MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automatic CT Segmentation for Radiotherapy Treatment Planning: How Good Is Good Enough?

W S Ingram*, L Dong, University of Pennsylvania, Philadelphia, PA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Automatic CT segmentation will always differ from manual segmentation, but metrics used to evaluate accuracy provide no information about the dosimetric impact. This study bridges that gap by simulating a large number of automatically-segmented structures to determine if there is any correlation.


Methods: Five similar head-and-neck VMAT plans were selected. Three dosimetric endpoints were evaluated: right parotid mean dose, brainstem D0.03cc, and nodal PTV D95%. To simulate automatic segmentation, random 3D deformations were applied to the reference structures. First, structure contours were converted to a polygon mesh. Next, a sparse grid of randomly-oriented deformation vectors was created. Finally, the deformations were interpolated and applied to each mesh vertex. This was repeated with varying deformation vector parameters (grid spacing 10,20,50mm; magnitude 2,5,10mm; 50 iterations each) to generate 400 deformed versions of each structure per plan. Two segmentation accuracy metrics were computed for the deformed structures: Dice similarity coefficient (DSC) and average surface-to-surface distance (ASD). These were compared to the difference in the dosimetric endpoints between the deformed and reference structures.


Results: Metric ranges were similar for each structure, with DSC=0.60-0.98 and ASD=0.12-3.74mm. The distributions were heavily skewed, with roughly 50% having DSC>0.9 and ASD<0.9mm. Dosimetric endpoint differences were similarly skewed, with median and maximum deviations of 2.6% and 47.0% for right parotid, 1.6% and 35.0% for brainstem, and 1.0% and 30.0% for nodal PTV. These changes were moderately correlated with DSC and ASD for right parotid (Spearman’s ? = -0.59 and 0.58) and brainstem (-0.55 and 0.54). The correlations were stronger for nodal PTV (-0.88 and 0.83).


Conclusion: Segmentation accuracy metrics do not correlate well with dosimetric impact, and the correlation varies between structures. Novel approaches must be developed to understand the dosimetric impact of segmentation accuracy. Future efforts should investigate metrics that characterize structure proximity to target volumes.

Keywords

Segmentation, Treatment Planning, Deformation

Taxonomy

IM/TH- image Segmentation: CT

Contact Email