Encrypted login | home

Program Information

From Big Data to Knowledge: Predicting Treatment Re-Planning in Proton Radiotherapy for Lung Cancer

C Teng

C Teng1*, G Janssens2 , C Ainsley1 , B Teo1 , P Gabriel1 , G Valdes4 , A Berman1 , W Levin1 , C Simone3 , T Solberg4 , (1) University of Pennsylvania, Philadelphia, Pennsylvania, (2) IBA, Louvain-la-neuve, Brabant Wallon, (3) University of Maryland, Baltimore, MD, (4) UCSF Comprehensive Cancer Center, San Francisco, CA


TU-RPM-GePD-J(B)-4 (Tuesday, August 1, 2017) 3:45 PM - 4:15 PM Room: Joint Imaging-Therapy ePoster Lounge - B

Purpose: In light of tumor regression and normal tissue changes, dose distributions can deviate undesirably from what was planned. Consequently, re-planning is sometimes necessary during the course of treatment to ensure continued tumor coverage or to avoid overdosing organs at risk (OARs). We identified most important features that are predictive of re-planning for lung cancer.

Methods: All consecutive lung cancer patients who received definitive proton radiotherapy and had more than two evaluation CT scans at our institution from 2011 to 2015 were included in this study (n=70, 333 CTs). The cohort included a variety of tumor sizes, locations, histologies, beam angles, as well as radiation-induced tumor and lung change. Dosimetric features included changes in the DVHs of the PTV, ITV, and OARs (heart, cord, esophagus, brachial plexus and lungs). Non-dosimetric features included tumor and lung change, characterized by changes in sizes, and in the Hounsfield numbers. Beam-specific features included beam angles and changes in water equivalent thickness (WET) along the beam path. Moreover, we engaged three lung physicians to independently evaluate the DVHs and made retrospective re-planning decisions. Random forests and decision tree methods were applied to discover the most important features that predicted physicians’ re-planning decisions.

Results: Table 1 summarized the re-planning decisions by lung physicians as well as by the criteria of initial planning. The much lower re-planning rates by physicians suggested a different decision process for re-planning. Figure 1 listed the most informative dosimetric predictors for re-planning and their action thresholds in a decision tree. Tumor volume and density change without beam-specific information were not informative predictors (Figure 2).

Conclusion: A Penn institute-wide retrospective proton re-planning study was conducted for lung cancer. Informative features were discovered by machine learning tools and may be used to guide automatic online decision-making.

Contact Email: