Tracking lung tumour motion using a dynamically weighted optical flow algorithm and electronic portal imaging device

Abstract

This paper investigates the feasibility and accuracy of using a computer vision algorithm and electronic portal images to track the motion of a tumour-like target from a breathing phantom. A multi-resolution optical flow algorithm that incorporates weighting based on the differences between frames was used to obtain a set of vectors corresponding to the motion between two frames. A global value representing the average motion was obtained by computing the average weighted mean from the set of vectors. The tracking accuracy of the optical flow algorithm as a function of the breathing rate and target visibility was investigated. Synthetic images with different contrast-to-noise ratios (CNR) were created, and motions were tracked. The accuracy of the proposed algorithm was compared against potentiometer measurements giving average position errors of 0.6 ± 0.2 mm, 0.2 ± 0.2 mm and 0.1 ± 0.1 mm with average velocity errors of 0.2 ± 0.2 mm s−1, 0.4 ± 0.3 mm s−1 and 0.6 ± 0.5 mm s−1 for 6, 12 and 16 breaths min–1 motions, respectively. The cumulative average position error reduces more rapidly with the greater number of breathing cycles present in higher breathing rates. As the CNR increases from 4.27 to 5.6, the average relative error approaches zero and the errors are less dependent on the velocity. When tracking a tumour on a patient's digitally reconstructed radiograph images, a high correlation was obtained between the dynamically weighted optical flow algorithm, a manual delineation process and a centroid tracking algorithm. While the accuracy of our approach is similar to that of other methods, the benefits are that it does not require manual delineation of the target and can therefore provide accurate real-time motion estimation during treatment.

Simultaneous object detection and segmentation for patient‐specific markerless lung tumor tracking in simulated radiographs with deep learning

“…Besides DL-based approaches, numerous other studies have investigated non-DL-based image tracking techniques for markerless lung tumor tracking, including well-established methods such as image registration, 40 template matching, 41 and optical flow 42,43 . Other less common methods like short arc tumor tracking 44 and hidden Markov model 45 have also been proposed.Some studies were exclusively conducted on digital or experimental phantoms with tumor motions that were mechanically controlled following either simulated breathing patterns 42 or patient-measured motion traces. 45,46 They were often able to achieve sub-millimeter accuracy, which might potentially be biased by the often simpler geometry of phantom anatomy and tumor shape/size.…”

“…Besides DL‐based approaches, numerous other studies have investigated non‐DL‐based image tracking techniques for markerless lung tumor tracking, including well‐established methods such as image registration, 40 template matching, 41 and optical flow 42,43 . Other less common methods like short arc tumor tracking 44 and hidden Markov model 45 have also been proposed.…”

“…Other less common methods like short arc tumor tracking 44 and hidden Markov model 45 have also been proposed. Some studies were exclusively conducted on digital or experimental phantoms with tumor motions that were mechanically controlled following either simulated breathing patterns 42 or patient‐measured motion traces 45,46 . They were often able to achieve sub‐millimeter accuracy, which might potentially be biased by the often simpler geometry of phantom anatomy and tumor shape/size.…”

Vector-field dynamic x-ray (VF-DXR) using optical flow method in patients with chronic obstructive pulmonary disease

“…Tanaka et al [ 14 ] have reported DXR had the potential to visualize vector-field of lung motion using the cross-correlation method. OFM was initially proposed as one of the markerless tracking methods for lung tumors in chest radiographs [ 26 ]. OFM can visualize motion velocity with higher spatial resolution than the cross-correlation method [ 27 ].…”

Development of AI-driven prediction models to realize real-time tumor tracking during radiotherapy

“…Although neural networks and fuzzy logic exhibit a satisfactory performance, in the research conducted with the use of these techniques, only 3 and 20 patients were tested, respectively [18,20,21]. In the work presented by Teo et al, only the tumor motion in the superior-inferior direction was predicted with the input of internal target position with an electronic portal imaging device (EPID) at the frequency at 7.5 Hz [19,22]. In addition, their approach cannot be performed if the internal target positions are invisible on EPID.…”