Unencrypted login | home

Program Information

Selecting Reference Patients for Automatic Treatment Planning Using Multiple Geometrical Features

no image available
M Karimi

M Karimi*, N Li, M Zarepisheh, L Cervino, X Jia, K Moore, S Jiang, Center for Advanced Radiotherapy Technologies (CART), University of California, San Diego, La Jolla, CA

SU-E-CAMPUS-T-2 Sunday 3:00PM - 6:00PM Room: Exhibit Hall

Purpose: To discover the geometric features that best identify an anatomically-similar reference patient from a library of previous treatments to predict achievable dose distributions for automatic IMRT treatment planning of a new patient.

Methods: Using a library of previously-treated patients, two similarity functions are defined in the domain of the dose distributions and geometrical features that match a patient with their most similar member from the cohort. The goal is to attain the same patient matches from these similarity functions, so that when a new patient without dosimetric information is input into the geometric similarity function a similar patient from the library can be identified with good dosimetric predictive power for the new patient. The dosimetric similarity function is defined by a weighted sum of the difference area between two DVHs for all organs and PTV, with the weights tuned to reflect clinical priorities. The generalized geometric similarity function was defined as a nonlinear combination of organ/PTV volume, organ/PTV overlapping volume, and Dice index between a pair of two organs/PTVs/intersections. A genetic-algorithm-based learning technique was then employed to find the optimal combination of geometrical features to predict patient similarity.

Results: For a database of 10 prostate cancer treated patients, a combination of bladder volume and (PTV, rectum) overlapping Dice index achieved 80% accuracy compared to the DVH similarity function. Validation on a library of 15 patients reduced the accuracy to 60%, implying that other geometric features could be needed.

Conclusion: We have developed a method using machine learning to identify the best match of a new patient to the prior patients to guide automatic IMRT planning. Initial results from this feasibility study have modest predictive power but future work will focus in the incorporation of other geometric features, e.g. mutual information and overlap-volume-histograms, to increase selection accuracy.

Contact Email: