A Cloud Computing Based Multi-Objective Optimization Method for Inverse Treatment Planning
Y Na1*, T Suh2, L Xing1, (1) Stanford University School of Medicine, Stanford, CA, (2) Catholic Univ Medical College, SeoulSU-E-T-628 Sunday 3:00:00 PM - 6:00:00 PM Room: Exhibit Hall
Purpose: Multi-objective (MO) plan optimization entails generation of an enormous number of IMRT or VMAT plans constituting the Pareto surface, which presents a computationally challenging task. The purpose of this work is to overcome the hurdle by developing an efficient MO method using emerging cloud computing platform.
Methods: As a backbone of cloud computing for optimizing inverse treatment planning, Amazon Elastic Compute Cloud with a master node (17.1 GB memory, 2 virtual cores, 420 GB instance storage, 64-bit platform) is used. The master node is able to scale seamlessly a number of working group instances, called workers, based on the user-defined setting account for MO functions in clinical setting. Each worker solved the objective function with an efficient sparse decomposition method. The workers are automatically terminated if there are finished tasks. The optimized plans are archived to the master node to generate the Pareto solution set. Three clinical cases have been planned using the developed MO IMRT and VMAT planning tools to demonstrate the advantages of the proposed method.
Results: The target dose coverage and critical structure sparing of plans are comparable obtained using the cloud computing platform are identical to that obtained using desktop PC (Intel Xeon® CPU 2.33GHz, 8GB memory). It is found that the MO planning speeds up the processing of obtaining the Pareto set substantially for both types of plans. The speedup scales approximately linearly with the number of nodes used for computing. With the use of N nodes, the computational time is reduced by the fitting model, 0.2+2.3/N, with r^2>0.99, on average of the cases making real-time MO planning possible.
Conclusions: A cloud computing infrastructure is developed for MO optimization. The algorithm substantially improves the speed of inverse plan optimization. The platform is valuable for both MO planning and future off- or on-line adaptive re-planning.