Volumetric imaging during head and neck radiation therapy using a Kalman filter tracking approach

Objective.Head and neck (H&N) volumetric imaging using a standard linear accelerator during treatment delivery could enable motion management strategies to improve geometric accuracy and avoid critical organs at risk (OARs). We developed a Kalman filter (KF) tracking approach to image 3D H&N anatomy in real-time using 2D x-ray images.Approach.Our method produces a 3D image corresponding to each 2D projection at the time the projection is acquired. A KF incorporates 3D-to-2D projection geometry to weight the prior 3D deformable vector field (DVF) and 2D DVF measurement generated by registering each 2D x-ray image to a 2D image generated using the planning computed tomography (CT) image. We performed a simulation study using the planning CT image and 5 minute surface motion traces for 12 H&N cancer patients. 2D intra-treatment images served as input to obtain the 3D DVF and resultant 3D images. We evaluated the KF tracking performance on the target and critical structures using centroid error, Hausdorff distance and Dice coefficient metrics. The results were compared to those from the current standard of care, no tracking.Results.Using KF tracking, we obtained a mean tumour centroid error of -0.2 ± 0.6 mm, 0.3 ± 0.8 mm, and 0.1 ± 0.4 mm in the left-right, anterior-posterior and superior-inferior axes respectively, compared to 0.9 ± 1.0 mm, -0.8 ± 1.1 mm, and -0.1 ± 1.2 mm with no tracking. The 95th percentile Hausdorff distance for the tumour contour was 1.2 ± 0.6 mm with KF tracking and 2.0 ± 1.2 mm without tracking. The Dice coefficients were 0.91 ± 0.08, 0.93 ± 0.04, 0.93 ± 0.05, 0.79 ± 0.18, and 0.89 ± 0.06 for the tumour, brainstem, left parotid gland, optic chiasm, and spinal cord respectively with KF tracking and 0.78 ± 0.17, 0.84 ± 0.06, 0.83 ± 0.08, 0.39 ± 0.14, and 0.70 ± 0.13 respectively with no tracking.Significance.Volumetric imaging with a standard linear accelerator was investigated using KF tracking with H&N patient data. The findings demonstrate accurate tracking of both the tumour and OARs, indicating potential applicability for the proposed method in a clinical translation setting.