Radiation Physics

PV QA 3 - Poster Viewing Q&A 3

TU_16_3271 - Fast and Automatic Segmentation of Multiple Organs from ViewRay MR Images Using Deep Densely Connected CNN for Adaptive Radiation Therapy Treatment Planning

Tuesday, October 23
1:00 PM - 2:30 PM
Location: Innovation Hub, Exhibit Hall 3

Fast and Automatic Segmentation of Multiple Organs from ViewRay MR Images Using Deep Densely Connected CNN for Adaptive Radiation Therapy Treatment Planning
Y. FU1, T. R. Mazur1, S. Liu2, X. Chang1, Y. Lu3, H. Li1, P. J. Parikh1, and D. Yang1; 1Washington University School of Medicine, St. Louis, MO, 2Stanford University, Stanford, CA, 3Washington University School of Medicine, Department of Radiation Oncology, St. Louis, MO

Purpose/Objective(s): To expedite the contouring process for fast adaptive radiotherapy treatment planning, a deep densely connected convolutional neural network (DeepDenseNet) is proposed to automatically segment multiple organs, including liver, kidneys, stomach and large bowel from ViewRay 3D MR images.

Materials/Methods: A densely connected convolutional neural network is proposed to segment liver, kidneys, stomach and large bowel from ViewRay 3D MR images voxel by voxel. A total of 120 patients’ datasets were collected from our department for training and testing of the DeepDenseNet. Treatment sites of those 120 patients include pancreas, liver, stomach, adrenal gland and prostate. The DeepDenseNet was trained using 100 datasets and tested on the remaining 20 datasets. DICE coefficient, median Hausdorff distance (HD) and max HD were calculated for the 20 testing datasets to evaluate the segmentation accuracy of the DeepDenseNet. To demonstrate the clinical relevance and usefulness of the proposed method, the trained DeepDenseNet model was used to expedite the contouring process of adaptive treatment planning by segmenting these organs prior to manual contouring. For comparison, the time cost to contour these organs with and without the help of the DeepDenseNet was reported.

Results: The DeepDenseNet was able to segment the liver, kidney, large bowel and stomach with good accuracy and speed. For the 20 testing patients, the average DICE coefficients were 0.87, 0.88, 0.95 and 0.92 for large bowel, stomach, liver and kidneys respectively. The average median HD were 2.59 mm, 2.72 mm, 2.34 mm and 2.31 mm for large bowel, stomach, liver and kidneys respectively. Contouring experiments were performed by four medical physics residents. Three randomly selected patients’ datasets were contoured by each resident. The experiments showed that the contouring process with the help of the DeepDenseNet was on average ×3.3 times faster than before.

Conclusion:

A new convolutional neural network called DeepDenseNet was proposed to segment multiple organs automatically from abdominal ViewRay 3D MR images with good segmentation accuracy and speed. The proposed method has the potential to greatly expedite the contouring process for fast adaptive radiotherapy treatment planning.

Author Disclosure: Y. FU: None. T.R. Mazur: None. S. Liu: None. X. Chang: None. P.J. Parikh: Research Grant; Varian Medical, Viewray. Honoraria; Viewray. Speaker's Bureau; Sirtex. Advisory Board; Sirtex. Stock; Holaira. D. Yang: Employee; Mercy Health. Research Grant; Agency for Healthcare Research and Quality, Viewray Inc.

Send Email for Yabo FU


Assets

TU_16_3271 - Fast and Automatic Segmentation of Multiple Organs from ViewRay MR Images Using Deep Densely Connected CNN for Adaptive Radiation Therapy Treatment Planning



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 Fast and Automatic Segmentation of Multiple Organs from ViewRay MR Images Using Deep Densely Connected CNN for Adaptive Radiation Therapy Treatment Planning