Dose-volume-based IMRT fluence optimization: A fast least-squares approach with differentiability

In intensity-modulated radiation therapy (IMRT) for cancer treatment, the most commonly used metric for treatment prescriptions and evaluations is the so-called dose volume constraint (DVC). These DVCs induce much needed flexibility but also non-convexity into the fluence optimization problem, which is an important step in the IMRT treatment planning. Currently, the models of choice for fluence optimization in clinical practice are weighted least-squares models. When DVCs are directly incorporated into the objective functions of least-squares models, these objective functions become not only non-convex but also non-differentiable. This non-differentiability makes it problematic that software packages designed for minimizing smooth functions are routinely applied to these non-smooth models in commercial IMRT planning systems. In this paper, we formulate and study a new least-squares model that allows a monotone and differentiable objective function. We devise a greedy approach for approximately solving the resulting optimization problem. We report numerical results on several clinical cases showing that, compared to a widely used existing model, the new approach is capable of generating clinically relevant plans at a much faster speed, with speedups above one-order of magnitude for some large-scale problems.