Primary Tumor Extension Probability-based Prediction of Individualized Clinical Target Volume for Nasopharyngeal Carcinoma

Abstract

To establish a primary tumor extension probability knowledge base and develop a probability-based clinical tumor volume (CTV) prediction model for nasopharyngeal carcinoma (NPC). One thousand patients with newly histologically confirmed, non-disseminated NPC between January 2010 and December 2011 were included in this study. For all patients, primary gross tumor volumes (GTVp) displayed on pre-treatment magnetic resonance (MR) images were manually superimposed onto a single computed tomography (CT) scan. Tumor extension probability of each CT voxel was calculated to form a knowledge base. Then association rules mining algorithm was implemented to develop a probability-based CTV prediction model. One hundred cases were randomly selected from the patient cohort to evaluate similarity and variance between predicted CTV and expert delineation, which were manually completed by 2 radiation oncologists by consensus. Similarity was evaluated using Dice similarity coefficient (DSC), and variance using mean distances from GTVp to CTV in anterior, posterior, lateral, superior and inferior directions. Five multicenter radiation oncologists and another randomly selected 16 cases were included in a multicenter setting to evaluate the similarity between edited predicted CTV and expert delineation against that between manual delineation and expert delineation. Besides, inter-observer variance of 5 oncologists’ delineation was assessed using multi-observer DSC, and time spent on manual delineation and in editing predicted CTV were also reported. For 1000 patients, 58 developed local recurrence after median follow up of 67.3 months, and all recurrences were in-field or marginal. Probability heat map and equiprobability contours were generated to demonstrate the tumor extension probability. Based on proposed prediction model, CTV10% and CTV5% were defined to encompass corresponding subclinical disease. For 100 cases, mean DSC between predicted CTV and expert delineation were 0.89 ± 0.031 and 0.90 ± 0.020 for high-risk and low-risk CTV, indicating high similarity. Distances from GTV to predicted CTV were comparable to or slightly smaller than that to expert delineation in most directions. In multicenter evaluation, with assistance of predicted CTV, increased delineation accuracy was observed in all 5 oncologists, with mean DSC increased from 0.85 to 0.91 for high risk CTV and 0.87 to 0.93 for low risk CTV (P < 0.001 for 5 oncologists). Moreover, mean multi-observer DSC increased from 0.73 to 0.80 for high risk CTV, and 0.78 to 0.83 for low risk CTV (both P < 0.001) and average time spent on CTV delineation decreased from 30.3 to 14.9 min (P < 0.001). The probability-based predicted CTV was comparable to experienced expert delineation and might be enough for treating NPC. Predicted CTV significantly improved delineation accuracy in less experienced radiation oncologist and reduced inter-observer variance and time spent on CTV delineation.