(Tuesday, 7/16/2019) 11:00 AM - 12:15 PM
Room: Stars at Night Ballroom 2-3
Purpose: To investigate a fully automatic framework based on deep dilated U-Net (dU-Net) for multiple OARs segmentation on CT images for prostate cancer patients, and to compare its performance with two experienced planners.
Methods: To address the challenges in automatic OAR delineations in treatment planning, we proposed dU-Net, a learning-based framework that can simultaneously segment multiple organs in 3D space. dU-Net aggregates contextual information by convolution and max-pooling, and restores the image details by transpose convolution and skip connections. In order to further improve the information recovery in the fine scale, dilated convolution is employed to replace the bottom of convolutional-deconvolutional pathway of a traditional U-Net architecture. We also generalize the Jaccard distance loss function, which was originally designed for single object segmentation, to multiple objects to further improve the learning efficiency. The performance was evaluated by comparing the auto-segmentation results with ground-truth contours that were manually outlined by a physician. In order to better understand its performance, we had two experienced planners manually contour the same OARs and compared their results with ground truth.
Results: We trained dU-Net using 120 prostate cancer patients, and employed the trained model to segment prostate, seminal vesicles (SVs), bladder and rectum on 28 testing patients (3 seconds per patient), resulting in a mean Dice-Similarity-Coefficient (DSC) of 0.83, 0.72, 0.93 and 0.82, respectively. While planner 1 performed similarly as dU-Net with DSCs of 0.84, 0.74, 0.93 and 0.80, planner 2 (DSCs: 0.86, 0.78, 0.95 and 0.85) achieved statistically better performance in SVs (p=0.023) and prostate (p=0.030).
Conclusion: Our preliminary results demonstrate that dU-Net has potential to achieve similar performance as planners in contouring multiple-OARs for prostate cancer patients. The DSC varieties among different OARs also reflect the intrinsic challenges in contouring these organs, which need to be considered when further improving the segmentation performance.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
