Generative model of phase space
Systems and methods are disclosed for simulating dose deposition. The systems and methods perform operations comprising: receiving a set of training data representing phase space of a radiotherapy treatment device comprising propagation and scattering of particles inside the radiotherapy treatment device; training a generative machine learning model based on the set of training data to generate one or more samples of the phase space of the radiotherapy treatment device; and simulating dose deposition at a particular region of interest based on the one or more samples of the phase space generated by the generative machine learning model.
1 . A computer-implemented method comprising:
receiving a set of training data representing a phase space of a radiotherapy treatment device comprising propagation and scattering of particles inside the radiotherapy treatment device;
training a generative machine learning model based on the set of training data to generate one or more samples of the phase space of the radiotherapy treatment device; and
simulating a dose deposition at a particular region of interest based on the one or more samples of the phase space of the radiotherapy treatment device generated by the generative machine learning model, wherein a radiotherapy treatment plan for a patient is generated based on the simulated dose deposition.
2 . The computer-implemented method of claim 1 , wherein the set of training data describes position and momentum vectors of relevant particles within a set of regions of the radiotherapy treatment device.
3 . The computer-implemented method of claim 1 , wherein the generative machine learning model comprises a neural network comprising at least one of a generative adversarial network (GAN), a variational autoencoder (VAE), a normalizing flow network, or a diffusion model.
4 . The computer-implemented method of claim 1 , wherein the set of training data represent the phase space of the radiotherapy treatment device as a collection of samples from a probability distribution.
5 . The computer-implemented method of claim 1 , further comprising:
generating the set of training data by performing a Monte Carlo simulation, wherein the set of training data is stored and retrieved from a non-volatile storage device or is computed in real time.
6 . The computer-implemented method of claim 1 , wherein the set of training data comprises a multi-dimensional surface representing physical information about the particles inside the radiotherapy treatment device.
7 . The computer-implemented method of claim 1 , wherein the trained generative machine learning model is stored on a same processing device that simulates the dose deposition when simulating the dose deposition to avoid accessing the one or more samples of the phase space of the radiotherapy treatment device from a remote storage location.
8 . The computer-implemented method of claim 1 , further comprising:
generating a three-dimensional (3D) volume of the dose deposition based on an interaction between the one or more samples of the phase space of the radiotherapy treatment device and an image detector, and wherein the generative machine learning model is part of a radiotherapy system that includes the image detector.
9 . The computer-implemented method of claim 1 , further comprising:
generating a three-dimensional (3D) volume of the dose deposition based on an interaction between the one or more samples of the phase space of the radiotherapy treatment device and an image detector, and wherein the generative machine learning model is separate from a radiotherapy system that includes the image detector.
10 . The computer-implemented method of claim 1 , wherein:
the generative machine learning model comprises a generative adversarial network (GAN) configured to train a generative model using a discriminative model; and
values applied by the generative model and the discriminative model are established using adversarial training between the discriminative model and the generative model.
11 . The computer-implemented method of claim 10 , wherein:
the adversarial training between the discriminative model and the generative model comprises:
training the generative model to generate a first synthetic sample of the phase space of the radiotherapy treatment device from a distribution of the particles inside the radiotherapy treatment device; and
training the discriminative model to classify the first synthetic sample as a synthetic or a real training particle inside the radiotherapy treatment device; and
an output of the generative model is used for training the discriminative model and an output of the discriminative model is used for training the generative model.
12 . The computer-implemented method of claim 11 , further comprising:
obtaining a particle from the set of training data; and
computing a training loss for the discriminative model based on a result of comparing an output of the discriminative model with the particle obtained from the set of training data.
13 . The computer-implemented method of claim 11 , wherein the first synthetic sample is generated based on a random point within the phase space or a specified region within the phase space.
14 . The computer-implemented method of claim 1 , wherein the generative machine learning model comprises a normalizing flow network, further comprising:
obtaining a particle from the set of training data;
computing a training loss based on a likelihood of the particle obtained from the set of training data based on a distribution provided by the normalizing flow network; and
updating the distribution provided by the normalizing flow network based on the training loss.
15 . The computer-implemented method of claim 14 , further comprising:
computing a new particle based on a distribution provided by the normalizing flow network, wherein the new particle is computed based on a random point within the phase space or a specified region within the phase space of the radiotherapy treatment device.
16 . A system for generating one or more radiotherapy treatment plans, the system comprising:
one or more processors configured to perform operations comprising:
receiving a set of training data representing a phase space of a radiotherapy treatment device comprising propagation and scattering of particles inside the radiotherapy treatment device;
training a generative machine learning model based on the set of training data to generate one or more samples of the phase space of the radiotherapy treatment device; and
simulating a dose deposition at a particular region of interest based on the one or more samples of the phase space of the radiotherapy treatment device generated by the generative machine learning model, wherein a radiotherapy treatment plan for a patient is generated based on the simulated dose deposition.
17 . The system of claim 16 , wherein the set of training data describes position and momentum vectors of relevant particles within a set of regions of the radiotherapy treatment device.
18 . The system of claim 16 , wherein the generative machine learning model comprises a neural network comprising at least one of a generative adversarial network (GAN), a variational autoencoder (VAE), a normalizing flow network, or a diffusion model.
19 . A non-transitory computer-readable medium comprising non-transitory computer-readable instructions for performing operations of claim 1 .
20 . A computer-implemented method comprising:
accessing a generative machine learning model that has been trained based on a set of training data to generate one or more samples of a phase space of a radiotherapy treatment device, the set of training data comprising propagation and scattering of particles inside the radiotherapy treatment device; and
simulating a dose deposition at a particular region of interest based on the one or more samples of the phase space of the radiotherapy treatment device generated by the generative machine learning model, wherein a radiotherapy treatment plan for a patient is generated based on the simulated dose deposition.
21 . The computer-implemented method of claim 20 , wherein the set of training data describes position and momentum vectors of relevant particles within a set of regions of the radiotherapy treatment device.
22 . The computer-implemented method of claim 20 , wherein:
the generative machine learning model comprises a generative adversarial network (GAN) configured to train a generative model using a discriminative model; and
values applied by the generative model and the discriminative model are established using adversarial training between the discriminative model and the generative model.
23 . The computer-implemented method of claim 22 , wherein:
the adversarial training between the discriminative model and the generative model comprises:
training the generative model to generate a first synthetic sample of the phase space of the radiotherapy treatment device from a distribution of the particles inside the radiotherapy treatment device; and
training the discriminative model to classify the first synthetic sample as a synthetic or a real training particle inside the radiotherapy treatment device; and
an output of the generative model is used for training the discriminative model and an output of the discriminative model is used for training the generative model.