MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Impact of a 3D Convolution Neural Network Method On Liver Segmentation: An Accuracy and Time-Savings Evaluation

NM Cole1*, H Wan1, J Niedbala2, YK Dewaraja3, A Kruzer1, D Pittock1, C Halley1, AS Nelson1, (1) MIM Software Inc., Cleveland, OH, (2) Michigan Medicine, Ann Arbor, MI, (3) University of Michigan, Ann Arbor, MI

Presentations

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

Room: AAPM ePoster Library

Purpose: The growing practice of treating primary and metastatic liver cancer with yttrium-90 microsphere selective internal radiation therapy (SIRT) and liver dominant neuroendocrine tumors with 177Lu-DOTATATE molecular radiotherapy has increased the need for fast, accurate, automated liver segmentation. Personalized dosimetry-based treatment planning requires accurate liver contours, however manual segmentation is time consuming. Previously, atlas-based segmentation was shown to greatly reduce the time burden, nonetheless, we sought to further decrease this time burden with a neural network approach. This study evaluates the accuracy and time-savings of a 3D Convolutional Neural Network (CNN) auto-segmentation method.

Methods: The CNN was trained with 108 contoured data sets then tested for auto-segmentation in 37 patient CTs from Y-90 and Lu-177 SPECT/CT and Y-90 PET/CT scans. The CNN outputs a liver contour from the entire CT volume. These contours were compared to the manually edited contours using the dice similarity coefficient (DSC), mean Hausdorff distance (HD), and the 95% max Hausdorff distance (95-HD).


Results: The average DSC was 0.97 ± 0.05. Averaged mean HD was 2.1 ± 3.7mm and averaged 95-HD was 12.2 ± 22mm with a median of 0mm. Time to edit the liver contour for each scan averaged 2.6 ± 2.4 minutes with 9/37 contours required no edits. The CNN also performed well with very low dose (=15 mAs) CTs with an average DSC of 0.94 ± 0.06 and editing time of 4.3 ± 2.4 minutes across 13/37 scans. Overall, the CNN showed reduced processing time compared to atlas-based (10.8 ± 4 minutes) and manual segmentation (34.8 ± 8 minutes) methods evaluated in a previous study.


Conclusion: The 3D Convolutional Neural Network can accurately auto-segment the liver, requiring little to no manual adjustment, leading to a significant decrease in processing time for dosimetry-based treatment planning in SIRT and molecular radiotherapy.

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

Keywords

Segmentation, Internal Dosimetry, Radiation Dosimetry

Taxonomy

IM/TH- Image Segmentation Techniques: Modality: CT

Contact Email