RADI-19. EVALUATION OF BRAIN METASTASIS LOCAL CONTROL POST RADIOSURGERY VIA MACHINE LEARNING AND RADIOMICS. (12th August 2019)

• Record Type: Journal Article
• Authors: Dise, Joseph; Abraham, Christopher; Caruthers, Douglas; Mutic, Sasa; Robinson, Clifford
• Is Part Of: Neuro-oncology advances. Volume 1(2019)Supplement 1
• Page Start: i25
• Page End: i25
• Publication Date: 2019-08-12
• Subjects: 616.99481
• DOI: 10.1093/noajnl/vdz014.111
• Languages: English
• ISSNs: 2632-2498

Abstract

Abstract: Stereotactic radiosurgery can be used to treat multiple, surgically inaccessible, metastatic brain lesions in a single, minimally invasive outpatient procedure. For brain metastasis, stereotactic radiosurgery can provide excellent local control depending on the robustness of the treatment plan. Previous studies have been performed correlating key radiation planning factors to higher local control probability such as tumor size and maximum dose. However, a separate non-inferiority study demonstrated that higher prescription isodose lines (in excess of 70% or higher) did not correlate to local failure. The previous works were limited to shallow feature levels regarding only the dicom plan information and lacked a predictive model. In order to address these conflicting conclusions and to support clinical decision making, we propose a radiosurgery informatics pipeline to support testing these hypotheses with observational data. First, a multidisciplinary team generated a mind-map of relevant information to inform database design. Portions of this mind-map were implemented in a relational database system (PorstgreSQL), and populated with information from 1024 patients treated for brain metastasis via stereotactic radiosurgery. Clinical information were derived from curated databases and the array of intervention variables were mined from the DICOM RT plans, structure sets, images and dose via MATLAB scripts. These factors include, but are not limited to, radiation dosimetry, prior whole brain radiation, radiomic imaging features, prior radiosurgery status, and physician determined local control status. From this pipeline, we plan to use a multi-level feature-based supervised machine learning approach that will be created via boosting to predict local control in patients using local failure timing, or lack thereof, provided by physician. To control for local failure observer bias, an unsupervised machine learning model via random trees will be created to predict clusters of patient parameters with similar local control rates.