Representing Knowledge for Radiation Therapy Planning with Markov Logic Networks

Abstract

Radiation oncologists rely on clinical guidelines and results of clinical research studies to design an effective Radiation Treatment (RT) plan with the goal of maximum damage to cancer cells and minimum effects on normal tissue. As a step toward computerizing the clinical guidelines and clinical trials results in RT planning, we propose an approach and investigated its feasibility of representing the complex and uncertain RT knowledge using Markov Logic Networks (MLNs). MLNs combines both probability and first-order logic in a single representation to encode uncertain knowledge and perform reasoning on incomplete data or evidence. Within this approach, different types of RT knowledge with associated uncertainty can be extracted from published clinical guidelines and research studies, then be represented into a computerized formal model, and reasoned with evidence for intelligent RT planning. As an example for demonstration we focus on the RT planning scenario for limiting the risk of radiation-induced effects and suggesting dosimetric criteria and prescription dosage. We tested the constructed MLNs by making inferences to predicate the risk of radiation-induced effects given RT dose-volume plan. The initial results show the MLNs prediction of risk is in the range of risk suggested in guidelines. The proposed approach can be generalized to represent and reason with uncertain knowledge in radiation oncology for decision support.