Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Three-Dimensional Dose Super-Resolution Using Implicit Neural Representation

V Vasudevan1*, O Pastor-Serrano2,4, C Huang3, C Chuang4, Z Perko2, P Dong4, L Xing4, (1) Stanford University, Stanford, CA, (2) Delft University of Technology, Delft, NL, (3) Stanford University, Stanford, CA, (4) Stanford University School of Medicine, Stanford, CA,

Presentations

PO-GePV-M-112 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Spatial dose distribution data plays a key role in radiation treatment planning. The data is typically stored in large 3D arrays representing the voxelated patient volume, with typically millions of voxels. A compact and continuous representation of dose data is of significance in treatment planning and downstream applications such as dose super-resolution, given the significant speed-up that working with low-resolution data offers. This work aims to reduce the computational time in calculating high-resolution doses by developing a technique for 3D dose super-resolution using an implicit representation of the low-resolution dose.

Methods: We train a multi-layer perceptron (MLP) with sinusoidal activations to learn an implicit representation of the low-resolution dose. The learned representation is a continuous function that maps voxel spatial coordinates to their dose values. We then use the learned representation to sample the dose distribution at a higher resolution. We evaluate the quality of the proposed super-resolution technique using a spine tumor dose distribution.

Results: In our experiments, we compare the dose predicted by the MLP implicit representation to reference 1% noise Monte Carlo doses via the gamma evaluation Γ(1%, 1 mm), using trilinear upsampling as a baseline. We additionally compute the average relative error among voxels receiving any dose. With a very high gamma pass rate of 98.77% and low mean error of 1.39±0.48% (vs. 89.09% and 1.56±0.70% for trilinear upsampling, respectively), the proposed model performs high-accuracy super-resolution.

Conclusion: The presented approach is the first step in introducing implicit neural representation for 3D dose super-resolution, which has the potential to speed up radiotherapy workflows by allowing fast dose calculation in low-resolution grids. Future research includes extending the approach to dose distributions from other treatment sites.

Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by the National Institutes of Health (NIH) under grants R01CA256890 and R01CA227713.

Keywords

Treatment Planning, Dose

Taxonomy

Not Applicable / None Entered.

Contact Email

Share: