SYSTEM AND METHOD FOR LEARNING MODELS OF RADIOTHERAPY TREATMENT PLANS TO PREDICT RADIOTHERAPY DOSE DISTRIBUTIONS
The present disclosure relates to systems and methods for developing radiotherapy treatment plans though the use of machine learning approaches and neural network components. A neural network is trained using one or more three-dimensional medical images, one or more three-dimensional anatomy maps, and one or more dose distributions to predict a fluence map or a dose map. During training the neural network receives a predicted dose distribution determined by the neural network that is compared to an expected dose distribution. Iteratively the comparison is performed until a predetermined threshold is achieved. The trained neural network is then utilized to provide a three-dimensional dose distribution.
1 . A method for training neural network model by performing a set of training operations comprising:
receiving a plurality of training data comprising one or more three-dimensional medical images and one or more three-dimensional anatomy maps and one or more three-dimensional dose distributions;
training the neural network model to predict both a fluence map and a dose map, the neural network model trained based on the one or more three-dimensional medical images, the one or more three-dimensional anatomy maps and the one or more three-dimensional dose distributions, the fluence map depicting a number of particles per second of applied radiation crossing a volume element in a patient, and the dose map depicting a radiation dose to be delivered to a patient from a radiotherapy device at a particular location; and
storing the trained neural network model.
2 . The method of claim 1 , further comprising:
generating a three-dimensional dose distribution based on both the fluence map and the dose map predicted by the neural network model.
3 . The method of claim 1 , wherein the one or more three-dimensional anatomy maps correspond to a subset of the one or more three-dimensional medical images and indicate locations of anatomical structures and treatment targets.
4 . The method of claim 3 , wherein the one or more three-dimensional anatomy maps comprise at least one of an image contour, a contoured surface in space, a map of functional anatomy, a binary mask corresponding to a structure, and a function of the structure comprising a signed distance map.
5 . The method of claim 1 , wherein the fluence map comprises a three-dimensional map of fluence.
6 . The method of claim 1 , wherein the set of training operations comprises:
initializing the neural network model with an initial layer configuration, an initial connection configuration, an initial set of weights, and an initial set of biases;
inputting a first set of the plurality of training data to the initialized neural network model, the first set of training data comprising patient records from a population of patients that include medical images, specific anatomical structures, and expected three-dimensional dose distributions;
receiving a predicted dose distribution from the neural network model;
comparing the predicted dose distribution from the neural network model with one of the expected dose distributions;
adjusting weights and biases of the neural network model to decrease differences between the predicted dose distribution and the one of the expected dose distributions;
repeating the set of training operations until the differences between the predicted dose distribution and the one of the expected dose distributions reach a predetermined threshold; and
storing the trained neural network model.
7 . The method of claim 6 , wherein the plurality of training data comprises a second set of training data, and the second set of training data comprises an updated set of new patient images corresponding to a particular patient.
8 . The method of claim 1 , wherein the plurality of training data comprises at least one of a dose distribution, a measure of quality based on a dose volume histogram, an image contour, a contoured surface in space, functional anatomy, and a signed distance map and combinations thereof to train the neural network model.
9 . The method of claim 1 , wherein the plurality of training data comprises a function of the predicted dose distribution, wherein the function is at least a square or an exponential.
10 . The method of claim 1 , further comprising:
receiving the trained neural network model;
inputting testing data into the trained neural network model, the testing data comprising new patient records from a new population of patients that include new medical images, new specific anatomical structures, and new expected dose distributions;
obtaining an additional new predicted dose distribution from the trained neural network model; and
determining an error factor by comparing the new expected dose distributions with the additional new predicted dose distribution.
11 . The method of claim 1 , wherein the neural network model comprises a deep convolutional neural network (DCNN).
12 . A non-transitory computer-readable medium comprising computer-readable instructions that, when executed by one or more processors, configure the one or more processors to perform a set of training operations comprising:
receiving a plurality of training data comprising one or more three-dimensional medical images and one or more three-dimensional anatomy maps and one or more three-dimensional dose distributions;
training a neural network model to predict both a fluence map and a dose map, the neural network model trained based on the one or more three-dimensional medical images, the one or more three-dimensional anatomy maps and the one or more three-dimensional dose distributions, the fluence map depicting a number of particles per second of applied radiation crossing a volume element in a patient, and the dose map depicting a radiation dose to be delivered to a patient from a radiotherapy device at a particular location; and
storing the trained neural network model.
13 . The non-transitory computer-readable medium of claim 12 , wherein the one or more three-dimensional anatomy maps correspond to a subset of the one or more three-dimensional medical images and indicate locations of anatomical structures and treatment targets.
14 . The non-transitory computer-readable medium of claim 13 , wherein the one or more three-dimensional anatomy maps comprise at least one of an image contour, a contoured surface in space, a map of functional anatomy, a binary mask corresponding to a structure, and a function of the structure comprising a signed distance map.
15 . The non-transitory computer-readable medium of claim 12 , wherein the fluence map comprises a three-dimensional map of fluence.
16 . The non-transitory computer-readable medium of claim 12 , wherein the set of training operations comprises:
initializing the neural network model with an initial layer configuration, an initial connection configuration, an initial set of weights, and an initial set of biases;
inputting a first set of the plurality of training data to the initialized neural network model, the first set of training data comprising patient records from a population of patients that include medical images, specific anatomical structures, and expected three-dimensional dose distributions;
receiving a predicted dose distribution from the neural network model;
comparing the predicted dose distribution from the neural network model with one of the expected dose distributions;
adjusting weights and biases of the neural network model to decrease differences between the predicted dose distribution and the one of the expected dose distributions;
repeating the set of training operations until the differences between the predicted dose distribution and the one of the expected dose distributions reach a predetermined threshold; and
storing the trained neural network model.
17 . A method for using a trained neural network model, the method comprising:
receiving a set of three-dimensional medical images from an image acquisition device;
applying the trained neural network model to the set of three-dimensional medical images received from the image acquisition device to predict both an individual fluence map and an individual dose map associated with the set of three-dimensional medical images;
generating a three-dimensional dose distribution based on both the individual fluence map and the individual dose map predicted by the trained neural network model; and
storing the three-dimensional dose distribution in association with the set of three-dimensional medical images.
18 . The method of claim 17 , wherein the set of three-dimensional medical images are associated with one or more three-dimensional anatomy maps that indicate locations of anatomical structures and treatment targets.
19 . The method of claim 18 , wherein the individual fluence map comprises a three-dimensional map of fluence.
20 . The method of claim 19 , wherein the one or more three-dimensional anatomy maps comprise at least one of an image contour, a contoured surface in space, a map of functional anatomy, a binary mask corresponding to a structure, and a function of the structure comprising a signed distance map.