Room: AAPM ePoster Library
Purpose: evaluate the accuracy of a commercial atlas-based auto-segmentation tool for pediatric patients receiving craniospinal irradiation (CSI).
Methods: total of 111 pediatric CSI patients with 18 structures manually contoured on planning CT by dosimetrists and approved by radiation oncologists were divided into 3 age groups for constructing 3 atlas libraries: Age 3-8, 9-14 and 15-21 years (48, 39 and 24 patients, respectively). CT images were iteratively reconstructed with 0.98 mm pixel and 1-2 mm slice spacing. Contours for each patient were auto-segmented and compared with manual contours without including the subject in the age-dependent atlas library. A statistical validation between manual and auto-segmented contours was performed by using the Dice similarity coefficient (DSC).
Results: mean DSC for all contours was 0.73±0.23. Structure volume and DSC were moderately correlated (R² = 0.51). Structure volumes = 550cc had average DSC values = 0.95. Structures with the highest mean DSC were the cranium CTV and the brain (mean volume = 1415cc, mean DSC = 0.99; mean volume = 1397cc, mean DSC = 0.98, respectively). The smallest mean DSCs were for the optic chiasm and the optic nerves (mean volume = 0.58cc, mean DSC = 0.37; mean volume = 0.59cc, mean DSC = 0.56, respectively). Structure volume was not the only determinant factor for atlas performance. Small structures, e.g., eyes (mean = 7.06cc), still achieved a high DSC (mean = 0.84). There was no significant difference in mean DSC between the 3-8 age group and the 9-14 age group (0.78 and 0.75, respectively; p = 0.1); however, the 15-21 age group had a significantly lower mean DSC (0.71; p < 0.01).
Conclusion: have created three age-dependent atlas libraries for auto-segmenting structures in children receiving CSI. The atlas segmentation tool performed well in terms of DSC for pediatric structures with volumes = 550cc.