Author: Gourav Jhanwar, Mojtaba Tefagh, Masoud Zarepisheh 👨🔬
Affiliation: The University of Edinburgh, Memorial Sloan Kettering Cancer Center 🌍
Purpose:
Radiotherapy treatment planning involves solving computationally-intensive and time-consuming optimization problems. Central to these problems is a large matrix known as the dose-influence matrix (AKA dose deposition matrix or dij matrix) which quantifies the radiation dose delivered from each beamlet to each voxel. In practice, this matrix is sparsified by omitting small elements, often representing scattering components. The matrix sparsification accelerates the optimization process but compromises accuracy and plan quality. Our findings demonstrate that the scattering components form a highly compressible matrix, enabling compressed radiotherapy treatment planning (CompressRTP). This allows the optimization problems to be solved rapidly yet accurately.
Methods:
We separately precompute primary and scattering dose contributions within the dose-influence matrix for IMRT planning. Our analysis reveals that the scattering matrix’s singular values decay exponentially, demonstrating its low-rank and compressible nature. Since the primary dose matrix is sparse, this supports the use of the well-established "sparse-plus-low-rank" matrix compression. We developed a novel compression algorithm that operates efficiently without direct access to the scattering matrix. This algorithm was used to optimize treatment plans for 20 patients (10 lung, 10 prostate) using both compressed and sparsified matrices. The technique was integrated into our automated planning system, integrated with Eclipse, and evaluated based on dosimetric quality and computational efficiency.
Results:
Plans generated with CompressRTP achieved comparable PTV coverage and improved OAR sparing. For prostate patients, mean bladder and rectum doses decreased by 8.8% and 12.5%, respectively. For lung patients, mean lung and heart doses dropped by 10.8% and 11.2%, respectively, compared to sparsified matrix plans. Planning optimization time was reduced by over 40%.
Conclusion:
The large dosimetric datasets in treatment planning optimization are highly compressible, enabling rapid treatment planning without compromising data integrity and plan quality. This redundancy in data primarily stems from correlations between machine parameters such as adjacent beams/beamlets/voxels.