Graph-based risk assessment and error detection in radiation therapy

Abstract

Purpose : The objective of this study was to formalize and automate quality assurance (QA) in radiation oncology. QA in radiation oncology entails a multistep verification of complex, personalized radiation plans to treat cancer involving an interdisciplinary team and high technology, multivendor software and hardware. We addressed the pretreatment physics chart review (TPCR) using methods from graph theory and constraint programming to study the effect of dependencies between variables and automatically identify logical inconsistencies and how they propagate. Materials and Methods : We used a modular approach to decompose the TPCR process into tractable units comprising subprocesses, modules and variables. Modules represent the main software entities comprised in the radiation treatment planning workflow and subprocesses group the checks to be performed by functionality. Module‐associated variables serve as inputs to the subprocesses. Relationships between variables were modeled by means of a directed graph. The detection of errors, in the form of inconsistencies, was formalized as a constraint satisfaction problem whereby checks were encoded as logical formulae. The sequence in which subprocesses are visited was described in a activity diagram. Results : The comprehensive model for the TPCR process comprised 5 modules, 19 subprocesses and 346 variables, 225 of which were distinct. Modules included ”Treatment Planning System” and ”Record and Verify System”. Subprocesses included ”Dose Prescription”, ”Documents”, ”CT Integrity”, ”Anatomical Contours”, ”Beam Configuration”, ”Dose Calculation”, ”3D Dose Distribution Quality” and ”Treatment Approval”. Variable inconsistencies, their source and propagation are determined by checking for constraint violation and through graph traversal. Impact scores, obtained through graph traversal, combined with severity scores associated with an inconsistency, allow risk assessment. Conclusions : Directed graphs combined with constraint programming hold promise for formalizing complex QA processes in radiation oncology, performing risk assessment and automating the TPCR process. Though complex, the process is tractable.