Purpose: Both human and computer optimization of treatment plans have advantages; humans are much better at global pattern recognition, and computers are much better at detailed calculations. A major impediment to human optimization of treatment plans by manipulation of beam parameters is the long time required for feedback to the operator on the effectiveness of a change in beam parameters. Our goal was to create a real-time dose calculation and display system that provides the planner with immediate (fraction of a second) feedback with displays of three-dimensional (3D) isodose surfaces, digitally reconstructed radiographs (DRRs), dose-volume histograms, and/or a figure of merit (FOM) (i.e., a single value plan score function). This will allow the experienced treatment planner to optimize a plan by adjusting beam parameters based on a direct indication of plan effectiveness, the FOM value, and to use 3D display of target, critical organs, DRRs, and isodose contours to guide changes aimed at improving the FOM value.
Methods and materials: We use computer platforms that contain easily utilized parallel processors and very tight coupling between calculation and display. We ported code running on a network of two workstations and an array of transputers to a single multiprocessor workstation. Our current high-performance graphics workstation contains four 150-MHz processors that can be readily used in a shared-memory multithreaded calculation.
Results: When a 10 x 10-cm beam is moved, using an 8-mm dose grid, the full 3D dose matrix is recalculated using a Bentley-Milan-type dose calculation algorithm, and the 3D dose surface display is then updated, all in < 0.1s. A 64 x 64-pixel DRR calculation can be performed in < 0.1 s. Other features, such as automated aperture calculation, are still required to make real-time feedback practical for clinical use.
Conclusion: We demonstrate that real-time plan optimization using general purpose multiprocessor workstations is a practical goal. Parallel processing technology provides this capability for 3D planning systems, and when combined with objective plan ranking algorithms should prove effective for optimizing 3D conformal radiation therapy. Compared to our earlier transputer work, multiprocessor workstations are more easily programmed, making software development costs more reasonable compared with uniprocessor development costs. How the dose calculation is partitioned into parallel tasks on a multiprocessor work station can make a significant difference in performance. Shared-memory multiprocessor workstations are our first choice for future work, because they require minimum programming effort and continue to be driven to higher performance by competition in the workstation arena.