TH-AB-BRB-02: Enabling Web-Based Treatment Planning Using a State-Of-The-Art Convex Optimization Solver

Authors

  • Ungun B,

    1. Stanford University, Stanford, CA
    2. UT Southwestern Medical Center, Dallas, TX
    3. The University of California San Diego, La Jolla, CA
    4. Stanford University School of Medicine, Stanford, CA
    Search for more papers by this author
  • Folkerts M,

    1. Stanford University, Stanford, CA
    2. UT Southwestern Medical Center, Dallas, TX
    3. The University of California San Diego, La Jolla, CA
    4. Stanford University School of Medicine, Stanford, CA
    Search for more papers by this author
  • Bush K,

    1. Stanford University, Stanford, CA
    2. UT Southwestern Medical Center, Dallas, TX
    3. The University of California San Diego, La Jolla, CA
    4. Stanford University School of Medicine, Stanford, CA
    Search for more papers by this author
  • Boyd S,

    1. Stanford University, Stanford, CA
    2. UT Southwestern Medical Center, Dallas, TX
    3. The University of California San Diego, La Jolla, CA
    4. Stanford University School of Medicine, Stanford, CA
    Search for more papers by this author
  • Xing L

    1. Stanford University, Stanford, CA
    2. UT Southwestern Medical Center, Dallas, TX
    3. The University of California San Diego, La Jolla, CA
    4. Stanford University School of Medicine, Stanford, CA
    Search for more papers by this author

Abstract

Purpose:

To develop an ultra-fast web-based inverse planning framework for VMAT/IMRT. To achieve high speed, we investigate the use of a simple convex formulation of the inverse treatment planning problem that takes advantage of recent developments in the field of distributed optimization.

Methods:

A Monte Carlo (MC) dose calculation algorithm was used to calculate the dose matrix (268228 voxels × 360 beams, 96M non-zeros) for a 360-aperture, 4-arc VMAT plan taken from the clinic. We wrote the objective for the inverse treatment planning problem as a sum of convex (piecewise-linear) penalties on the dose at each voxel in the planning volume. This convex voxel-separable formulation allowed us to apply a new, open-source, CPU- and GPU-capable optimization solver (http://foges.github.io/pogs/) to calculate our solutions of optimal beam intensities. In each planning session, after performing one full optimization we accelerated subsequent runs by “warm-starting”: for run k, the optimal solution from run k-1 was used as an initial guess. We implemented the treatment planning application as a Python web server running on a standard g2–2xlarge GPU node on Amazon EC2.

Results:

Our method formed optimal treatment plans in 5–15 seconds. Warm-start times ranged from 100ms–8s (mean 3s) while sweeping out a 5-log range of tradeoffs between target coverage and OAR sparing in 1000 total optimizations. Satisfactory plans were reached in 1–10 iterations of the optimization, with total planning time <10 minutes. Dosimetric characteristics such as the DVH curves showed that the resultant plans were comparable or superior to the clinically delivered plan.

Conclusion:

This work demonstrates the feasibility of high-quality, low-latency treatment planning using a convex problem formulation and GPU- based convex solver, making it practical to manipulate treatment objectives and view DVH curves and dose-wash views in nearly real-time in a web application.

Funding support for this work is provided by the Stanford Bio-X Bowes Graduate Fellowship and NIH Grant 5R01CA176553

Ancillary