Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Evaluation of a Commercial Convolution Neural Network Based Auto-Segmentation Software

M Leyva*, D Wang, N Mcallister, A Gutierrez, R Tolakanahalli, Miami Cancer Institute, Miami, Florida


PO-GePV-M-16 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: The purpose of this study is to evaluate the accuracy of AI generated normal structure contours on various organs and to perform a comparative analysis to peer-reviewed clinical organ contours.

Methods: The AutoContour™ tool from Radformation™ was used to segment organs-at-risk (OAR) for head and neck, brain, prostate and mediastinum disease sites. The structure models used by AutoContour are trained using a deep-learning convolution neural network (CNN). AI generated structure sets were exported to Velocity (Varian Medical Systems) for evaluation. A total of 224 AutoContour structures were compared to clinically approved organ structures using Dice similarity coefficient and surface distance metrics. The metrics were generated in Velocity for site specific clinically relevant normal structures.

Results: Good agreement was found between the AI generated contours and manually drawn clinical contours. A mean surface distance of <5mm was found for 90% of the segment sample. Mean Dice coefficients of greater than 0.7 were found for 60% of the sample included in the analysis. A larger variance in Dice values was noted for structures with small volume (< 5cc) such as pituitary, chiasm, cochlea, etc. as well as for structures that were manually drawn solely in the area of interest such as spinal cord and esophagus. This variation is due to the differing superior-inferior extent of AutoContour generated structures as compared to manually drawn contours.

Conclusion: The Radformation AutoContour tool generates clinically acceptable normal structure contours and is efficient in removing inter-user segmentation variability that occurs with manual segmentation. The software completes segmentation in 15-60 seconds depending on the disease site saving valuable clinical resources. Although good quantitative agreement was found, further analysis will be performed using a randomized qualitative scoring method and a larger segment sample.


Segmentation, Validation, Treatment Planning


IM/TH- image Segmentation: CT

Contact Email