Commissioning a Next-Generation Treatment Planning System for Adaptive Radiation Therapy

Abstract

Image-guided radiotherapy allows monitoring patients’ tumour response, weight loss, organ filling and shape changes (e.g. neck flexion, lung motion). Incorporating anatomic deformations into dose distributions may enable informed adaptive re-planning and improved therapeutic ratios. Deformable image registration is needed to track tissue motion on a voxel-by-voxel basis prior to adaptive re-planning. The purpose of this study was to quantify the geometric accuracy of novel deformable registration algorithms. A new treatment planning system (RayStation, RaySearch Laboratories), undergoing testing at our institution, contains 2 novel deformable image registration algorithms: a contour-based algorithm (driven by organ contours between image pairs) and a hybrid algorithm (additionally driven by maximizing the similarity between voxel grey-values). The algorithms were tested as follows. First, digital phantoms were created by warping 5 CT images (3 pelvis, 2 head and neck, 1 lung) with known deformations. The original images were then recovered using RayStation’s registration algorithms. Accuracy was quantified as the voxel-by-voxel displacement differences throughout the images. Second, patient datasets (n=74, liver/lung 4DCT and MR/MR, liver CT-MR, prostate MR-MR) were registered using both algorithms. Accuracy of the patient data was quantified by measuring the residual distances of contours and point landmarks between the deformed (i.e. registered) images and the actual patient images. Typical image resolution was 2-3 mm for CT, and 3-6 mm for MR. For the phantom CT data, the 95th percentile of errors for the contours distance-to-agreement ranged from 0.5-1.9 mm for the hybrid algorithm and 0.5-3.9 mm for the contour-based algorithm. Both algorithms had internal displacement errors >5 mm in nearly all cases, indicating contour alignment does not guarantee accurate internal organ registration. For the patient data, both registration algorithms achieved mean contour distance-to-agreements <2 mm for the liver/lung 4DCT, liver CT-MR and prostate MR, and <3.5 mm for the liver/lung MR. The internal organ target registration error was measured using anatomic point landmarks (median: 5 liver, 60 lungs, 3 prostate). The internal accuracy for the contour-based algorithm was consistently within the image resolution across all data (mean errors: 2.7-4.4 mm). The hybrid algorithm’s accuracy varied widely (means errors: 2.7-3.3 mm for liver/lung 4DCT, 3.9 mm for prostate MR, 8.2-12.0 mm for liver CT-MR and liver/lung 4DCT), exceeding the image resolution when MR data was used. RayStation’s novel deformable registration technology was evaluated. To ensure the registration errors are within the image resolution, the choice of algorithm depends on the anatomic site and modality of image-guidance (e.g. CT, cone-beam CT, MR). Deformable image registration technology allows dynamic patient motion to be incorporated into dose distributions, facilitating the clinical implementation of adaptive radiotherapy.