Texture Analysis Feature Optimization for Intensity-Modulated Radiation Therapy

Automated Intensity-Modulated Radiation Therapy (IMRT) planning is based on a set of constraints and objectives, called a wish list. This thesis concerns automated IMRT planning with a non-smooth, non-convex, discontinuous, multivariable objective function. Specifically, constrained optimization using the Long Run High Gray-level Emphasis (LRHGE) texture analysis feature as objective was tested for IMRT planning of prostate cancer treatment. Incorporation of the LRHGE in the in-house developed, non-linear, gradient based solver Erasmus-iCycle by using finite difference derivatives showed that gradient based optimization does not lead to convergence, most likely due to the non-smooth, non-convex, discontinuous behavior of the LRHGE. Two global optimization solvers that do not require gradients were tested as well and both did converge. However, the resulting plans were not clinically optimal. An attempt was made to speed-up optimization using reduced order modeling. The relative inaccuracy of multivariate regression approximations and radial basis function approximations of the LRHGE was used as a benchmark for performance. Using multivariate regression approximations of the LRHGE led to 2.5 orders of magnitude speed-up, while maintaining median relative error below 2% for two out of four data sets. Radial basis function approximations performed better than multivariate regression in terms ofmedian relative error for three out of four data sets.However, in terms of speed multivariate regression approximations were superior to radial basis function approximations, since multivariate regression approximations were calculated around 15 times faster than radial basis function approximations. Despite the improvement in computation time, it is still impractically long for clinical practice, especially for realistic problem sizes. Also all attained solutions from optimization did not show improvement compared to current practice. All in all, using the LRHGE in treatment planning did not show any benefits