Room: AAPM ePoster Library
Purpose: To report on the dose-prediction approach that uses a search engine to find cases of similar anatomy in a database of peer-reviewed cases. We show that contrary to the traditional knowledge-based algorithms, the presented approach retains all information within the original plans and can be used in clinical practice to automate the planning process.
Methods: This work uses a recent mathematical formalism developed for shape analysis to quantify segmentations. To construct the search data, all cases were aligned to a reference dataset. Variance for each PTV point was quantified to obtain a shape’s eigenvectors describing its discrepancies from the reference dataset. When planning a new case, the procedure is repeated to align the new case to the reference dataset. Corresponding eigenvalues are compared to each database case using a Mahalanobis distance quantifying their similarities. The search procedure outputs the list of cases in the database ordered by anatomical similarity to a new patient’s volume. The previously attained dose is used as a prediction. Additionally, the strategy to create the dose fall-off with ring structures is replicated from the database match to the new patient CT dataset, with an ESAPI script that automatically creates a treatment plan based on settings reproduced from the previous treatment.
Results: The approach was tested on lung cases treated with SBRT, where a script using common arcs and optimization options did produce sub-optimal results when compared to the clinical plans. However, when using settings learned from the database, plans of comparable quality were obtained across a wide variety of tumor locations and sizes.
Conclusion: From a practical point of view, the system delivered feedback on beam/arc directions that worked in the past for tumors of similar shapes, guiding plan parameters selection that improved the accuracy of automated planning methods.