Percolation Based Cluster Models Fully Incorporating Spatial Dose Distribution in Assessment of Parotid Gland Radiation Induced Complication in Head and Neck Cancer Treatment

Abstract

Radiotherapy (RT) induced xerostomia is a common complication resulting from head and neck (HN) radiotherapy and can impact HN cancer survivor’s quality of life. The current normal tissue complication probability (NTCP) models are based on the degenerate dose volume histograms (DVHs), having suboptimal discriminative power and thus not routinely used for effective treatment planning although it may provide certain insights. We aimed to generate a novel cluster model that incorporates the 3D spatial parotid gland dose distribution to provide clinicians with reliable decision making tools. Seventy-five HN patients treated with IMRT and chemotherapy were retrospectively included in the study. The patients were chosen such that the full range of mean dose in the parotid gland was included to ensure no bias was introduced in NTCP modeling. The prescription was 66Gy to the PTV in 30 fractions. Clusters are aggregates of connected voxels whose doses are larger than a threshold and the model is specified with connectivity. The common choices are 1-, 2-, and 3-connectivity. Cluster metrics including the mean cluster size (MCS) and the largest cluster size normalized by the gland volume (NSLC) were evaluated and scrutinized against the benchmark NTCP. Two fitting options using maximum likelihood method to the Lyman-Kutcher-Burman (LKB) model were made: the volume parameter n fixed at 1.0 (mean dose model) and unrestricted (full LKB model). The fitted parameters TD50 (uniform dose leading to 50% complication probability) and m (steepness of the curve) were assessed with the LKB NTCP models using the available xerostomia data. Statistical analyses including bootstrapping were performed to evaluate the statistical significance of the results. NSLC was found to be a better metric than MCS with reference to the LKB model and strong correlation (average r=0.67) was observed between NTCP and NSLC for all three connectivity choices. The mean dose model returned the fitted parameters TD50 (39.9Gy) and m (0.4) from the NSLC distribution close to the published data. The threshold dose determining the clusters was around 40Gy, different from the clinically adopted mean dose of 26Gy. Drastically different TD50 and m values were extracted from the fits with the full LKB model, where the threshold dose would be near 27Gy if parotid was deemed as a parallel organization. Bootstrapping estimates confirmed the fitting outcomes. Cluster models incorporating the spatial dose information could serve as useful predictors for parotid gland RT induced complication. Parameterization of the clustering patterns showed different predictions than what current clinical protocols recommended. Further investigation is needed to validate the cluster models as useful clinical decision support tools.