Click here to


Are you sure ?

Yes, do it No, cancel

Learning for Daily Pancreas SBRT Plan Adaptation Decision

Y Min1*, M Tyran2 , M Cao3 , P Lee4 , M Steinberg5 , D Ruan6 , (1) UCLA School of Medicine, Los Angeles, CA, (2) Institut Paoli Calmettes, Marseille, France, (3) UCLA School of Medicine, Los Angeles, CA, (4) UCLA, Los Angeles, CA, (5) University of California, Los Angeles, Los Angeles, CA, (6) UCLA School of Medicine, Los Angeles, CA


(Thursday, 7/18/2019) 11:00 AM - 12:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: Advance in on-board imaging provides valuable information to make daily adaptation decisions. However, manually assessing the similarity level of the per-fraction image to the plan reference is difficult, and the challenge for adaptation decision is further composited with dose impact. We hypothesize that a deep learning approach may circumvent the contouring process and derive adaptation decisions from fraction-wise imaging data and priors.

Methods: The overall decision system consists of two sequential modules: a 2D convolutional network that obtains intermediate likelihood from slice triplets and an aggregator. Specifically, the deep network performs volume-to-slice conversion from reference dose map, reference and per-fraction image volumes, and generates a likelihood estimator for adapt of that specific triplet. The network consists of 9 residual-keeping convolutional blocks, a feature vector concatenation step, and a fully connected network. A classic logistic regressor is used to fuse the 2D-triplet based likelihood to reach a volume-wise decision. Dose and image records for 35 clinical pancreas SBRT were collected and the optimal adapt decisions were made retrospectively by meticulously recontouring and assessment without effort constraint. To augment the training data, clinical plans and reference images were used in combination with simulated diffeomorphic B-spline based deformation. The adaptation endpoint was generated by calculating the dose variation and the pass/fail pattern according to clinical criteria for PTV/GTV coverage and duodenum/small-bowel/stomach/spinal-cord sparing.

Results: The final plan adapt decision estimator obtained accuracy of 87.56% over 900 test cases, with 80.6% sensitivity and 99.4% specificity. Decision on a volume can be generated in 700 milliseconds with NVIDIA GTX 1080 TI GPU.

Conclusion: The propose approach is able to generate SBRT plan adaptation decision without performing segmentation, circumventing its associated limitations in both quality and speed. The current study is being extended to address the plan adaptation strategy once an online adapt decision is made.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email