Validation of automatic segmentation of ribs for NTCP modeling

Abstract

Background and purposeDetermination of a dose–effect relation for rib fractures in a large patient group has been limited by the time consuming manual delineation of ribs. Automatic segmentation could facilitate such an analysis. We determine the accuracy of automatic rib segmentation in the context of normal tissue complication probability modeling (NTCP). Materials and methodsForty-one patients with stage I/II non-small cell lung cancer treated with SBRT to 54Gy in 3 fractions were selected. Using the 4DCT derived mid-ventilation planning CT, all ribs were manually contoured and automatically segmented. Accuracy of segmentation was assessed using volumetric, shape and dosimetric measures. Manual and automatic dosimetric parameters Dx and EUD were tested for equivalence using the Two One-Sided T-test (TOST), and assessed for agreement using Bland–Altman analysis. NTCP models based on manual and automatic segmentation were compared. ResultsAutomatic segmentation was comparable with the manual delineation in radial direction, but larger near the costal cartilage and vertebrae. Manual and automatic Dx and EUD were significantly equivalent. The Bland–Altman analysis showed good agreement. The two NTCP models were very similar. ConclusionsAutomatic rib segmentation was significantly equivalent to manual delineation and can be used for NTCP modeling in a large patient group.