Digital Health Innovation and Informatics

SS 16 - Digital Health Innovation & Informatics 1

120 - Automatic Segmentation Using Convolutional Neural Networks In Prostate Cancer

Monday, October 22
4:55 PM - 5:05 PM
Location: Room 007 A/B

Automatic Segmentation Using Convolutional Neural Networks In Prostate Cancer
C. C. Vu1, L. Zamdborg1, Z. A. Siddiqui1, G. S. Gustafson2, D. J. Krauss1, and T. M. Guerrero3; 1Dept. of Radiation Oncology, Beaumont Health, Royal Oak, MI, 2Dept. of Radiation Oncology, Beaumont Health, Troy, MI, 3Beaumont Health (Department of Radiation Oncology), Royal Oak, MI

Outline/Objective(s): Accurate delineation of tumor volumes and organs-at-risk is critical in radiation therapy treatment planning. In prostate cancer, commonly segmented organs include the prostate, seminal vesicles, bladder, and rectum. The purpose of this study was to evaluate a deep convolutional neural network algorithm for automatic segmentation in prostate cancer.

Materials/Methods: Treatment plans generated during normal clinical operations were restored. Patients were divided into training, validation, and test datasets according to a 70/20/10 split, and four additional patients served as a second test dataset for comparison against manual segmentation by resident physicians. MRI data was not used for training. We employed a modified 2D U-Net, applying transfer learning from the VGG16 image classification model. The model was then trained end-to-end online on four NVIDIA K80 GPUs using the ADAM optimizer to minimize the categorical cross-entropy loss. Real-time data augmentation was used during training. We employed dropout for regularization and batch normalization to reduce internal covariate shift. The model was implemented using Keras with a TensorFlow back-end.

Results: 18,624 CT slices from 114 patients were included in this study. Model training lasted approximately 2.7 days. Evaluation took 0.11 seconds per CT slice. On the resident test dataset, the mean Dice similarity coefficient was 82% for the prostate, 63% for the seminal vesicles, 85% for the rectum, and 86% for the bladder. These results were comparable to the contours generated by four radiation oncology residents of all training levels (PGY 2-5), who had a mean Dice similarity coefficient of 78% for the prostate, 67% for the seminal vesicles, 87% for the rectum, and 91% for the bladder (Table 1).

Conclusion: Convolutional neural networks can achieve auto-segmentation results in prostate cancer that are comparable to human manual segmentation by radiation oncology residents.

Abstract 120 - Table 1:

Machine

PGY-2

PGY-3

PGY-4

PGY-5

Prostate

82%

78%

80%

81%

75%

Seminal Vesicles

63%

69%

65%

68%

66%

Rectum

85%

86%

87%

88%

86%

Bladder

86%

92%

91%

92%

92%

Mean Dice similarity coefficient by organ for deep learning model and four radiation oncology residents. PGY: post-graduate year

Author Disclosure: C.C. Vu: None. L. Zamdborg: None. Z.A. Siddiqui: None. G.S. Gustafson: Partnership; GREATER MICHIGAN GAMMA KNIFE. MEMBER; ACR. D.J. Krauss: Partnership; Greater Michigan Gamma Knife. T.M. Guerrero: None.

Charles Vu, MD

William Beaumont Hospital

Disclosure:
Employment
Beaumont Health: Resident Physician: Employee

Presentation(s):

Send Email for Charles Vu


Assets

120 - Automatic Segmentation Using Convolutional Neural Networks In Prostate Cancer



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Automatic Segmentation Using Convolutional Neural Networks In Prostate Cancer