Implicit neural representation for radiation therapy dose distribution.

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Imaging

Background:

• Radiotherapy treatment planning relies on 3D dose distribution data, often stored in large voxel grids.
• Efficient representation of this data is crucial for planning and downstream applications.

Purpose of the Study:

• To develop an implicit neural representation for 3D dose distribution data in radiotherapy.
• To create a concise and accurate method for characterizing dosimetric data.

Main Methods:

• Utilized a coordinate-based multilayer perceptron (MLP) with sinusoidal activations to map spatial coordinates to dose values.
• Trained MLPs to implicitly represent dose distributions, evaluating performance across various tumor sites and model architectures.
• Reconstructed dose distributions by evaluating the trained MLP at each voxel location.

Main Results:

• Achieved a mean-squared error of 10^-6 and peak signal-to-noise ratio >50 dB with a compression ratio of ~32.
• Optimal accuracy was observed with model sizes corresponding to bitrates of 1-2, with performance degradation at lower bitrates.
• Demonstrated effective compression and accurate reconstruction for prostate, spine, and head and neck tumor cases.

Conclusions:

• Implicit neural representations offer a low-dimensional, continuous alternative to voxel grids for 3D dose data.
• The proposed method provides an accurate and compact way to represent radiotherapy dose distributions.
• This work establishes a foundation for using neural representations in radiation oncology applications.