MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Multi-Disciplinary Scientific Session
  5. Automatic Detection of Contouring Errors Using Convolutional Neural Networks

Automatic Detection of Contouring Errors Using Convolutional Neural Networks

D Rhee1*, C Cardenas2 , H Elhalawani3 , R McCarroll4 , L Zhang5 , J Yang6 , B Beadle7 , L Court8 , (1) MD Anderson Cancer Center, Houston, TX, (2) University of Texas MD Anderson Cancer Center, Houston, TX, (3) UT MD Anderson Cancer Center, Houston, TX, (4) University of Maryland Medical Center, Baltimore, MD, (5) MD Anderson Cancer Center, Houston, TX, (6) MD Anderson Cancer Center, Houston, TX, (7) Stanford University, Stanford, CA, (8) UT MD Anderson Cancer Center, Houston, TX

Presentations

(Monday, 7/15/2019) 4:30 PM - 6:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: To (1) develop a deep learning-based contouring tool for head-and-neck normal tissues and (2) use this tool to detect contouring errors in a clinically used autocontouring system.

Methods: (1) We developed a two-step convolutional neural network (CNN)-based autocontouring tool for 11 normal structures of the head-and-neck on CT images. In the training data (3,495 previously treated patient CT scans in total, 825-1702 patients per structure), incorrect contours were identified using a semi-automated method to minimize the curation time. The Inception-ResNet-V2 architecture was used to classify the existence of organs-at-risk in each CT slice. Then, a combination of 2D (FCN-8s) and 3D (V-Net) architectures was used to segment the normal structures from the CT slices that had been classified to contain these structures. We evaluated the accuracy of our tool using 12 institutional and 24 publically available CT scans. (2) We identified 19 patients where our clinical multi-atlas-based autocontouring tool had failed to produce acceptable contours. We compared the unacceptable contours to those generated by our CNN-based tool using Dice and Hausdorff distance. These metrics were investigated to automatically detect contouring errors.

Results: (1) The Dice/Hausdorff distance between manual contours and those from our tool was 0.98/0.99mm for brain, 0.89/0.80mm for mandible, 0.89/0.32mm for eyes, 0.87/0.55mm for brainstem, 0.84/1.09mm for parotids, 0.84/0.46mm for spinal cord, 0.77/0.96mm for esophagus, 0.73/0.47mm for optic nerves, 0.71/0.27mm for lenses, 0.70/0.33mm for cochleas, and 0.38/0.69mm for optic chiasm. (2) The average Dice between the multi-atlas-based tool and the CNN-based tool was 0.97 for acceptable and 0.46 for unacceptable contours for brain, and 0.85 and 0.35 for brainstem. We established criteria for 10/11 structures that detect all contouring errors with an average false positive rate of 7.8%.

Conclusion: We developed a reliable CNN-based autocontouring tool and showed that this tool can effectively detect contouring errors.

Keywords

Quality Assurance, Segmentation, CT

Taxonomy

IM/TH- image segmentation: CT

Contact Email

Additional Links