A method of generating a radiotherapy treatment plan for a patient, a computer program product and a computer system

Abstract

A machine learning-based method of generating a radiotherapy treatment plan for a patient, comprises dose prediction and dose mimicking, wherein the dose prediction step involves using a machine learning system that has been trained to consider at least one optimality criterion related to physical or technical restrictions that will affect the delivery of the treatment plan. Thus, at least one of the factors that are normally taken into account in the dose mimicking step is introduced in the dose prediction step. The invention also relates to a method of training such a machine learning system for use in radiotherapy treatment planning, a computer program product and a computer system.

Claims

1. A computer-based method for generating a radiotherapy treatment plan for a patient, the treatment plan comprising providing radiation to a patient from a radiation source, the method comprising: generating, by using a machine learning system, a predicted dose distribution or a set of irradiation parameters that can be used to generate a dose prediction for the patient dependent on patient geometry, based on a medical image and structure data comprising at least one delineated structure of the patient, wherein the machine learning system has been trained to consider at least one optimality criterion related to physical or technical restrictions that will affect the delivery of the treatment plan; and generating, by using the predicted dose distribution and/or set of irradiation parameters, a deliverable treatment plan.

2. The method of claim 1, wherein the machine learning system has been trained to consider the physicality of the delivery process including the interaction of the patient's geometry and the properties of the radiation source as the at least one optimality criterion, and the output from the dose prediction step is expressed as a dose distribution or a set of irradiation parameters, or both.

3. The method of claim 1, wherein the machine learning system has been trained to consider the constraints of the system used for delivering the radiotherapy, such as the machine constraints of a delivery system.

4. The method of claim 1, wherein the machine learning system comprises a non-trainable layer based on a dose function for the patient and is arranged to output a fluence distribution for each beam, spot positions and weights for each beam and energy or a set of brachy seed positions and dwell times, depending on the type of radiotherapy, and/or a dose distribution.

5. The method of claim 1, wherein the optimization problem comprises at least one constraint related to the at least one optimality criterion.

6. The method of claim 1, wherein the machine learning system is a neural network that has been designed to consider at least one technical constraint of a delivery system to be used when delivering the treatment as the optimality criterion.

7. A method for training a machine learning system for use in radiotherapy treatment planning, comprising: providing a set of input data related to a medical image of a patient, structure date comprising at least one delineated structure of the patient, and a desired dose distribution dependent on patient geometry, determined on the basis of the medical image to a machine learning system, comparing the output from the machine learning system to the desired dose distribution and feeding the result of the comparison to the machine learning system, wherein the set of input data further comprises at least one optimality criterion related to physical or technical restrictions that will affect the delivery of the treatment plan, and the training comprises training the machine learning system to consider the at least one optimality criterion related to physical or technical restrictions that will affect the delivery of the treatment plan.

8. The method of claim 7, wherein the at last one optimality criterion is related to the combined effect of the properties of the radiation source and the geometry of the patient.

9. The method of claim 7, wherein the machine learning system includes a non-trainable layer based on a dose function for the patient and the machine learning system is arranged to output a predicted dose distribution or a set of irradiation parameters that can be used to generate a dose prediction for the patient to be used as input data to the dose mimicking stage.

10. The method of claim 8, wherein the training comprises refining an optimization process, wherein the optimization procedure is based on an optimization problem comprising an optimization function related to the at least one optimality criterion.

11. The method of claim 7, wherein the at least one optimality criterion is related to the technical constraints of a delivery system to be used when delivering the treatment.

12. A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith which, when run in a computer cause the computer to perform the method according to claim 1.

13. A computer system comprising a processor, and a program memory, wherein the program memory comprises computer program code configured to, when executed by the processor, cause the computer system to perform the method of claim 1.

14. A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith which, when run in a computer cause the computer to perform the method according to claim 7.

15. A computer system comprising a processor and a program memory, wherein the program memory comprises computer program code configured to, when executed by the processor, cause the computer system to perform the method of claim 7.

Description

BRIEF DESCRIPTION OF DRAWINGS

[0027] The invention will be described in more detail in the following, by way of examples and with reference to the appended drawings.

[0028] FIG. 1 is an overview of a procedure of dose prediction and dose mimicking.

[0029] FIG. 2 illustrates the structure of a U-Net neural network adapted for use in an embodiment of the invention.

[0030] FIG. 3 is a flow chart of a method according to an embodiment of the invention.

[0031] FIG. 4 is a schematic drawing of a computer system which may be used according to the invention.

DETAILED DESCRIPTION OF EMBODIMENTS

[0032] FIG. 1 is an overview of a prior art method of dose prediction and dose mimicking. Input data to the procedure include at least one of a medical image S11 and structure data S12. The medical image S11 is a medical image of a portion of a patient, such as a CT image or an MR image. Structure data include delineated structures of the patients, of one or more regions of interest (ROI). These input data are used in a dose prediction step S13 that is performed in a machine learning system to generate a predicted dose which is a spatial dose adapted to the patient's geometry. The output from the dose prediction step S13 is a predicted dose S14, or a set of irradiation parameters that may be used in a subsequent step (not shown) to generate a predicted dose. The predicted dose S14 is used as input data to a dose mimicking step S15, which generates a deliverable treatment plan S16 for the patient. As is common in the art, the dose mimicking step is designed to ensure that the treatment plan comes as close as possible to producing the desired dose distribution, taking into account the actual physiological and technical conditions. This means that the dose mimicking step is typically designed to enforce the following criteria: [0033] the dose distribution resulting from the interaction of the irradiation with the patient tissue, taking the geometry and anatomy of the patient into account, [0034] the physicality of the radiation source, [0035] the deliverability of the plan with respect to the machine constraints of the delivery system that is to be used, [0036] the optimality criteria with respect to the predicted dose, [0037] other optimality criteria such as radiobiological effect (RBE)

[0038] The radiation source may be any known radiation source used for radiotherapy, including a device for providing an external beam to the patient or a radiation source introduced into the patient's body for brachytherapy. The radiation provided may be photon or ion based.

[0039] According to the invention at least one of the factors that are normally taken into account in the dose mimicking step S15 is introduced in the dose prediction step S13 to increase that chance that the predicted dose is close to an actually deliverable dose.

[0040] In a first preferred embodiment, the invention involves considering physical descriptors in the dose prediction stage, together with the physicality of the source of radiation. That is, the combined properties of the source of radiation and how it interacts with the patient tissue. The physical descriptors are given by the dose function

d=f(x)

[0041] where x is the set of treatment parameters, d is the resulting dose distribution in the patient, and f is the mapping function from treatment parameters to dose. The physical descriptors of the patient may come from any suitable source, typically from a medical image of the patient taken at an earlier point in time or in connection with the treatment planning. The medical image may be, for example, a planning image acquired for the initial treatment planning, or a fraction image taken before the delivery of a fraction, or a combination of two or more such images.

[0042] As a first possible implementation, an approximation of the dose mapping function f(x) could be introduced as a non-trainable layer in the machine learning system. For this embodiment, the machine learning system is preferably a neural network, which in the case of photon radiotherapy planning is arranged to output a desired fluence instead of the desired dose, as is done conventionally. In this case, accordingly, the output S14 from the dose prediction step, would be a desired fluence map instead of a predicted dose distribution. The dose mimicking step could in this case be adapted to work either on the dose distribution or on the fluence map, or on both.

[0043] A second possible implementation of the first embodiment would be to introduce a metric related to the difference between the output of the dose function and the dose, that is, |f(x)−d| as a penalty term part of the loss function for the machine learning system. The metric may, for example, be the absolute value of the difference, but any suitable metric may be used. For this second option, any optimization-based machine learning system could be used. Alternatively, a method based on a combination of machine learning and constraint satisfaction programming could be used. In this case d=f(x) would be imposed as a hard constraint in the machine learning system during training, so that the constraint would be an intrinsic part of the trained model.

[0044] In a second preferred embodiment, the invention involves considering machine constraints of the delivery system in the dose prediction stage. The machine constraints involve such factors as how much and how fast components of the delivery system, such as gantry and collimator leaves, can move. The machine constraints could be added to the optimization problem of any optimization-based machine learning system as an objective function or, preferably, as a constraint, to be treated analogously to the physics constraints discussed above.

[0045] The first and second preferred embodiments can be implemented independently of each other but can also be used together, to ensure that the predicted dose takes into account both the physicality of the patient and the technical constraints of the delivery system.

[0046] FIG. 2 illustrates a simplified architecture of a neural network known as a U-Net, adapted for use according to embodiments of the invention. The basic function of a U-Net is known to the skilled person and will only be discussed in general terms. As is common in the art, the U-Net comprises a number of layers; FIG. 2 shows a model with four layers. As will be known to the skilled person, the U-Net has a contracting path 21, the leftmost leg of the U and an expanding path 23, the rightmost leg of the U shape shown in FIG. 2. The contracting path consists of repeated application of convolutions 211, 213, 215, each followed by a rectified linear unit (ReLU) and a max pooling operation, as will be known to the skilled person. During the contraction, the spatial information is reduced while feature information is increased. The fourth layer is the fully contracted layer 220. The expansive pathway 23 combines the feature and spatial information through a sequence of up-convolutions 231, 233, 235 and concatenations with high-resolution features from the contracting path. The uppermost layer 235 of the expanding path includes a concatenation function and a convolution function as is common in the art, and the output of this according to embodiments of the invention will be a predicted dose 236. The predicted dose is input to a non-trainable layer 237 arranged to calculate irradiation parameters 239 based on the predicted dose. The irradiation parameters may be used to control dose delivery.

[0047] As explained above, input data to the machine learning model will include a medical image and structure data related to the patient.

[0048] FIG. 3 is a flow chart of a method of training the machine learning system according to embodiments of the invention. The training preferably comprises refining an optimization process, wherein the optimization procedure is based on an optimization problem comprising an optimization function related to at least one optimality criterion. In a first step S31, a set of input data is provided to a machine-learning system that is to be trained. The set of input data is related to a medical image of a patient, a structure data set, and a desired dose distribution dependent on patient geometry, determined on the basis of the medical image. According to embodiments of the invention, the set of input data also comprises at least one optimality criterion related to physical or technical restrictions that will affect the delivery of the treatment plan.

[0049] In step S32, the machine learning system produces an output based on the input data. In step S33, the output from step S32 is compared to a desired dose distribution already available and in step S34 the result of the comparison is fed to the machine-learning system as feedback.

[0050] The at least one optimality criterion may be related to the combined effect of the properties of the radiation source and the geometry of the patient. Alternatively, or in addition, the at least one optimality criterion may be related to the technical constraints of a delivery system to be used when delivering the treatment. Preferably, as shown in FIG. 2, the machine learning system includes a non-trainable layer based on a dose function for the patient and the machine learning system is arranged to output a predicted dose distribution or a set of irradiation parameters that can be used to generate a dose prediction for the patient to be used as input data to the dose mimicking stage.

[0051] FIG. 4 is a schematic representation of a computer system in which the inventive method may be performed. A computer 31 comprises a processor 33, a data memory 34 and a program memory 35. Preferably, a user input means 37, 38 is also present, in the form of a keyboard, a mouse, a joystick, voice recognition means or any other available user input means. The user input means may also be arranged to receive data from an external memory unit. It should be understood that FIG. 4 is only provided to illustrate different components that may be used to implement the method according to the invention and that the actual computer architecture can vary. One or more of the components, or the whole system, may be implemented in a cloud environment.

[0052] The data memory 34 typically comprises the necessary input data and the output data from the dose prediction step, as well as the treatment plan. The treatment plan may be generated in the computer 31, or received from another storage means in any way known in the art. As will be understood, the data memory 34 is only shown schematically. There may be several data memory units, each holding one or more different types of data, for example, one data memory for the input data, one for the dose distribution or irradiation parameters, etc.

[0053] The program memory 35 holds a machine learning system as discussed above and a dose mimicking program. As will be understood, the data memory 34 and/or program memory 35 are not necessarily part of the same computer as the processor 33, and may be located in any computer that is reachable from the processor, such as in a cloud environment.