https://orcid.org/0009-0003-8068-0279
Figures
Abstract
Background
Scale-Invariant Feature Transform (SIFT) features are widely used in target recognition and tracking. This study aimed to exploit the robust performance of the SIFT algorithm to accurately calculate tumor tissue displacement during respiratory motion.
Materials and methods
A thoracic phantom was employed in this study. Eight synthetic nodules with different shapes and gray values were inserted into the phantom. First, the nodules were displaced upward, downward, leftward, and rightward by 10 pixels to simulate motion in different directions during respiration. The nodules were then rotated clockwise and counterclockwise by 5°, 10°, 30°, and 45°. Subsequently, the SIFT and cross-correlation algorithms were applied to analyze the phantoms. Finally, a t-test was used to assess differences among motion directions.
Results
The t-test result indicated no significant difference in the detection of moving phantom across all motion directions (p > 0.05). No statistically significant difference was observed between SIFT and cross-correlation in detecting translational motion. In contrast, the results obtained from rotating phantoms demonstrated that SIFT could effectively detect rotational distortion.
Citation: Pan Q, Luan S, Zhang F, Tian L, Wu G, Fan Y (2026) An improved Scale-Invariant Feature Transform-based method for detecting tumor tissue motion induced by respiration. PLoS One 21(6): e0330578. https://doi.org/10.1371/journal.pone.0330578
Editor: Mohamed Yacin Sikkandar, Majmaah University College of Applied Medical Sciences, SAUDI ARABIA
Received: September 9, 2025; Accepted: April 6, 2026; Published: June 24, 2026
Copyright: © 2026 Pan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data underlying the findings of this study are available from The Cancer Imaging Archive (TCIA): Gavrielides MA, Kinnard LM, Myers KJ, Peregoy J, Pritchard WF, Zeng R, Esparza J, Karanian J, Petrick N (2015). Data From Phantom FDA [dataset]. The Cancer Imaging Archive. DOI: 10.7937/k9/TCIA.2015.orbjkmux.
Funding: This study was supported by the Key Science and Technology Research Plan of the HeBei Provincial Health Commission (Scientific Research Fund Project No. 20201329). This study was approved by the Ethics Committee of Qinghe Central Hospital. The relevant code resources can be found at https://mp.csdn.net/mp_download/manage/download/UpDetailed.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Lung cancer is a disease with a high incidence and mortality rate. Nearly 70% of patients with cancer require radiotherapy during the course of treatment [1]. However, accurately determining tumor motion during respiration remains challenging. If tumor motion during breathing can be precisely quantified, radiation exposure to patients can be significantly reduced. At present, four-dimensional computed tomography (4DCT) is most commonly used in clinical practice to assess tumor motion; however, this approach is time-consuming and labor-intensive [2].
Previously, numerous studies have attempted to improve the accuracy of automatic motion detection. Emilie developed a tumor tracking technique capable of following a moving tumor in real time [3]. Christoph Hoog Antink used a high–sampling-rate terahertz homodyne spectroscopy system to estimate thoracic movement [4]. Evan et al. adapted the Microsoft Kinect v2 sensor to trace and record patient respiratory motion [5]. Mari Honda employed a breathing motion sensor for chest radiography during inspiration [6]. Melanie et al. used a self-gating 4D magnetic resonance imaging (MRI) workflow to assess the motion of the breast and organs at risk [7].
These studies have substantially improved the accuracy of tumor treatment and significantly reduced damage to surrounding organs at risk. However, chest tissue motion during respiration is not limited to simple translational displacement; a certain degree of rotational deformation also occurs with breathing. Owing to its invariance to rotation, the Scale-Invariant Feature Transform (SIFT) is widely used in target recognition and tracking [8,9], as well as in image segmentation and registration [10–12]. In this study, the robust performance of the SIFT algorithm was leveraged to accurately calculate tumor tissue displacement during respiratory motion. Both translational displacement and rotational angle were quantified to further enhance the accuracy of tissue motion assessment.
2. Materials and methods
The thoracic phantom employed in this study is shown in Fig 1. The phantom was scanned using a Philips 16-row CT scanner (Mx8000 IDT, Philips Healthcare, Andover, MA). Scans were acquired with a pitch of 1.00 and a slice thickness of 3.0 mm. Images were reconstructed with a slice thicknesses of 3.0 mm and a reconstruction increment of 1.5 mm. A total of eight synthetic nodules with different shapes and gray values were adopted in this experiment. The nodules included: (1) spherical, HU 100; (2) spherical, HU 100; (3) spherical, HU −630; (4) spherical, HU −630; (5) elliptical, HU 100; (6) lobulated, HU 100; (7) spiculated, HU −630; and (8) lobulated, HU −630. The nodules were displaced upward, downward, leftward, and rightward by 10 pixels to simulate translational motion, and the resulting configurations are shown in Table 1. The phantom position was repositioned between different nodule layouts. All CT images had a matrix size of 512 × 512 pixels for each slice [13]. Ten consecutive slices were analyzed for each nodule. Tumor position changes during respiration were simulated by moving the synthetic nodules by 10 pixels in each direction. This displacement was selected because the tumor motion range in this region is approximately 1 cm. The translated nodules are shown in Fig 2. In addition to translational motion, rotational motion was simulated by rotating the nodules by 5°, 10°, 15°, 30°, and 45°. The rotated nodules are shown in Fig 3. These rotation angles were selected because rib rotation during respiration typically ranges from 25° to 45°, whereas chest tissue rotation is generally smaller than that of the ribs.
A The phantom B Schematic diagram of Nodule Layout. C-F are CT images for the nodules, C shows nodule 1, D shows nodule 2, E shows nodule 3, F shows nodule 4.
(A) original image of nodule, (B) nodule moves downward, (C) nodule moved upward, (D) nodule moves rightside, (E) nodule moves leftside.
(A) original image of nodule, (B) nodule rotate to left 5°, (C) nodule rotate to right 5°, (D) nodule rotate to left 30° (E) nodule rotate to left 30°, (F) nodule rotate to left 45.
2.1 SIFT
The SIFT is based on gradient orientation histograms constructed in Gaussian scale space. The overall calculation process is shown in Fig 4. The main steps of the algorithm are described as follows:
A. Gaussian scale-space construction.
Gaussian scale space is used to identify extreme points for the calculation of feature descriptors:
where x and y denote the pixel coordinates of the image, ,
represents the Gaussian filter,
= 1, 2, 3, 4, 5 denotes the scale coefficient, and
represents the difference of Gaussian image. In this scale space, extrema are detected within the neighborhood of each point across the current, previous, and subsequent scale-space images.
In this study, the edge point of the known synthetic nodules were manually delineated. The contour points of the phantom nodules were then directly selected as candidate extrema for descriptor computation and feature matching.
B. Orientation assignment and descriptor generation.
The dominant orientation of each point is computed based on a gradient orientation histogram. The highest peak of the histogram is assigned as the main orientation of the key point.
Subsequently, a SIFT descriptor is constructed using the gradient magnitudes and orientations in the neighborhood of the detected key point. The scale of the key points is used to determine the radius of the search region. Gradient orientations are rotated according to the main orientation of the key point, and gradient histograms are then constructed around the key point. As a result, a 128-dimensional descriptor is obtained. In this study, a 30 × 30 pixel neighborhood around each key point was analyzed to identify the corresponding point in the translated phantom.
2.2 Cross-correlation
The cross-correlation function is a measure of similarity between two images. It calculates the cross-correlation in either the spatial or frequency domain to determine image movement in the spatial domain. In this study, the native MATLAB functions xcorr2 was used to compute 2D cross-correlation based on the dot-product method (Eq. 2), where the inputs are matrices.
where f denotes the original image, t denotes the template, represents the mean of the template, and
represents the mean of the template of
.
Because nodules 1–4 were spherical, no difference was observed before and after rotation. The images of translated nodules are shown in Fig 2. Nodules 5–8 were rotated, and the corresponding images are shown in Fig 3. The parameters of the rotated nodules are summarized in Table 2. Both the SIFT and cross-correlation algorithms were independently applied to detect these nodules.
2.3 Evaluation criteria
The matching rate (MR) of two corresponding algorithms was calculated to evaluate the accuracy of the two algorithms and is defined as follows:
where N1 denotes the number of detected points in image I1, N2 denotes the number of detected points in image I2, and Matchnum represents the number of matched points.
All programs were developed using MATLAB 2024a (The MathWorks, Inc., Natick, MA, USA). Statistical analysis were performed using SPSS software, version 20.0 (IBM Corp., Armonk, NY, USA).
3. Results
The statistical results of this experiment are summarized in Table 3. Table 3 illustrates the influence of different motion directions on the detection performance of the cross-correlation and SIFT algorithms. Most t-test results showed P < 0.05, indicating that no statistically significant differences were observed among the detection results of moving nodules in different directions. The t-test results for nodules showed no statistically significant differences for either the SIFT or cross-correlation algorithms (Nodule 5, rightward motion, P = 0.128; Nodule 6 downward motion, P = 0.312; Nodule 7, leftward motion, P = 0.534), indicating comparable detection performance across these movement conditions Fig 5 and 6. Fig 7 presents a comparison of detection results between the SIFT and cross-correlation algorithms across eight synthetic nodules. Fig 7 shows box plots of the detection results for rotating nodules at various angular positions. The evaluated nodules included Nodule 5 (elliptical, HU 100), Nodule 6 (lobulated, HU 100), Nodule 7 (spiculated, HU −630), and Nodule 8 (lobulated, HU −630).
A shows comparision of 45°, B shows comparision of 30°, C shows comparision of 10°, D shows comparision of 5°.
4. Discussion
The SIFT algorithm has been widely applied across multiple fields, with numerous studies reporting significant results [14–19]. In this study, the SIFT algorithm was applied to calculate chest tissue motion during respiration. Unlike previous approaches, which primarily focused on translational displacement, the present method additionally accounts for rotational distortion. By leveraging the rotation invariance of the SIFT algorithm, distorted changes in chest tissue motion during respiration were quantitatively assessed.
Table 3 presents the t-test results for different movement directions. Downward, rightward, and leftward translations were compared with upward translation. As shown in Table 3, for most nodules and movement directions, no statistically significant differences were observed in the detection results. However, statistically significant differences were identified in a limited number of cases for specific nodules and directions. Further inspection of the phantom images revealed that, during motion, some Nodules overlapped with rib structures, resulting in severe boundary blurring. This overlap likely contributed to inaccurate detection results.
As shown in Fig 6, under a translational displacement of 10 pixels, the detected motion distances for all nodules fluctuated around the expected value of 10 pixels. The fluctuation observed for Nodules 1, 2, 5, and 6 were significantly smaller than those for Nodules 3, 4, 7, and 8, indicating higher detection accuracy for Nodules 1, 2, 5, and 6. This difference can be attributed to their relatively higher gray values (HU 100), which resulted in greater contrast with the surrounding tissues. During the calculation process, conversion to a grayscale image was performed, which resulted in a significant difference in nodule gray values. The cross-correlation detection results for Nodules 1–4 are also superior to those for Nodules 5–8. This is likely because Nodules 1–4 are circular, and their boundaries undergo minimal deformation during motion, resulting in more accurate detection.
The detection results obtained using the SIFT algorithm exhibited greater fluctuations compared to cross-correlation. This is because the key point filtering step was omitted to ensure accurate determination of all key-point positions. The calculation of all key points was performed to capture the known edges of the selected synthetic nodule. As shown in Fig 7, the SIFT algorithm was able to compute the rotation angles of Nodules 5–8. For right-side synthetic nodules with relatively low gray values (HU −630), such as Nodules 7 and 8, the calculated results fluctuated more. In contrast, left-side synthetic nodules with higher gray value (HU 100) exhibited smaller fluctuations. Future work may improve stability by optimizing this aspect of the calculation.
In addition, smaller rotation angles were associated with relatively low errors. This is primarily because the rotation angle influences changes in gray values around the nodules. Larger rotation angles result in greater changes around the nodules, which increase the likelihood of error during detection.
Another important consideration is that neither the SIFT algorithm nor the cross-correlation algorithm provides satisfactory detection results in uniform media. Detection in such regions requires further development of specialized algorithms. In this study, thoracic tissue motion was simplified to translational and rotational components, which imposes certain limitations for future clinical applications. Because the model does not represent biological tissue, the algorithm may not capture all possible complexities of real tissue movement. Advanced imaging techniques, such as 4DMRI, may offer improved accuracy in these scenarios.
5. Conclusion
This study quantified both movement and rotation of a thoracic phantom and demonstrated that the SIFT algorithm can be used to calculate distortion caused by thoracic tissue during respiratory motion. Although the method showed promising performance, the accuracy of rotation angle estimation still requires further improvement.
Acknowledgments
The relevant code resources can be found at https://mp.csdn.net/mp_download/manage/download/UpDetailed.
The relevant data resources can be found at huo.yun.2008@163.com.
References
- 1. Wu F, Wang L, Zhou C. Lung cancer in China: current and prospect. Curr Opin Oncol. 2021;33(1):40–6. pmid:33165004
- 2. Habatsch M, Schneider M, Requardt M, Doussin S. Movement assessment of breast and organ-at-risks using free-breathing, self-gating 4D magnetic resonance imaging workflow for breast cancer radiation therapy. Phys Imaging Radiat Oncol. 2022;22:111–4. pmid:35619641
- 3. Carpentier EE, Karan T, McDermott RL, et al. Four-dimensional treatment planning strategies for dynamic tumor tracking. J Appl Clin Med Phys. 2024;25:e14269. pmid:38235952
- 4. Hoog Antink C, Schulz R, Rohr M, Wenzel K, Liebermeister L, Kohlhaas R, et al. Estimating thoracic movement with high-sampling rate THz technology. Sensors (Basel). 2023;23(11):5233. pmid:37299961
- 5. Silverstein E, Snyder M. Comparative analysis of respiratory motion tracking using Microsoft Kinect v2 sensor. J Appl Clin Med Phys. 2018;19(3):193–204. pmid:29577603
- 6. Honda M, Haruyama R, Sugiura Y, Ohara K, Mochigi K, Kono Y, et al. A breathing movement sensor for chest radiography during inspiration in children aged less than 3 years: a prospective randomized controlled study. Biosci Trends. 2020;14(3):200–5. pmid:32418922
- 7. Park C, Bae K, Jung J-H. Object recognition in infrared image sequences using scale invariant feature transform. Proceedings Volume 6968, Signal Processing, Sensor Fusion, and Target Recognition XVII; 2008. 69681P p.
- 8. Bal A, Alam MS. Automatic target tracking in FLIR image sequences using intensity variation function and template modeling. IEEE Trans Instrum Meas. 2005;54(5):1846–52.
- 9. Tomasi C, Kanade T. Detection and tracking of point features. Int J Comput Vis. 1991;9(3):137–54.
- 10. Lee PG, Heo J-Y, Suk B-Y, Yeou H, Park H. Scale-invariant object tracking method using strong corners in the scale domain. Opt Eng. 2009;48(1):017204.
- 11. Schwering PBW, Lensen HA, van den Broek SP, den Hollander RJM, van der Mark W, Bouma H, et al. Application of heterogeneous multiple camera system with panoramic capabilities in a harbor environment. SPIE Proceedings, 7481; 2009. 74810C p. http://doi.org/10.1117/12.833469
- 12.
Gavrielides MA, Kinnard LM, Myers KJ, Peregoy J, Pritchard WF, Zeng R, et al. Data From Phantom FDA [Dataset]. The Cancer Imaging Archive. 2015. https://doi.org/10.7937/k9/TCIA.2015.orbjkmux
- 13. Liu J, Yu X, Gong L, Yu W. Automatic matching of infrared image sequences based on rotation invariant. 2009 International Conference on Environmental Science and Information Application Technology; Wuhan, China; 2009. p. 365–8.
- 14. Wang A, Camargo A, Fevig R, Martel F, Schultz RR. Image mosaicking from uncooled thermal IR video captured by a small UAV. Proceedings of the IEEE Southwest Symposium on Image Analysis and Interpretation. IEEE; 2008. p. 161–4.
- 15. Zhang W, Zhao Y. An improved SIFT algorithm for registration between SAR and optical images. Sci Rep. 2023;13(1):6346. pmid:37072451
- 16. Gao J, Sun Z. An improved ASIFT image feature matching algorithm based on POS information. Sensors (Basel). 2022;22(20):7749. pmid:36298100
- 17. Deng Y, Deng Y. Two-step matching approach to obtain more control points for SIFT-like very-high-resolution SAR image registration. Sensors (Basel). 2023;23(7):3739. pmid:37050799
- 18. Huang H, Li X, Xiong W. A survey of image registration based on SIFT. Softw Guide. 2019;18:1–4.
- 19. Xia X, Xiang H, Cao Y, et al. Feature extraction and matching of humanoid-eye binocular images based on SUSAN-SIFT algorithm. Biomimetics. 2023;8(2):139. pmid:37092391