MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Using Very Small Contour Sets to Train High-Quality Deep-Learning Segmentation Models

Y Zhao*, D Rhee, C Cardenas, L Court, J Yang, The University of Texas MD Anderson Cancer Center, Houston, TX

Presentations

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

Room: AAPM ePoster Library

Purpose: To generate high-quality deep-learning segmentation models when the availability of contoured cases is severely limited (e.g.~10 patients).

Methods: Ten head-and-neck CT scans with well-defined contours were deformably registered to 200 CT scans of the same anatomical site without contours. The acquired deformation vector fields were used to train a principal component analysis (PCA) model for each of the 10 well-contoured CT scans by capturing the mean deformation and the most prominent variations. In total, we trained 10 PCA models, where each model can produce an infinite number of synthetic CT scans and corresponding contours by applying random deformations.
We used 300, 600, 1000, and 2000 synthetic CT scans and contours generated from one PCA model to train a V-net, a 3D CNN architecture, for left parotid autosegmentation. We repeated this by using the same numbers of training cases generated from 7 and 10 PCA models with the training data distributed evenly between each PCA model. We evaluated the performance of the trained segmentation models by calculating the Dice similarity coefficients between the auto-generated contours and the physician-drawn contours on 30 independent test CT scans.

Results: The average Dice for left parotid of the VNet model trained with 1 PCA model (i.e. generated from one well-contoured patient) were 45.4%, 48.2%, 53.4%, and 56.2% for the 300, 600, 1000, and 2000 synthetic training cases, respectively. As the number of PCA models was increased to 7 and 10, the average Dice increased to 61.7%, 66.7%, 73.2%, 74.7%; and 65.8%, 70.0%, 75.2%, 77.8%, respectively.

Conclusion: We demonstrated an effective data augmentation approach to train high-quality deep-learning segmentation models from a very limited number of well-contoured patients. This work could potentially greatly reduce the effort in data curation for deep-learning based autosegmentation.

Funding Support, Disclosures, and Conflict of Interest: Disclosures: Our research group receives funding from the NCI and Varian Medical Systems.

Keywords

Segmentation, Data Acquisition, CT

Taxonomy

IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email