UCLA

Adaptive therapies (ART) have potential for improving treatment efficacy, reducing unnecessary exposure of normal tissues, and improving patient quality of life. Ideally, every patient could receive on-line ART, fully optimizing the treatment to their daily anatomy as they lie on the treatment table. Additionally, daily on-line ART would allow reductions in the planned error margins by more certainly locating the tumor targets, providing another avenue for reducing exposure to normal tissues. To date, the computational complexity, labor, and time required to perform the additional tasks necessary for on-line ART has made it an infeasible option for clinical implementation. Accelerating and automating these processes as much as possible will be imperative for clinical integration.

Towards this goal, software was developed for performing fast dose calculations, dose accumulation, contour propagation and analysis, deformable image registration (DIR) validation and error quantification, and biomechanical modeling. Each of these processes were accelerated for near real-time performance by parallelization and optimization for the architecture of graphics processing units (GPUs). Brief descriptions of the major contributions are given below.

A non-voxel-based dose convolution optimized for GPU architecture achieved over 4000x acceleration compared to a single-threaded implementation. Expanding this algorithm to a multi-GPU cloud-based implementation further increased the acceleration by a factor of two, despite the additional overhead associated with a distributed, cloud-based solution.

A DIR and dose accumulation framework was developed to track anatomical changes over the treatment course and estimate the actual delivered dose distribution. This framework was employed in retrospective studies to analyze the dose to the parotid glands for head-and-neck patients, and determine the feasibility of reducing error margins during planning.

A biomechanical modelling framework was developed to create patient-specific models from diagnostic imaging. Through GPU implementation, the high-resolution model maintained interactive framerates, for both linear elasticity and the subsequent evolution to hyper-elasticity. To validate the DIR algorithm employed in the dose accumulation framework, clinically realistic deformations were induced in patient-specific biomechanical models, which output simulated imaging volumes with known, ground-truth deformation vector fields.

Similar model-generated deformations supplied annotated training data for the development of a neural network able to infer a quantified error estimates for clinical DIR, requiring only image similarity information as input. This methodology delivers a fully automated, fast technique to replace a process that was historically time-consuming, user-biased, and subject to small sample sizes.

The works presented focused on head-and-neck patients, but were developed with a general approach and the intent to expand to other sites. With future integration, these tools provide a foundation for building an automated, accelerated pipeline for clinical implementation of on-line ART.