Radiation Physics

PV QA 3 - Poster Viewing Q&A 3

TU_16_3275 - External Validation of a Deep Learning-Based Auto-Segmentation Method for Radiation Therapy

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

External Validation of a Deep Learning-Based Auto-Segmentation Method for Radiation Therapy
M. Kuo1, T. Zhang2, H. Zhong1, M. Huang1, H. Geng1, C. Cheng1, Y. X. Li2, J. Dai2, and Y. Xiao1; 1University of Pennsylvania, Philadelphia, PA, 2National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China

Purpose/Objective(s): Automated delineation of the organs-at-risk (OARs) has high application value for treatment planning in radiotherapy due to the efficiency and consistency. We previously reported a Deep Dilated Convolutional Neural Networks (DDCNN)-based auto-segmentation method for radiotherapy of rectal cancer. The DDCNN model was constructed and implemented with data from a single institution. In this study, we investigated the reproducibility of DDCNN using independent validation samples from a public dataset for external validation.

Materials/Methods: The DDCNN was trained using data of 218 patients from institution A. We performed model testing with 60 patients from institution A (internal validation dataset), and 18 patients with expert consensus delineations from a public dataset (external validation dataset). All patients from institution A were set up with the prone position. While the patients from the external validation dataset were set up in the supine position. The images were processed geometrically to be consistent with the positioning required by the model (prone vs. supine). A histogram normalization operation was applied to ensure the consistency of the imaging parameters. The performance of DDCNN was evaluated on segmentation of the bladder, left femoral head, and right femoral head.

Results: The segmentation accuracy was consistent for both internal and external cases. The Dice Similarity Coefficient (DSC) values were 0.93±0.03 vs. 0.89±0.07 for the bladder, 0.92±0.04 vs. 0.90±0.05 for the left femur, and 0.92±0.03 vs. 0.91±0.05 for the right femur, respectively. For the bladder, the accuracy of segmentation from external validation was significantly worse than that from internal validation (p<0.05). However, the lower dice value was due to the different filling levels of the bladder and the contrast agents remaining in the bladder in some cases of external validation set. We selected the cases that had high consistency with the training set through visual inspection and evaluated the segmentation accuracy. The DSC value was 0.92±0.03, which was comparable with that of the internal validation.

Conclusion: Despite different patient positioning, patient sizes, and imaging parameters, DDCNN showed overall comparable agreements on segmentation of the OARs with clear boundaries between the internal and external validation sets. The external validation demonstrates that the deep learning method has the potential for general application in radiotherapy, even though its performance in segmentation of organs with considerable inter- and intra-observer variability needs further investigation.

Author Disclosure: M. Kuo: None. T. Zhang: None. H. Geng: None. Y. Li: Chairman, Department of Nuclear Medicine; Cancer Hospital & Institute, CAMS & PUMC. Chairman, Department of Radiation Oncology; Cancer Hospital & Institute, CAMS & PUMC.

Men Kuo, PhD

Presentation(s):

Send Email for Men Kuo


Assets

TU_16_3275 - External Validation of a Deep Learning-Based Auto-Segmentation Method for Radiation Therapy



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 External Validation of a Deep Learning-Based Auto-Segmentation Method for Radiation Therapy