Opinion - Grids for optimizing cancer radiotherapy treatment

(The author, Dimitri Dimitroyannis, is a board-certified, clinical medical physicist at Kansas City Cancer Centers, who trained in experimental high-energy physics. This summary is rewritten from his longer article in Medical Physics Web.)

Transverse slice from an actual computer-aided tomography image of a patient with cancer at the base of the tongue. Soft tissue appears as light grey, while bone is whitish. Red area is the cancerous target that needs to be irradiated. The three yellow areas correspond to uninvolved tissues that need to receive as little radiation as possible — the spinal cord (center) and the two parotid glands (left and right). Image courtesy of Dimitri Dimitroyannis

About 4 million European and North American citizens will be diagnosed with cancer this year. More than half of these new cancer patients will receive radiotherapy for the treatment of their disease, along with surgery and chemotherapy as appropriate.

However, treating cancer patients with radiation is inherently contradictory: The goal is to apply a large, uniform dose of ionizing radiation to a small target area inside the patient's body. At the same time, we want to minimize collateral damage to healthy tissue nearby. This can be quite complicated (see image at right). To hit the precise target without spilling over requires careful planning and the use of 3D imaging — a computationally intensive process.

What's more, advances in treatment planning and radiation delivery have exacerbated the problem, as we move from traditional radiation-treatment to more complex schemes, such as new variations upon intensity-modulated radiation therapy (IMRT) — an advanced mode of high-precision radiotherapy that uses computer-controlled x-ray accelerators to deliver precise radiation doses to a malignant tumor, or even specific areas within a tumor. IMRT allows for the radiation dose to conform more closely to the three-dimensional shape of the tumor by modulating — or controlling — the intensity of the radiation beam over time and space.

Practical IMRT treatment-planning is based upon delivering the maximum radiation dose to the target while minimizing dose delivery elsewhere, paying special attention to at-risk, non-involved areas. In practice, the planner must guess at the values of certain starting parameters and achieve an overall optimum plan by making repeated iterations. Unfortunately, every patient is different, so there are no universally applicable "golden" starting parameters.

Finding the optimum IMRT radiotherapy plan is a "Pareto problem," also known as "optimization under constraints." One way to solve such problems is by generating a large number of competing solutions, and then using a mathematical technique known as multi-objective optimization to sort out the really promising ones.

There's a downside, however: the severe computational cost of generating hundreds of individual plans in order to find the handful that best hit the target.

To generate the large number of radiotherapy plans needed in a reasonable time, we have introduced a technique initially developed for the needs of the high-energy physics community: the computational grid.

We used the equivalent uniform dose (EUD) to represent the dose delivered to organs-at-risk (OAR). For visualization simplicity, we plotted just two OARs,  the spinal cord and the right parotid. The objective here is to minimize the dose to both these organs.  Blue: plans designated as no good; Red plans designated as good. Image courtesy of Dimitri Dimitroyannis

A test case

Using the computational engine of Pinnacle — a commercial system by Philips Medical Systems — we built a computational grid designed to generate a large number of radiotherapy plans. We tested it upon a challenging case in which a tumor was growing at the base of the tongue; the nearby organs-at-risk were the spinal cord and the salivary glands known as the left and right parotids. Starting from a plausible set of initial conditions, our grid created 150 individual plans in about 3 hours. We estimate that executing this task on a typical workstation would have taken more than 60 hours.

Our computational grid comprises 55 commodity workstations (HP dc7600, Pentium4 2.8 GHz) running on ScientificLinux 4.4 (CERN/Fermilab) and gLite. For convenience, data security, and to protect the patient's confidentiality, our grid was physically under our control. The grid was activated after-hours or during weekends.

At the end of the computation, all of the completed plans — technically known as integral dose-volume histograms (DVH) — were saved for analysis. They  were then examined by at least three experienced radiotherapy planners, who graded them as "good" or "no good" and awarded them 1 or 0 points respectively (see graph above). The result was a set of treatment plans that optimized the dosage for the cancer growing at the base of the tongue, while minimizing the dosage to nearby organs that were healthy.

Multi-objective optimized radiotherapy planning attempts to offer an intriguing solution to our current inability to produce certified optimal IMRT plans. We have demonstrated that computational grids are both feasible and — as the only existing realistic generators of large number of quasi-optimal plans — essential to success. We envision the deployment of geographically dispersed, load-balanced, time-shifted computational grids for our clinical needs, following the infrastructure lessons gained from observing high-energy physics labs.

Unfortunately, the utility of such computational grids is not yet widely known to IMRT practitioners. However, the radiotherapy community commands ample resources that, if harnessed as a focused computational grid virtual organization, could produce true personalized medicine for the immediate benefit of our patients.

—Dimitri Dimitroyannis