MENU

Click here to

Ă—

Are you sure ?

Yes, do it No, cancel

Improved Auto-Segmentation for CT Male Pelvis: Comparison of Deep Learning to Traditional Atlas Segmentation Methods

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

Presentations

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

Room: AAPM ePoster Library

Purpose:
To compare the accuracy of a 3D Convolutional Neural Network (CNN) based automated contouring method to an atlas for male pelvis on CT.

Methods:
A CNN trained on 320 expert-segmented images from multiple institutions was compared to an atlas containing 35 separate images. Both algorithms were run on the same 35 atlas images and compared to gold-standard manual segmentations for 6 anatomical structures: prostate, bladder, rectum, both femurs, and seminal vesicles. A leave-one-out analysis was used on the atlas to avoid using images to segment themselves. The Dice score, mean distance to agreement (MDA), and the Hausdorff 95th percentile distance (HD95) were calculated for both methods. Statistically significant improvement was calculated via a two-sample t-test on each structures’ statistics.

Results:
The CNN had statistically significant improvements as follows (CNN, Atlas): Prostate Dice (0.82, 0.71), MDA (2.33mm, 3.50mm), and HD95 (6.23mm, 9.21mm); Bladder Dice (0.94, 0.82), MDA (1.09mm, 3.96mm), and HD95 (3.88mm, 15.74mm); Seminal Vesicles Dice (0.71, 0.51), MDA (2.24mm, 4.17mm), and HD95 (7.17mm, 11.64mm); and Rectum Dice (0.77, 0.69). However, the atlas performed better for the Rectum HD95 (27.47mm, 15.24mm) and Right Femur MDA (1.20mm, 0.93mm) and HD95 (5.87mm, 3.98mm). No significant change was found for the remaining statistics.

Conclusion:
The CNN method was frequently more accurate than the traditional atlas method. However, the atlas performed better on the rectum and right femur. This was felt to be related to variability in contouring styles between the CNN outputs and the referenced gold-standard segmentations. 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: Funding for this project was provided by MIM Software Inc. Aaron Nelson is a part-owner of MIM Software Inc. All authors are employees of MIM Software Inc.

Keywords

CT, Segmentation, Pattern Recognition

Taxonomy

IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email