MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep Learning-Driven Target Volume Delineation for Prostate Cancer Radiation Therapy

Y Wu1*, N Kovalchuk1 , A Hsu1 , B Han1 , W Zhao1 , L Wang1 , Y Rong2 , L Xing1 , (1) Stanford Univ, School of Medicine, Stanford, CA (2) University of California-Davis, Sacramento, CA

Presentations

(Sunday, 7/29/2018) 3:00 PM - 3:30 PM

Room: Exhibit Hall | Forum 6

Purpose: Target volume delineation is a time consuming task. Automatic delineation of target volume increases efficiency and mitigates the inter-observer variation. In this work, we aimed to develop a deep learning based technique to automatically delineate target volume for prostate cancer.

Methods: The retrospective study comprised of 25 patients. The target volumes, prostate and lymph nodes, were manually delineated by radiation oncologists. A volumetric deep convolutional neural network was trained end-to-end to provide automatic segmentation of the whole 3D image set. Unlike conventional methods that work on 2D images, our neural network does all image processing volumetrically. The network architecture consisted of two paths, contracting and expanding, each having five hierarchical levels that were connected via global shortcuts. Within each level on a single path, there were several convolutional blocks, where local shortcuts were employed to facilitate residual learning. The foreground/background prediction of every voxel concluded the task of delineation at the last level. The model was trained with the Adam optimization method, using the dice similarity coefficient as the loss function. Data from 20 patients were selected for training, and the remaining five were used for validation. Data sets were substantially enlarged using image augmentation methods, such as rotation, flipping and scaling.

Results: The prostates and lymph nodes were automatically segmented with the output comparable to the manual segmented results. The dice similarity coefficient for prostate delineation was 87%. Fast convergence was achieved in training, taking only 1000 iterations. The delineation of volumes was instantaneous in validation, requiring one second per patient.

Conclusion: The volumetric deep convolutional neural networks with local and global shortcut connections is a promising method for accurate and efficient automatic delineation of target volumes. The end goal of this project will be delineation of all targets and organs at risk simultaneously.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email