Introduction: [¹⁸F]-Fluorodeoxyglucose PET has become an essential technique in oncology. Accurate segmentation is important for treatment planning. With the increasing number of available methods, it will be useful to establish a reliable evaluation tool.
Method: Five methods for [F]-fluorodeoxyglucose PET image segmentation (MIP-based, Fuzzy C-means, Daisne, Nestle and the 42% threshold-based approach) were evaluated on non-Hodgkin's lymphoma lesions by comparing them with manual delineations performed by a panel of experts. The results were analyzed using different similarity measures. Intraoperator and interoperator variabilities were also studied.
Results: The maximum of intensity projection-based method provided results closest to the manual delineations set [binary Jaccard index mean (SD) 0.45 (0.15)]. The fuzzy C-means algorithm yielded slightly less satisfactory results. The application of a 42% threshold-based approach yielded results furthest from the manual delineations [binary Jaccard index mean (SD) 0.38 (0.16)]; the Daisne and the Nestle methods yielded intermediate results. Important intraoperator and interoperator variabilities were demonstrated.
Conclusion: A simple assessment framework based on comparisons with manual delineations was proposed. The use of a set of manual delineations performed by five different experts as the reference seemed to be suitable to take the intraoperator and the interoperator variabilities into account. The online distribution of the data set generated in this study will make it possible to evaluate any new segmentation method.