Encrypted login | home

Program Information

Direct Aperture Optimization Utilizing L0-Penalty Based Total Count Variation Regularization

D Nguyen

D Nguyen*, D O'Connor , Q Lyu , D Ruan , K Sheng , Department of Radiation Oncology, University of California Los Angeles, Los Angeles, California


SU-I-GPD-T-323 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall

Purpose: Converting optimized fluence to deliverable apertures has resulted in unpredictable plan quality degradation, even with TV regularization with the L1-norm to smooth the fluence prior to conversion. To address this limitation, we introduce a direct aperture optimization (DAO) using an L0-based total count variation (L0-TV) regularization to directly penalize the cardinality of fluence discontinuities.

Methods: The DAO problem was formulated with an L2-norm dose fidelity term and an anisotropic L0-penalty on fluence differentials. The L0-TV regularization enforces piecewise constant regions, resulting in fluence maps that are discretized into apertures whose number is regularized in the same optimization function. The non-convex optimization problem can be efficiently solved with a special variant of the Primal Dual Hybrid Gradient (PDHG) algorithm. For evaluation, the algorithm was tested on a head-and-neck (H&N) with 3 PTVs, a lung (LNG), and a glioblastoma multiforme (GBM) patient, utilizing 20 non-coplanar beams. The method was compared to the fluence map optimization problem with L1-TV regularization, which was MLC segmented to match the number of deliverable fluence segments of the L0-TV plans.

Results: The average number of segments per beam for each plan was 10.1 (H&N), 9.75 (LNG), and 2.6 (GBM). The L0-TV method reduced the average mean and max dose by 4.48% and 2.81% of the prescription dose, respectively. The van’t Riet conformation number was superior with the L0-TV method for all 3 patient cases, increasing from 0.891 to 0.878 (H&N), 0.737 to 0.770 (LNG), and 0.629 to 0.747 (GBM). On average, PTV homogeneity increased, changing from 0.906 to 0.919 (H&N), 0.935 to 0.933 (LNG), and 0.891 to 0.946 (GBM) with the L0-TV method. PTV coverage (D95,D98,D99) differed less than 1% of the prescription dose.

Conclusion: The novel DAO approach utilizing the non-convex L0-penalty produces plans superior to its L1-counterpart in PTV conformity, heterogeneity, and OAR sparing.

Funding Support, Disclosures, and Conflict of Interest: DOE DE-SC0017057 NIH R44CA183390 NIH R01CA188300 NIH R43CA183390 NIH U19AI067769

Contact Email: