Impact of the accuracy of automatic tumour functional volume delineation on radiotherapy treatment planning

Over the past few years several automatic and semi-automatic PET segmentation methods for target volume definition in radiotherapy have been proposed. The objective of this study is to compare different methods in terms of dosimetry. For such a comparison, a gold standard is needed. For this purpose, realistic GATE-simulated PET images were used. Three lung cases and three H&N cases were designed with various shapes, contrasts and heterogeneities. Four different segmentation approaches were compared: fixed and adaptive thresholds, a fuzzy C-mean and the fuzzy locally adaptive Bayesian method. For each of these target volumes, an IMRT treatment plan was defined. The different algorithms and resulting plans were compared in terms of segmentation errors and ground-truth volume coverage using different metrics (V-95, D-95, homogeneity index and conformity index). The major differences between the threshold-based methods and automatic methods occurred in the most heterogeneous cases. Within the two groups, the major differences occurred for low contrast cases. For homogeneous cases, equivalent ground-truth volume coverage was observed for all methods but for more heterogeneous cases, significantly lower coverage was observed for threshold-based methods. Our study demonstrates that significant dosimetry errors can be avoided by using more advanced image-segmentation methods.