Analysis of lung tumour motion is important to inform treatment margins in radiotherapy planning. With the developments of MRI simulators and MRI-guided linear accelerators (MR-Linac)^1,2,3 real time lung tumour tracking is one of the key applications of interest.

A number of studies have used MRI for target tracking: Rapid imaging in a single plane can be used to monitor motion during free breathing but this cannot account for through-plane motion⁴. Simultaneous excitation of two orthogonal scan planes can be used to inform 3D location albeit at a slower rate⁵. Another approach is to use template matching of the real-time dataset to a priori 3D image data as has been done in the liver⁶. Lung tumour motion is more challenging due to respiration and variations in motion that are dependent on location and changes during treatment. Nevertheless, template matching has been shown to be useful in lung tumours⁷. The purpose of this study was to examine an advanced template matching method called Texture Reformatted Angle Correlation (TRAC), which utilises a semi-automated selection region for the target and the production of a library of multi-angled templates.

All imaging was conducted on a 3T MRI scanner (Siemens Skyra) using an 18 channel body receiver coil and a 32 channel spine coil. The methodology involves 5 steps outlined in Figure 1. As part of our existing lung protocol, a 3D volume (HASTE) sequence is first acquired during free breathing with a phase navigator echo (TE/TR=92/2000ms, resolution = 1.25 x1.25 x 4.0 mm). This dataset is used to generate a library of target templates from multi-planar reconstructions. In contrast to previous studies, local rigidity of the template matching method⁵ is maximised by using a segmented region of interest (Figure 2). The second improvement of the TRAC method is extracting tumour templates from both orthogonal planes with angle α = 0° and oblique planes (1°, 5°, 10°, -3°, -5°, -10°).

A real-time motion sequence is subsequently acquired using a rapid single sagittal plane 2D TrueFISP acquisition (TE/TR = 1.44/419 ms; in-plane resolution 0.70 x 0.70mm and 4 mm slice thickness) with 30 phases at 500 ms per frame. A searching area was defined which was twice the size of the tumour in two respective dimensions to determine a template per frame. This was then matched to the appropriate library template by calculating a cross-correlation coefficient⁸ to determine the tumour location at each frame.

An in-house developed MR compatible motion platform was used to simulate breathing and a piece of melon was used as the tracking target. A stationary 3D scan of the phantom was acquired as the gold standard volume. The phantom was then imaged during motion (±2cm, 0.1 Hz) to verify the methodology.

The tests with both patient and phantom datasets produced accurate tracking results. Figure 3 shows examples of the TRAC method; the green rectangle is the locus of interest and the red oval indicates the matched target. The individual matched templates for each frame are used to measure motion in three dimensions. For this patient, viewing these images in a cine mode revealed predominantly superior-inferior with some out-of-plane motion. Figure 4 plots the slice position and angle for each frame. In all frames except frame 1 and 2, the tumour stayed within the planes located at +0.77(mm) and -0.48(mm) and variations between -5°/-10° rotation.

Results from the phantom test showed repeatable in-plane and through-plane tracking over the course of three cycles.

The cross-correlation coefficient ranged from 0.74 to 0.83 in the patient test, and 0.90 to 0.92 in the phantom test, which is explained by the simpler phantom texture.

In this study, an improved tracking method (TRAC) has been described for 3D motion tracking of lung tumours and tested using in vivo and phantom data. The method can cope with through-plane translations and rotation of the target to accurately measure motion from a single plane acquisition. The method can be adopted for MRI planned radiotherapy and shows promise for MR-Linac tracking. Future work is ongoing to look at speeding up the post-processing efficiency and the dependency on temporal and spatial resolution. This approach is particularly attractive for MR-Linac systems where a fast single plane acquisition image can be used as the basis for tracking.