Accuracy-first optimization for dose calculation & planning workflows

Introduction

When optimizing medical software, speed gains mean nothing if they compromise clinical accuracy. In radiotherapy dose calculation and treatment planning workflows, a 1% dose difference can have clinical consequences. Here’s how we approach optimization with a validation-first methodology—ensuring every performance improvement is validated against strict accuracy requirements before it reaches production.

The Challenge of Data Transfer

Dose calculation pipelines move large volumes of data: CT images, structure sets, beam configurations, and dose grids. When moving these workloads to the GPU or optimizing existing implementations, the PCIe bus often becomes the bottleneck. Time spent transferring data between host and device memory can negate the compute benefits of parallelization.

We recommend a streaming architecture where data is transferred in chunks asynchronously while the GPU processes the previous chunk. This hides transfer latency and keeps the GPU fed. For dose calculation specifically, consider pre-loading beam geometries and structure masks while the previous beam’s dose is being accumulated.

Precision Matters

While single precision (FP32) is often sufficient for visualization, dose calculation and planning workflows frequently require double precision (FP64) or mixed precision to maintain numerical stability. Compiler optimizations that reorder floating-point operations can change results; reduced precision can cause clinically significant dose differences in regions with steep gradients.

Key Takeaways

Before optimizing for speed, establish a reference implementation and automated regression suite. Every optimization must pass the same tolerance-based tests the reference passes. Document any intentional numerical differences and ensure they remain within acceptable clinical bounds.

Conclusion

Accuracy-first optimization is non-negotiable in medical imaging. By profiling to find real bottlenecks, creating bit-exact or tolerance-based validation, and optimizing incrementally while validating at each step, teams can deliver both performance gains and confidence in clinical correctness.