Radiotherapy planning based on neural network

Abstract

The embodiment of the invention relates to radiotherapy planning based on a neural network. The neural network is trained using a training corpus having a plurality of information features, each of the information features including both a reference radiotherapy dose and at least one corresponding post-treatment patient data. By one method, at least one corresponding post-treatment patient data includes a patient image, such as, but not limited to, one or more computed tomography images. The trained neural network facilitates radiotherapy planning by generating resultant therapy efficacy probability information and resultant therapy complication probability information.

Description

Technical Field

[0001] These teachings generally relate to treating a planned target volume of a patient with energy according to an energy-based treatment plan, and more particularly, to optimizing an energy-based treatment plan. Background Art

[0002] The use of energy to treat diseases includes known areas of the prior art. For example, radiation therapy includes an important part of many treatment plans for reducing or eliminating unwanted tumors. Unfortunately, the applied energy cannot inherently distinguish between unwanted materials and adjacent tissues, organs, etc., which are desirable or even crucial for the continued survival of the patient. Therefore, energy such as radiation is usually applied in a carefully administered manner to at least attempt to confine the energy to a given target volume. So-called radiotherapy plans often function in the above aspects.

[0003] A radiotherapy plan typically includes specified values of each treatment platform parameter during each of a plurality of successive fields. The treatment plan for a radiotherapy session is often automatically generated by a so-called optimization process. As used herein, "optimization" will be understood to mean improving a candidate treatment plan, and not necessarily ensuring that the optimization result is actually the only best solution. Such optimization often includes automatically adjusting one or more physical treatment parameters (often while observing one or more corresponding constraints in these aspects), and mathematically calculating possible corresponding treatment outcomes (such as dosing levels) to identify a set of given treatment parameters that represents a good compromise between a desired therapy outcome and avoiding unwanted side effects.

[0004] The tumor control probability (TCP) refers to the likelihood of complete tumor destruction when radiation is administered according to a specific radiotherapy plan, and the normal tissue complication probability (NTCP) quantifies the probability of complications occurring in the surrounding normal tissue when the target area is exposed to radiation. These two probabilities can be used to measure the expected benefit of a given treatment plan and the expected damage to normal tissue. However, achieving an appropriate balance in those aspects can be challenging when formulating a radiotherapy plan. In practice, radiotherapy plans are usually formulated by experienced oncologists and medical physicists, and their views may differ, and as a result, there is often a significant difference among planners. Subsequently, the latter may make it difficult to quantify the TCP and NTCP of a given radiotherapy plan. Brief Description of the Drawings

[0005] At least in part, the above needs are met by providing the neural network-based radiotherapy plan described in the following detailed description, particularly when studied in conjunction with the accompanying drawings. In the

[0006] drawings:

[0007] Figure 1 Include block diagrams configured according to various embodiments of these teachings;

[0008] Figure 2 Include flowcharts configured according to various embodiments of these teachings;

[0009] Figure 3 Include schematic diagrams configured according to various embodiments of these teachings;

[0010] Figure 4 Include schematic diagrams configured according to various embodiments of these teachings; and

[0011] Figure 5 Include schematic diagrams configured according to various embodiments of these teachings.

[0012] The elements in the figures are shown for simplicity and clarity and are not necessarily drawn to scale. For example, the size and / or relative positioning of some of the elements in the figures may be exaggerated relative to other elements to assist in improving the understanding of the various embodiments of these teachings. Additionally, common but well-understood elements that are useful or necessary in commercially viable embodiments are often not depicted in order to more readily understand these various embodiments of the teachings. Certain actions and / or steps may be described or depicted in a particular order of occurrence, and those skilled in the art will understand that such specificity of sequence is not actually required. Unless a different specific meaning is otherwise set forth herein, the terms and expressions used herein have the ordinary technical meaning given to such terms and expressions by those skilled in the above-described technical field. Unless otherwise specifically indicated, the word "or" as used herein shall be interpreted in a disjunctive rather than a conjunctive construction. Detailed Description

[0013] Generally, many of these various embodiments can provide for training a neural network that is configured to facilitate radiotherapy planning by generating result treatment efficacy probability information and result treatment complication probability information. These teachings can include accessing a training corpus having a plurality of information features, each of the information features including both a reference radiotherapy dose and at least one corresponding post-treatment patient data. These teachings can then include training a neural network using the training corpus. By one method, the at least one corresponding post-treatment patient data includes patient imagery, such as but not limited to one or more computed tomography images.

[0014] By one method, the neural network includes a probability model. In such a case, the probability model can include a latent space of variations in a radiation dose prediction pipeline.

[0015] By a method, a neural network includes both an encoder and a decoder, which are combined into a U-net shape structure that receives multi-channel radiotherapy input of a specific patient and outputs a corresponding feature map.

[0016] By a method, a neural network can be configured to generate a prior latent space, where at least some positions in the prior latent space encode corresponding radiation dose variants to provide a plurality of radiation dose variant codes. If desired, the neural network can be configured to output a predicted radiation dose result from the plurality of radiation dose variant codes and the feature map.

[0017] By a method, the above-mentioned plurality of information features can include information representing a plurality of different patients and / or a plurality of different radiotherapy methods.

[0018] Configured in this way, these teachings will allow, for example, training a radiotherapy planning option space, which allows users to benefit from a better quantifiable trade-off between TCP and NTCP. These teachings also facilitate the use of probability models in the field of radiotherapy dose planning.

[0019] After a thorough review and study of the following detailed description, these and other benefits may become clearer. Now referring to the drawings, and particularly referring to Figure 1 , an illustrative device 100 compatible with many of these teachings will first be presented.

[0020] In this particular example, the enabling device 100 includes a control circuit 101. As a "circuit", the control circuit 101 thus includes a structure that includes at least one (and typically multiple) conductive paths (such as paths including conductive metals such as copper or silver) that conduct electricity in an ordered manner, and the paths typically also include corresponding electrical components (both passive (such as resistors and capacitors) and active (such as any of various semiconductor-based devices)) to allow the circuit to implement the control aspects of these teachings.

[0021] Such a control circuit 101 can include a fixed-purpose hardwired hardware platform (including but not limited to an application-specific integrated circuit (ASIC) (which is an integrated circuit customized by design for a specific purpose rather than intended for general use), a field-programmable gate array (FPGA), etc.), or can include a partially or fully programmable hardware platform (including but not limited to a microcontroller, a microprocessor, etc.). These architectural options for such structures are well known and understood in the art and do not require further description here. The control circuit 101 is configured to perform one or more of the steps, actions, and / or functions described herein (e.g., by using corresponding programming well understood by those skilled in the art).

[0022] It should be understood that a reference to the singular form "control circuit" can refer to literally a single discrete circuit or to multiple discrete control circuits that can be considered collectively as the "control circuit". Thus, in the absence of any contrary language, a reference to the "control circuit" should be understood to include a single such device or multiple such devices.

[0023] The control circuit 101 is operably coupled to a memory 102. The memory 102 may be integrated into the control circuit 101 or may be physically separated from the control circuit 101 (in whole or in part) as desired. The memory 102 may also be local to the control circuit 101 (where, for example, both share a common circuit board, chassis, power supply, and / or enclosure), or may be partially or fully remote from the control circuit 101 (where, for example, the memory 102 is physically located in another facility, metropolitan area, or even country compared to the control circuit 101).

[0024] In addition to information such as optimization information for a particular patient and information about a particular radiotherapy platform as described herein, the memory 102 may also be used, for example, to non - transitorily store computer instructions that, when executed by the control circuit 101, cause the control circuit 101 to operate as described herein. As an example, according to the following description, the control circuit 101 may be at least partially configured as a neural network. (As used herein, this reference to "non - transitory" will be understood to refer to the non - transient state of the stored content (and thus except when the stored content only constitutes a signal or wave), rather than the volatility of the storage medium itself, and thus includes both non - volatile memory (such as read - only memory (ROM)) and volatile memory (such as dynamic random access memory (DRAM)).

[0025] By an alternative method, the control circuit 101 is also operably coupled to a user interface 103. The user interface 103 may include any of a variety of user input mechanisms (such as, but not limited to, a keyboard and keypad, cursor control device, touch - sensitive display, voice recognition interface, gesture recognition interface, etc.) and / or user output mechanisms (such as, but not limited to, a visual display, audio transducer, printer, etc.) to facilitate receiving information and / or instructions from the user and / or providing information to the user.

[0026] If desired, the control circuit 101 may also be operably coupled to a network interface (not shown). The control circuit 101 configured as such may communicate with other elements (both inside and outside the device 100) via the network interface. Network interfaces, including both wireless platforms and non - wireless platforms, are well understood in the art and need not be described in particular here.

[0027] By a method, some or all of any desired patient-related imaging information can be acquired by a computed tomography apparatus 106 and / or other imaging apparatus 107 known in the art.

[0028] In this illustrative example, the control circuit 101 is configured to ultimately output an optimized energy-based treatment plan (such as, for example, an optimized radiotherapy plan 113). The energy-based treatment plan generally includes specified values of each of the various treatment platform parameters during each of a plurality of successive exposure fields. In this case, the energy-based treatment plan is generated by an optimization process, examples of which are further provided herein.

[0029] By a method, the control circuit 101 can be operatively coupled to an energy-based treatment platform 114 that is configured to deliver a treatment performance energy 112 to a corresponding patient 104 in accordance with the optimized energy-based treatment plan 113, the corresponding patient having at least one treatment volume 105 and one or more organs at risk (represented by first through Nth organs at risk 108 and 109 in Figure 1 ). These teachings generally apply to any one of a variety of energy-based treatment platforms / devices. In a typical application setting, the energy-based treatment platform 114 will include an energy source, such as a radiation source 115 of ionizing radiation 116.

[0030] By a method, the radiation source 115 can be selectively moved along an arcuate path via a gantry (where the path includes at least to some extent the patient itself during treatment delivery). Desirably, the arcuate path can include a complete or near-complete circle. By a method, the control circuit 101 controls the movement of the radiation source 115 along the arcuate path and can correspondingly control when the radiation source 15 starts moving, stops moving, accelerates, decelerates, and / or the speed at which the radiation source 115 travels along the arcuate path.

[0031] As an illustrative example, the radiation source 115 can include, for example, an X-ray source based on a radio frequency (RF) linear particle accelerator (linac). A linac is a particle accelerator that greatly increases the kinetic energy of charged subatomic particles or ions by subjecting the charged particles to a series of oscillating electric potentials along a linear beam line, which can be used to generate ionizing radiation (such as X-rays) 116 and high-energy electrons.

[0032] A typical energy-based treatment platform 114 may also include one or more support devices 110 (such as a reclining chair) for supporting the patient 104 during a treatment session, one or more patient fixation devices 111, a gantry or other movable mechanism that allows selective movement of the radiation source 115, and one or more energy shaping devices (e.g., beam shaping device 117, such as jaws, multi-leaf collimator, etc.) to provide selective energy shaping and / or energy modulation as desired.

[0033] In a typical application setting, it is herein assumed that the patient support device 110 can be selectively controlled by the control circuit 101 to move in any direction (i.e., any X, Y, or Z direction) during an energy-based treatment session. Since the above elements and systems are well understood in the art, further details of these aspects are not provided here unless otherwise related to the description.

[0034] Now referring Figure 2 to, a process 200 that can be performed, for example, in conjunction with the above application setting (and more specifically, via the above control circuit 101) will be described. Generally, the process 200 is used to facilitate the generation of an optimized radiotherapy plan 113, thereby facilitating the treatment of a specific patient with therapeutic radiation according to the optimized radiotherapy plan using a specific radiotherapy platform.

[0035] For the purposes of this example, it will be assumed that the control circuit 101 is at least partially configured as a neural network that facilitates radiotherapy planning by generating resultant treatment efficacy probability information and resultant treatment complication probability information. Without intending to imply any specific limitations in these aspects, the neural network may include a probability model. A probabilistic neural network model is an artificial neural network that can be used for classification and pattern recognition tasks. The term "probabilistic" applies because such a network uses probability theory to make predictions and classify inputs.

[0036] By one method, the above probability model can be configured to include a variational latent space in a radiation dose prediction pipeline. The variational latent space in a neural network prediction pipeline refers to modeling and representing complex data in a lower-dimensional space. In this context, the term "latent" refers to hidden variables or factors that capture the underlying structure of the data. The variational aspect comes from the probabilistic nature of the model, which combines probability distributions to model the latent space. In this example, the variational latent space can learn a low-dimensional representation of the input data that captures the essential features of the latter. The above may include using an encoder-decoder architecture, where the encoder network maps the input data to the latent space and the decoder network reconstructs the input data based on the latent representation. If desired, such an encoder and decoder are combined into a U-network-shaped structure that receives multi-channel radiotherapy inputs for a specific patient and outputs the corresponding feature maps as described herein.

[0037] In place of, or in combination with, the foregoing, if desired, the neural network can be configured to generate a prior latent space, where at least some locations in the prior latent space will encode radiation dose variants to provide a plurality of radiation dose variant codes. In such a case, the neural network can be configured to output a predicted radiation dose result from the plurality of radiation dose variant codes and the feature map. If desired, these teachings will also provide the user with an opportunity (e.g., via the user interface 103 described above) to select a particular plan based on their preferences, guidelines, clinical best practices, etc. In such a case, such preferences can be indicated in the latent space code (e.g., indicating a preference for high TCP or high NTCP, where such preference can be measured in any of a variety of ways, including by the ratio of a reference population, e.g., designating the top 50% TCP / NTCP of a given population as "high").

[0038] Reference Figure 2 Referring to, at block 201, the process 200 provides the control circuit 101 described above access to a training corpus 202 (e.g., stored in the memory 102 described above). The training corpus 202 can include, but is not limited to, a plurality of information features. By one method, each of these information features includes both a reference radiation treatment dose and at least one corresponding post-treatment patient data. The information features described above can include any of a variety of other things. Examples include (optionally or necessarily as desired) but are not limited to patient-specific and / or radiotherapy platform-specific information, such as computed tomography content, contouring / segmentation content, and radiotherapy platform geometry content.

[0039] By one method, the plurality of information features includes information representing a plurality of different patients (including radiotherapy plan instructions and details). These different patients may all have been treated at the same radiotherapy facility / clinic, or at least some of the patients may have been treated at different facilities. By one method, all those patients were treated using the same radiotherapy method. By another method, at least some of the patients were treated using any of a plurality of different radiotherapy methods.

[0040] As a non-limiting example, the post-treatment patient data described above can include patient images, such as but not limited to at least one patient computed tomography image. The latter can include computed tomography images of the patient captured over a period of time (such as within a plurality of spaced-apart diagnostic and / or treatment sessions). These teachings will allow for annotation (e.g., providing segmentation and / or contouring, or highlighting certain features or conditions of interest) of some or all of the images in some respects.

[0041] At block 203, the process 200 then provides to train a neural network using the training corpus 202 described above. Training a neural network with a training corpus includes known areas of the prior art and generally includes using a large dataset of (labeled and / or unlabeled) examples to teach the neural network to recognize patterns and make accurate predictions. Thus, for the sake of brevity, no further description of such training is provided here.

[0042] As described above, the information features can include post-treatment patient data corresponding to the reference radiotherapy dose information described above. The post-treatment patient data can include, for example, patient images, such as one or more computed tomography images. Thus, as an example, a particular pre-administered radiotherapy plan aimed at delivering a radiotherapy dose of X Gy can be associated in the training corpus 202 with one or more post-treatment images depicting the post-treatment state of the (multiple) target (and / or non-target) volumes.

[0043] The result of the above is a trained neural network that can output predicted radiation dose results and can be usefully exploited in the radiotherapy plan option space.

[0044] Continuing to refer Figure 2 , at optional block 204, the trained neural network receives an input 205. The input 205 corresponds to a particular patient and / or radiotherapy platform and can include, for example, computed tomography content (e.g., depicting relevant patient volumes), contouring / segmentation content (identifying particular patient volumes, such as the (multiple) target volumes and / or one or more organs at risk), and radiotherapy platform geometry content, either one, some, or all. As an output, the trained neural network can produce a radiation dose (for one or more patient volumes) that can then be utilized when optimizing a particular radiotherapy plan for a particular patient.

[0045] Also, as shown in optional block 206, these teachings will readily accommodate optimizing radiotherapy plans using neural networks and, more specifically, using the predicted outputs of neural networks. At optional block 207, a so-called optimized radiotherapy plan (e.g., using the radiotherapy platform 114 described above) can then be used to administer therapeutic radiation to the corresponding patient. Since both radiotherapy plan optimization and administering radiotherapy using an optimized plan are well-understood areas of the prior art, no further explanation of those aspects is provided here.

[0046] Further examples consistent with these teachings will now be presented. It should be understood that the specific details of these examples are for illustrative purposes only and are not intended to imply any particular limitations on these teachings.

[0047] Figure 3Illustrates an illustrative framework 300 of a method in these aspects. In this example, the input 301 (including, for example, computed tomography images of (a) patient target volume(s) and / or one or more organs at risk) passes through a probabilistic deep learning model 302 to obtain a predicted dose space 303. The approved dose 304 can be used as a reference for the predicted dose during training. Subsequent computed tomography images 305 provide image-based information 306, which can reflect the effectiveness of one or more previously approved doses in terms of TCP and NTCP. (By one method, the reference dose 304 and the image-based information 306 can be used during training, but not necessarily during inference activities.)

[0048] Figure 4 Illustrates the inference phase 400, and Figure 5 Illustrates the training phase 500 of a probabilistic model that can be used in the above example to predict the dose space. These figures include an encoder 401 (E), a decoder 402 (D), a prior network 403 (PrN), and a posterior network 501 (PoN).

[0049] Regarding Figure 4 the illustrated inference phase 400, the depicted method includes a variational latent space in the dose prediction pipeline. Specifically, the encoder 401 and the decoder 402 are combined into a convolutional neural network or a transformer structure based on the U-net shape, which takes a multi-channel three-dimensional input 404 and outputs a three-dimensional feature map 405. The prior network 403 constructs a prior latent space 406, where (in this example) each position in this space 406 encodes a dose variant. The final predicted dose 407 is derived from the dose variant code and the three-dimensional feature map 405 from the U-net structure.

[0050] Referring Figure 5 to the illustrated training phase 500, the posterior network 501 uses a three-dimensional (and multi-channel) input 503 (denoted as X in the following equations) and a reference dose map (denoted as Y in the following equations) to generate another latent space 502 (here called the posterior latent space and denoted as z in the following equations). There is a KL divergence loss that minimizes the difference between the prior latent distribution 406 (denoted as P in the following equations) and the posterior latent distribution 502 (denoted as Q in the following equations). Thus,

[0051] L KLD = F1(TCP(Y), NTCP(Y)) * D KL (Q(z|Y, X) || P(z|X)),

[0052] D KL (Q || P) = E z～Q (log Q - log P)

[0053] where TCP(Y) and NTCP(Y) are the TCP / NTCP reflections from the planned reference dose map Y based on subsequent computed tomography images, and F1(·, ·) is an activation function (or, if desired, a neural network-based module) for transferring the effects from TCP(Y) and NTCP(Y) to L KLD Calculation.

[0054] In addition to L KLD , there are other functions that can minimize the difference between the predicted dose Y and the reference dose Y. For example,

[0055]

[0056] where measures the distance between Y and , and F2(·, ·) is an activation function (or, if desired, a neural network-based module) for transferring the effects from TCP(Y) and NTCP(Y) to L d Calculation. (By one method, these teachings will make TCP(Y) and NTCP(Y) vector-valued quantities that can be evaluated based on the analysis of subsequent computed tomography images and other clinical information from the patient regarding the degree of local control and side effects.)

[0057] Configured as such, these teachings will support, for example, the construction of an automated dose planning space that can reflect and / or leverage toxicity factors from subsequent computed tomography images. These teachings can also provide a latent space that captures meaningful variations, which in turn can allow the user to balance the outcome with toxicity by considering different values during the inference phase described above, thereby selecting a plan.

[0058] Those skilled in the art will recognize that various modifications, changes, and combinations can be made to the above embodiments without departing from the scope of the present invention, and such modifications, changes, and combinations should be considered within the scope of the inventive concept.

Claims

1. A method of training a neural network, the neural network being configured to facilitate radiation therapy planning by generating outcome treatment efficacy probability information and outcome treatment complication probability information, the method comprising: accessing a training corpus comprising a plurality of information features, each of the information features comprising both a reference radiation therapy dose and at least one corresponding post-treatment patient data; The neural network is trained using the training corpus.

2. The method of claim 1, wherein the neural network comprises a probabilistic model.

3. The method of claim 2, wherein the probabilistic model comprises a potential space of variations in a radiation dose prediction pipeline.

4. The method of claim 2, wherein the neural network comprises an encoder and a decoder, the encoder and the decoder being combined into a U-network shape structure, the U-network shape structure receiving a multi-channel radiotherapy input for a specific patient and outputting a corresponding feature map.

5. The method of claim 2, wherein the neural network is configured to generate an a priori latent space, wherein at least some positions in the a priori latent space encode corresponding radiation dose variants to provide a plurality of radiation dose variant codes.

6. The method of claim 5, wherein the neural network is configured to output predicted radiation dose results from the plurality of radiation dose variant codes and feature maps.

7. The method of claim 1, wherein the at least one corresponding post-treatment patient data comprises a patient image.

8. The method of claim 7, wherein the patient images include at least one patient computed tomography image.

9. The method of claim 1, wherein the plurality of information features includes information representative of a plurality of different patients.

10. The method of claim 1, wherein the plurality of information features includes information representative of a plurality of different radiation therapy approaches.

11. A device comprising: A control circuit configured as a neural network that facilitates radiation therapy treatment planning by generating outcome treatment efficacy probability information and outcome treatment complication probability information, wherein the neural network is trained using a training corpus including a plurality of information features, each of the information features including both a reference radiation therapy dose and at least one corresponding post-treatment patient data.

12. The apparatus of claim 11, wherein the neural network comprises a probabilistic model.

13. The apparatus of claim 12, wherein the probabilistic model comprises a potential space of variations in a radiation dose prediction pipeline.

14. The apparatus of claim 12, wherein the neural network comprises an encoder and a decoder, the encoder and the decoder being combined into a U-network shape structure that receives a multi-channel radiotherapy input for a specific patient and outputs a corresponding feature map.

15. The apparatus of claim 12, wherein the neural network is configured to generate an a priori latent space, wherein at least some positions in the a priori latent space encode corresponding radiation dose variants to provide a plurality of radiation dose variant codes.

16. The apparatus of claim 15, wherein the neural network is configured to output predicted radiation dose results from the plurality of radiation dose variant codes and feature maps.

17. The apparatus of claim 11, wherein the neural network is configured to receive input comprising at least one of: computed tomography content, contouring / segmentation content, and radiation therapy platform geometry content.

18. The apparatus of claim 17, wherein the neural network is configured to receive inputs comprising each of: computed tomography content, contouring / segmentation content, and radiation therapy platform geometry content.

19. The apparatus of claim 11, wherein the plurality of information features includes information representative of a plurality of different patients.

Images