Neural Representation for Three-Dimensional Dose Distribution and its Applications in Precision Radiation Therapy

Abstract

<h3>Purpose/Objective(s)</h3> Spatial dose distribution plays an important role in radiation treatment planning. Dose data is expressed in the voxelated grid representing the patient volume, and depending on the case, the total number of voxels can be of the order of 10^6–10^8. A compact representation of 3D dose data is of significance in facilitating treatment planning and downstream applications. This work aims to develop a technique to compress the 3D dose data using neural representation and demonstrate its potential in facilitating radiation therapy dose calculation and treatment planning. <h3>Materials/Methods</h3> In contrast to storing the dose values at each voxel, we propose using the weights of a multi-layer perceptron (MLP) to represent the dose data implicitly. We train a coordinate-based MLP with sinusoidal activations to map the voxel spatial coordinates to their dose values. We first identify the best architecture for a given parameter budget and use that to train a model for each patient in our dataset. The trained model is queried at each coordinate to reconstruct the 3D dose distribution at inference. We systematically evaluate the quality of the proposed representation by performing experiments on dose distributions of varying complexity from different disease sites. <h3>Results</h3> In our experiments, we generate implicit neural representations for 3D dose distributions of prostate, spine, and head and neck tumor cases. The learned representations achieve a peak signal-to-noise ratio greater than 50 dB and a compression ratio of ∼32, at a target bitrate of ∼1 for dose data from all three sites. The number of parameters in the trained network is less than 4% of the average number of voxels in dose data. Our results also show that model sizes with a bit rate of 1–2 are optimal for the task, and performance drops significantly for bitrates smaller than that. <h3>Conclusion</h3> We show how to learn a low-dimensional implicit neural representation of 3D dose data methodically and accurately. The learned representation is a continuous function and can accurately model the high-frequency information in the dose data. The continuous nature of the representation allows us to sample the dose distribution at arbitrary spatial resolutions. This study lays the groundwork for future applications of neural representations of dose data in radiation oncology.