A robust-CVaR optimization approach with application to breast cancer therapy

Timothy C. Y. Chan, Houra Mahmoudzadeh, Thomas G. PurdieEur. J. Oper. Res 2014 · DOI

Abstract

This paper introduces a novel robust optimization framework utilizing Conditional Value-at-Risk (CVaR) to manage uncertainty in systems with unknown state durations. Applied to breast cancer radiation therapy, the approach accounts for unpredictable breathing patterns, ensuring the heart is spared from excess radiation while the tumor receives the prescribed dose.

Introduction

In radiation therapy for left-sided breast cancer, clinicians face a critical challenge: delivering a lethal dose of radiation to the Clinical Target Volume (CTV) while minimizing exposure to the heart, which lies directly behind the breast. The primary source of uncertainty is patient respiration, which causes the chest to move and shifts the heart in and out of the radiation field.

Existing clinical standards, such as "average scans" or Active Breathing Control (ABC), often fall short. Average scans can lead to significant discrepancies between planned and realized doses, while ABC is resource-intensive and difficult for many patients to tolerate. This research addresses the gap by combining robust optimization—designing for worst-case scenarios—with CVaR to handle the uncertainty of time spent in various breathing phases.

Methodology

The authors propose two robust-CVaR formulations to manage systems fluctuating between finite breathing states:

  1. Weighted CVaR (w-CVaR): Calculates the CVaR for each individual state and takes a weighted average based on the time spent in each phase.
  2. Accumulated CVaR (CVaR-a): Generates a single, combined dose distribution across all states before calculating the CVaR of the total distribution.

To ensure computational tractability, the authors utilize duality-based reformulations to transform these non-convex risk problems into Linear Programs (LPs), solvable via standard commercial tools like CPLEX. Furthermore, the model employs "deformable registration" (voxel mapping) to track specific tissue points as they move across different breathing phases.

Key Insights and Ideas

A central thesis of the paper is that CVaR serves as a superior clinical proxy compared to traditional dose-volume criteria. While standard criteria are essentially Value-at-Risk (VaR) metrics, CVaR is mathematically convex and accounts for the severity of the dose in the distribution's "tails" (extreme values), rather than just the threshold.

Dose distribution for the Clinical Target Volume (CTV), highlighting the underdose and overdose tails targeted by CVaR optimization.

The study also explores the trade-off in conservatism. The CVaR-a model is mathematically proven to be less conservative than w-CVaR. By analyzing the accumulated distribution, CVaR-a allows for greater flexibility in heart sparing without over-protecting against physically impossible scenarios. This framework is a generalization of current methods; by shrinking the uncertainty set, one can recover standard "average" or "ABC" methods.

Clinical Application and Results

The study utilized a clinical 4D-CT dataset from the Princess Margaret Cancer Centre, encompassing five distinct breathing phases. To manage computational load, the authors sampled one out of every 32 voxels, maintaining accuracy in Dose-Volume Histograms (DVHs).

Key results include:

  • Stability: Robust models produced planned dose distributions that remained stable across 300 simulated breathing scenarios, whereas the average method often failed to meet clinical safety standards under variability.
  • Clinical Success: The CVaR-a model met all CTV coverage criteria while significantly reducing heart dose compared to w-CVaR.
  • Performance Frontier: The robust solution pushed treatment quality toward the ABC "gold standard" without requiring expensive breath-holding equipment.

Limitations and Future Work

While promising, the approach requires 4D-CT scans, which deliver higher imaging radiation than standard CTs. Additionally, the CVaR-a model is computationally intensive, requiring 164 minutes to solve compared to 68 minutes for w-CVaR. Future research will focus on validating these findings across a larger patient cohort and exploring ellipsoidal uncertainty sets to further refine treatment plans.

Summary updated

BibTeX
@article{chan2014robust,
  author = {Timothy C. Y. Chan and Houra Mahmoudzadeh and Thomas G. Purdie},
  title = {{A robust-CVaR optimization approach with application to breast cancer therapy}},
  journal = {Eur. J. Oper. Res.},
  year = {2014},
  doi = {10.1016/j.ejor.2014.04.038},
}
Figure