Encrypted login | home

Program Information

Evaluation of the Quality of Auto Segmentation with Potential Application in Auto Treatment Planning

no image available
C Yan

C Yan*, M Huq , University of Pittsburgh Cancer Institute, Pittsburgh, PA


SU-I-GPD-J-21 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall

Purpose: Automatic segmentation can reduce interobserver variability and results in dramatic increase in the efficiency and consistency of individual plans. It can also make treatment planning faster and more cost effective. The aim of this study is to evaluate the quality of auto-generated contours by comparing their dosimetric characteristics with manually delineated contours.

Methods: Three lung patients’ CT images have been selected and contours for three critical organs (lung, heart and spinal cord) have been automatically generated. The quality of the contours were compared with the manually-delineated contours. Clinically used treatment plans for the lung patients have been used to evaluate the dosimetric characteristics of the auto-generated contours and these dosimetric characteristics were compared with the manual delineations. Maximum cord dose, V20 of lung and mean lung dose (MLD), and V25 and V30 of heart were compared.

Results: Maximum cord dose for auto contours were 49.2Gy, 20.7Gy and 21.2 Gy, and were 46Gy, 21.3Gy and 16.3Gy for manual contours. V20 for auto contours were 11.9%, 10.7% and 20.2%, and were 11.9%, 10% and 20.7% for manual contours. MLD for auto contours were 6.4, 8.8 and 11.3Gy, for manual contours, the dose were 6.3, 8.5 and 11.5Gy. V25 and V30 of heart for patient 2 differ the most between auto contours (bot are 0s) and manual contours (6.8% and 3.6%, respectively).

Conclusion: Auto contours for lung matched manual contour best when V20 and MLD were considered. Congruence of maximum cord dose between auto contours and manual contours was worse than lung, but V25 and V30 of heart differ the most between auto and manual contours.

Contact Email: