Click here to


Are you sure ?

Yes, do it No, cancel

Comparison of a 3D Convolutional Neural Network Segmentation Method to Traditional Atlas Segmentation for CT Head and Neck Contours

A Kruzer*, H Wan, M Bending, C Halley, D Darkow, D Pittock, N Cole, P Jacobs, AS Nelson, MIM Software Inc., Cleveland, OH


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

Room: AAPM ePoster Library

The purpose of this study was to compare the efficacy of a 3D convolutional neural network (CNN) segmentation algorithm with a traditional atlas segmentation approach when applied to head and neck contours on CT.

A CNN algorithm was trained on a group of 721 subjects from multiple institutions. Ten structures were included: mandible, brainstem, eye (L and R), optic nerve (L and R), optic chiasm, parotid (L and R), and spinal cord. Separately, an atlas image database was built from 20 subjects with the same structures included (these images were not included in the CNN training set). Both the CNN and atlas were used to segment the group of 20 atlas subjects from multiple institutions using manual segmentations as ground truth. A leave-on-out method was used for the atlas segmentation to prevent any image being used to segment itself. Mean Dice Similarity coefficient (DSC), mean distance to agreement (MDA), and mean 95% Hausdorff distance (HD95) were calculated for both segmentation methods.

The CNN had statistically significant improvements as follows (CNN, Atlas): Right Eye Dice (0.88, 0.85); Optic Chiasm Dice (0.30, 0.10), MDA (2.29mm, 3.47mm), and HD95 (5.22mm, 8.23mm); Left Parotid Dice (0.79, 0.70); and Right Parotid Dice (0.81, 0.74). However, the atlas performed better for the Brainstem MDA (1.56mm, 1.21mm) and for the Mandible MDA (0.70mm, 0.56mm) and HD95 (2.57mm, 1.91mm). No significant change was found for the remaining statistics.

For eight out of ten structures, the CNN method was found to be the same or better than atlas segmentation. The results seen in the mandible and brainstem seemed to be caused by a stylistic difference between CNN output vs. atlas output. In the future, we plan to analyze potential solutions to stylistic differences and analyze time-savings as well as accuracy.

Funding Support, Disclosures, and Conflict of Interest: Aaron Nelson is a part-owner of MIM Software Inc. Alexandria Kruzer, Hanlin Wan, Michael Bending, Christopher Halley, Dan Darkow, Dane Pittock, Natalie Cole, and Paul Jacobs are employees of MIM Software Inc.


Segmentation, CT, Pattern Recognition


IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email