Purpose/Objective(s): This work introduces a concise representation of treatment plan(s) with consideration of all major spatial and dosimetric features of the corresponding isodose distribution(s), and establish a novel plan optimization framework to effectively utilize the features derived from prior knowledge for fast and autonomous treatment planning.

Materials/Methods: A new voxelization scheme, referred to as the isodose-feature preserving voxelization (IFPV), is introduced for sparse representation of a treatment plan. In the IFPV domain, conventional voxels that are spatially and dosimetrically close are aggregated by using a K-means algorithm into physically meaningful clusters while respecting the delineated anatomical structures of the case. For an assemble of plans with similar anatomies and treatment prescriptions, the IFPV features, such as the maximum/minimum/mean doses of each IFP voxel, are extracted from each plan to characterize the plans. The collected data also define the variation range of the dose at the IFP and set realistic upper and lower bounds for the dose at the IFP voxel. To use the population-based IFPV domain data to facilitate the search for a clinically sensible solution, we extended a previously published inverse planning algorithm that parameterizes the dosimetric tradeoff among the involved structures with physically meaningful quantities such as the permissible dose variation ranges to the IFPV domain. The approach is applied to plan three head-and-neck (HN) cancer cases, and the results are compared quantitatively with the plans generated independently by a dosimetrist and by using a conventional DVH-based algorithm.

Results: A novel inverse planning framework in sparse IFPV domain has been established to summarize prior planning knowledge and to automate the treatment planning process. The new scheme reduced the number of voxels by 3 orders of magnitudes, from tens of millions of voxels to thousands of IFP voxels, thus substantially reduces the computational time needed for inverse planning. We found that the proposed inverse planning strategy is capable of provide plans comparable or even favorable than that obtained by a dosimetrist. Because IFPVs contain much richer information than that of DVHs, a significant improvement in planning quality is also found relative to the alternative DVH-based planning strategy.

Conclusion: The IFPV framework combines physically more meaningful modeling of the inverse planning problem and smart consolidation of voxels and substantially improves the planning process with greatly reduced the need for manual trial and error determination of the model parameters.