Gastrointestinal Cancer

PV QA 1 - Poster Viewing Q&A 1

SU_13_2134 - Predicting survival after liver SBRT by deep learning-based analysis of treatment dose plans

Sunday, October 21
1:15 PM - 2:45 PM
Location: Innovation Hub, Exhibit Hall 3

Predicting survival after liver SBRT by deep learning-based analysis of treatment dose plans
B. Ibragimov1, D. A. S. Toesca2, D. T. Chang2, A. C. Koong3, and L. Xing4; 1Stanford University, Stanford, CA, 2Stanford University, Palo Alto, CA, 3Department of Radiation Oncology, Division of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 4Stanford University School of Medicine, Palo Alto, CA

Purpose/Objective(s): Optimizing dose delivery during liver radiation therapy (RT) is critical for the success of the treatment. Modern high-dose stereotactic body RT (SBRT) has shown an increase of 3-year survival to 73% from 38% for standard RT. In practice, finding the optimal SBRT plan with all patient specific factors considered is challenging, as optimization is based on population-averaged dose-volume histograms (DVHs). In this work, we, for the first time, harness the power of deep learning to learn the relationships between 3D dosimetric treatment plans of liver patients and their survival data and then use the deep neural networks-based model to predict post-SBRT survival for new patients.

Materials/Methods: A database of 125 liver SBRTs with complete follow-ups was collected at our institution. Among 125 patients, 58 were treated for liver metastases, 36 for hepatocellular carcinoma, 27 for cholangiocarcinoma and 4 for other primary cancers. The aim was to study SBRT dose plans over liver volumes and predict which patients will be alive 2 years after the treatments. All dose plans were rescaled to 1.5 cm3 resolution and superimposed with liver mask so that only doses delivered to the livers remain. To analyze the resulting 3D dose images, we utilized the concept from deep learning – convolutional neural networks – specially developed to discover the consistent patterns in digital images. Individually for primary and metastatic cancers, the SBRTs were separated into 20 folders, and CNN performance was validated following 20-folder cross validation experiment.

Results: The deep learning model was able to discover consistent patterns in dose plans associated with negative survival prognosis. The model performance was of 0.738 and 0.660 measured in terms of area under the receiving operator characteristic curves (AUC of ROC) for primary and metastatic cases, respectively. When separated by deep learning, 58% and 27% of patients with positive and negative survival prognosis, respectively, were alive 2 years after SBRT for primary liver cancer. Similarly, 55% and 30% of patients with positive and negative survival prognosis, respectively, were alive 2 years after SBRT for metastatic liver cancer. The most negative survival prognosis for primary liver cancer is associated with irradiation of the liver segments II (risk score (RS): 0.57), V (RS: 0.68) and VIII (RS: 0.70). The most negative survival prognosis for metastatic liver cancer is associated with irradiation of the liver segments IV (RS: 0.73), V (RS: 0.64) and VI (RS: 0.61).

Conclusion: A novel framework pioneering the concept of deep analysis of dose plans has been developed for predicting liver post-SBRT survival. The obtained results are very promising and opens an new avenue towards highly-personalized RT planning.

Author Disclosure: B. Ibragimov: None. D.A. Toesca: None. D.T. Chang: None. L. Xing: Research Grant; Varian Medical Systems. Honoraria; Varian Medical Systems. Royalty; Varian Medical Systems, Standard Imaging Inc.

Bulat Ibragimov, PhD

Presentation(s):

Send Email for Bulat Ibragimov


Assets

SU_13_2134 - Predicting survival after liver SBRT by deep learning-based analysis of treatment dose plans



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 Predicting survival after liver SBRT by deep learning-based analysis of treatment dose plans