MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

DoseNet: A Volumetric Dose Prediction Algorithm Using 3D Fully-Convolutional Neural Networks

V Kearney*, J Chan , S Haaf , M Descovich , T Solberg , University of California San Francisco, San Francisco, CA

Presentations

(Wednesday, 8/1/2018) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 4

Purpose: Convolutional neural network (CNN) based dose prediction has been successful in predicting radiotherapy doses on 2D planes based on simple IMRT co-planar plans. However, the perceptive fields of such approaches are limited to 2 dimensions, which render them disadvantaged when handling non-coplanar dose predictions. To mitigate this, the present study suggests a novel deep learning volumetric non-coplanar dose prediction model that utilizes a 3D fully-convolutional neural network architecture.

Methods: This study evaluates the root mean squared error (RMSE), and common clinical dosimetric criteria differences between clinically approved dose (CAD) and a stacked fully-connected network (FC), a 2D CNN (2D-CNN), and a 3D CNN (DoseNet). This study considers 146 SBRT prostate patients treated using non-coplanar plans on a CyberKnife machine. Volumetric dose is predicted based on the planning CT, prostate, urethra, bladder, rectum, and penile bulb. The network accepts each input independently, creating interdependent hierarchical relationships between the structures and the planning CT, which allows the network to consider tissue inhomogeneity and structural geometries simultaneously.

Results: DoseNet achieved superior dosimetric congruence with the clinically approved dose than FC and 2D CNN for the Dmax, Dmean, D95, and D99 of the prescribed tumor volume. DoseNet also predicted more dosimetrically similar organ at risk maximum dose values, for the rectum, penile bulb, urethra, and bladder. DoseNet required 600 epochs to train, which took ~14 hours on two Nvidia 1080 Ti graphics processing units (GPU)s. DoseNet makes a full volumetric prediction in ~1 second.

Conclusion: This study has demonstrated that 3D fully-convolutional neural networks are capable of achieving superior dosimetric prediction accuracy than fully-connected networks and 2D based CNNs. The highest dosimetric improvement was seen in high dose regions.

Funding Support, Disclosures, and Conflict of Interest: Vasant Kearney and Samuel Haaf work with Nimble Therapy LLC.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email