Autonomous Treatment Plan Optimization Strategy Augmented By Using a Knowledge-Guided and Irregularly Downsampled Voxelization Scheme

Abstract

Current IMRT/VMAT planning is tedious and labor intensive due to its involvement of manual trial-and-error adjustment of model parameters. Despite of intensive research devoted to autonomous treatment planning, a truly efficient and reliable approach in a clinical environment has yet to be developed. A major shortcoming of the existing approaches is their reliance of DVH domain for characterization of a treatment, which is degenerate and does not always provide sufficient guidance to the solution search process. the use of voxel domain, on the other hand, is hampered by the huge dimensionality and thus intractable computational efficiency. The purpose of the work is to mitigate the problem by implementing an autonomous planning strategy in a novel dose-anatomy guided downsampling voxel scheme and showcase the novel approach by planning several clinical head and neck and pancreatic VMAT treatments in a commercial planning system. A sparse voxelization scheme, referred to as the isodose-feature preserving voxelization (IFPV), introduced recently by our group for concise representation of a treatment plan, is used to represent prior knowledge and to facilitate the optimization. Briefly, anatomical and dosimetric information of a treatment plan is characterized by IFPV, which partitions the voxels into subgroups according to their geometric, anatomical and dosimetric features. For a given patient, the IFPV clusters are generated based on prior treatment reference plans with similar anatomy. A plan is then generated using a commercial treatment planning system (TPS) with its planning parameters initialized with the features of the clusters. API script programming is employed to interact with the TPS to implement autonomous plan evaluation and automated optimization parameter update. The procedure proceeds in an iterative fashion until reaching an acceptable tradeoff among target dose coverage/homogeneity and sparing of critical structures. In each iteration, the IFPV clusters are updated based on current dose distributions. Clinical cases with different sites are tested to demonstrate the proposed approach. Compared with the organ DVH-based approach, it is found that the optimization with IFPV voxelization is much easier to be guided toward the desired solution. This especially favors our autonomous planning approach and makes autonomous optimization more efficient. In addition, because of more effective use of both geometric and dose information during plan optimization, the IFPV voxelization-based inverse planning typically yields better treatment plans, as indicated by the testing clinical cases. This work presents the first implementation of IFPV voxelization into a commercial TPS platform for autonomous treatment planning. The algorithm developed here is directly translatable to widespread clinical used, thus the approach has the potential to significantly improve the radiation therapy workflow and patient care.