MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Auto Segmentation of Male Pelvis On CBCT Using 3D U-Net

R L.J. Qiu1*, T Ma1 , K Stephans1 , C Shah1 , A Godley2 , P Xia1 , (1) Cleveland Clinic, Cleveland, OH, (2) Miami Cancer Institute, Miami, FL

Presentations

(Sunday, 7/14/2019) 3:00 PM - 3:30 PM

Room: Exhibit Hall | Forum 6

Purpose: To generate and evaluate auto contours of male pelvis on daily pre-treatment cone-beam (CBCT) images using 3D U-Net

Methods: 16 IGRT prostate patients and 10 SBRT prostate patients with rectum balloons, had their bladder/prostate/rectum (B/P/R) contoured on the daily CBCTs. Manual contouring was assessed by the intra-observer reliability. Intra-observer compared initial contours with the repeated ones by the same observer on the same 10 CBCTs. The entire data set (300 CBCTs), was divided into the training set and testing test. A 3D U-Net convolution neural network (CNN) algorithm with fine-tuning was used. It connects the image features with the ground truth prostate, bladder and rectum contours on the training CBCT data, which provides the feedback to tune the parameters in the neural network. The testing CBCT images were used to check the accuracy of the CNN training. Once trained, its generated contours were compared with the manual contours. The comparisons were evaluated based on the dice similarity coefficient (DSC). The contouring accuracy was compared with the results from our previous atlas-based auto-segmentation method. A single GPU station was used.

Results: The DSC of B/P/R in the intra-observer tests are 0.95/0.91/0.90. The DSCs of the auto-contours generated by 3D U-Net improve from B/P/R : 0.9/0.9/0.83 to 0.95/0.9/0.89 (atlas-based method). The difference are statistically insignificant for all three organs while comparing to the intra-observer results. The computation time is around 24 seconds. However, if only using IGRT patients data to train the model and test on SBRT patients, the B/P/R are 0.93/0.75/0.65.

Conclusion: The auto-contours delineated by 3D U-Net CNN algorithms reach the same quality as the human expert contours. However, cautions need to be taken when using the trained neural network segment CBCTs with new patient anatomy.

Funding Support, Disclosures, and Conflict of Interest: This study was supported by a research grant from Elekta Inc and a grant from Nvidia Corporation.

Keywords

Segmentation, Cone-beam CT, Image-guided Therapy

Taxonomy

IM/TH- image segmentation: General (Most aspects)

Contact Email