatically compares image blocks to identify the best match. Both methods can effectively identify chest wall movement, but the semi-automatic approach offers greater flexibility and control over the registration process, while the block-matching algorithm is more efficient and less dependent on user input. Given the limitations of this algorithm in discovering each block size, it is recommended to evaluate and compare other motion estimation algorithms in future investigations.
Conclusion
We estimated the intra-fraction motion of the skin surface and the chest wall using cine EPID images during breast radiotherapy. A MATLAB algorithm was used to automatically identify the skin surface and chest wall, thereby reducing human error in the selection of anatomical landmarks. The steps are fully automatic, and due to the low contrast of the EPID images, there is no need for human intervention in drawing contours on consecutive frames. This is an automatic, repeatable and independent method. Our results showed that for breast cancer patients, both the range of intra-fraction motion of the skin surface and the chest wall was estimated to be approximately 1 mm.
Acknowledgements
All authors have approved the manuscript and agree with its submission.
The authors confirm that written informed consent has been obtained from the involved patients or if appropriate from the parent, guardian or power of attorney of the involved patients, and they have given approval for this information to be published in this case report.
Open access
