Deep regression estimation of arterial input function for dynamic pet

Abstract

According to various embodiments, machine learning techniques for determining a pharmacokinetic parameter of a radiotracer in a patient are presented. The techniques include: obtaining a time sequence of positron emission tomography (PET) scans of a target tissue of the patient, where the time sequence of PET scans includes images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, which provides features extracted from the time series of the PET scans; providing the features extracted from the time series of the PET scans to a second neural network system, which provides an estimated arterial input function; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

Description

DEEP REGRESSION ESTIMATION OF ARTERIAL INPUT FUNCTION FOR DYNAMIC PETGovernment Support

[0001] This invention was made with government support under grants EB031023, EB031798, and CA140204 awarded by the National Institutes of Health. The government has certain rights in this invention.Related Application

[0002] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 532,718, entitled “Deep Regression Estimation of Arterial Input Function for Dynamic PET,” filed August 15, 2023.Field

[0003] This disclosure relates generally to positron emission tomography (PET) imaging.Background

[0004] Positron emission tomography (PET) is a functional imaging modality that measures a local concentration or amount of a radiolabeled tracer molecule in a target tissue. The power of PET lies in its molecular specificity and high sensitivity. By using a particular radiolabeled molecule (a radiotracer) in nanomolar or picomolar quantities, one can monitor the interaction of that molecule with the body’s physiological processes such as blood flow, metabolism, and neurotransmission, etc. Dynamic PET is a technique where a dynamic time sequence of PET images isacquired following radiotracer injection. Based on the dynamic PET data, one can perform pharmacokinetic modeling to quantify physiological and biochemical parameters such as receptor density in a volume of tissue, rate of trapping / influx of a radiotracer, competition between endogenous and exogenous ligands, etc.

[0005] In pharmacokinetic modeling, sets of ordinary differential equations are used to relate the dynamics of the radiotracer and all its possible states to the resultant PET images. Models for each PET radiotracer are typically derived from a one-, two-, or three-compartment model in which the concentration of radiotracer in the blood as a function of time serves as the model’s input function. The coefficients of the differential equations in the models are then solved from the data to reflect inherent kinetic properties of the radiotracer.

[0006] Having an accurate input function is important for a successful kinetic modeling analysis. Prior techniques for obtaining an input function involve repeatedly drawing blood samples from patient’s radial artery from the time of the radiotracer injection (i.e. , start of dynamic PET) to the end of the time sequence of PET scans. The blood samples are then centrifuged to separate plasma, and the concentration of the radiotracer in plasma is measured in a well counter. The radioactive metabolites of the radiotracer are also analyzed and measured from blood samples from multiple time points using high-performance liquid chromatography (HPLC). The plasma activities are then corrected for radiolabeled metabolite fraction to produce an arterial input function (“AIF”) that is then metabolite-corrected. This correction adds a step to the process, which can increase inaccuracy of the final kinetic analysis.

[0007] While it is reasonably safe to perform serial arterial sampling, a typical dynamic PET scan lasts 90 to 180 minutes, and tolerance for such procedures can vary, especially in ailing or elderly populations. This could slow subject accrual orcontribute to subject dropout rates in longitudinal studies. The risks of arterial cannulation also increase in aging populations, and some older patients may require medications, which affect the blood clotting cascade and pose a contraindication to arterial line placement. These considerations could dissuade researchers from using otherwise promising radiotracers due to avoidance of AIF measurement in multicenter, longitudinal trials. In addition, some hospitals may not have the capability to acquire the blood measurements.

[0008] An alternative to an AIF determined from repeated arterial blood draws is to use activities measured from the vascular compartment in PET images to form an image-derived AIF. For example, in cardiac PET, radioactivity measures from left ventricle blood pool are used to estimate the arterial input function. In brain PET, dynamic activity in the carotid arteries can be measured from images and used as the input function. Accurate image-derived AIF estimation is contingent on precise blood pool segmentation in dynamic PET images, which may be compromised by limited accuracy due to the partial volume effect and noise. Further, the high noise and limited spatial resolution of PET can contribute to observed variability in an image-derived AIF, which can introduce error into the kinetic modeling results. Yet further, deriving an AIF from an image usually does not account for the metabolite of the radiotracer. Other methods, such as blind source separation, have been investigated, and yet are prone to failure due to high image noise.Summary

[0009] According to various embodiments, a machine learning method of determining a pharmacokinetic parameter of a radiotracer in a patient is presented.The method includes: obtaining a time sequence of positron emission tomography(PET) scans of a target tissue of the patient, wherein the time sequence of PET scans comprises images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, from which features extracted from the time series of the PET scans are obtained, wherein the first neural network system comprises a self-supervised learning neural network; providing the features extracted from the time series of the PET scans to a second neural network system, from which provides an estimated arterial input function is obtained, wherein the second neural network system comprises a deep regression network, wherein the estimated arterial input function obtained from the second neural network is metabolite-corrected; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

[0010] Various optional features of the above method embodiments include the following. The method may include: adjusting a dosing of a therapeutic for the patient based at least in part on the estimate of the pharmacokinetic parameter of the radiotracer; and administering the therapeutic to the patient. The pharmacokinetic parameter of the radiotracer may include an amount of the radiotracer that is bound to a target receptor site. The pharmacokinetic parameter of the radiotracer may include an amount of the radiotracer in the target tissue. The self-supervised learning neural setwork system may include a variational autoencoder. The variational autoencoder may include a vector-quantized variational autoencoder. The features extracted from the time series of the PET scans may be obtained from bottleneck layers of the variational autoencoder. The features extracted from the time series of the PET scans may include compact latent representations of the time series of PET scans. The deep regression network may include: a multi-layer perceptron layer applied to a spatialdimension to learn from features extracted from the time series of PET scans within a single point in time; and a multi-layer perceptron layer applied to a temporal dimension to learn relationships between the features extracted from the time series of PET scans across different points in time. The estimated arterial input function may be for a parent fraction.

[0011] According to various embodiments, a non-transitory computer-readable medium comprising instructions that, when executed by an electronic processor, configure the electronic processor to determine a pharmacokinetic parameter of a radiotracer in a patient by performing actions is presented. The actions include: obtaining a time sequence of positron emission tomography (PET) scans of a target tissue of the patient, wherein the time sequence of PET scans comprises images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, from which features extracted from the time series of the PET scans are obtained, wherein the first neural network system comprises a selfsupervised learning neural network; providing the features extracted from the time series of the PET scans to a second neural network system, from which provides an estimated arterial input function is obtained, wherein the second neural network system comprises a deep regression network, wherein the estimated arterial input function obtained from the second neural network is metabolite-corrected; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

[0012] Various optional features of the above computer-readable medium embodiments include the following. The actions may further include: adjusting a dosing of a therapeutic for the patient based at least in part on the estimate of thepharmacokinetic parameter of the radiotracer; and administering the therapeutic to the patient. The pharmacokinetic parameter of the radiotracer may include an amount of the radiotracer that is bound to a target receptor site. The pharmacokinetic parameter of the radiotracer may include an amount of the radiotracer in the target tissue. The self-supervised learning neural setwork system may include a variational autoencoder. The variational autoencoder may include a vector-quantized variational autoencoder. The features extracted from the time series of the PET scans may be obtained from bottleneck layers of the variational autoencoder. The features extracted from the time series of the PET scans may include compact latent representations of the time series of PET scans. The deep regression network may include: a multi-layer perceptron layer applied to a spatial dimension to learn from features extracted from the time series of PET scans within a single point in time; and a multi-layer perceptron layer applied to a temporal dimension to learn relationships between the features extracted from the time series of PET scans across different points in time.

[0013] According to various embodiments, a system for determining a pharmacokinetic parameter of a radiotracer in a patient is presented. The system includes an electronic processor and a non-transitory computer-readable medium comprising instructions that, when executed by the electronic processor, configure the electronic processor to determining a pharmacokinetic parameter of a radiotracer in a patient by performing actions comprising: obtaining a time sequence of positron emission tomography (PET) scans of a target tissue of the patient, wherein the time sequence of PET scans comprises images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, from which features extracted from the time series of the PET scans are obtained, wherein the first neural network system comprises a self-supervised learning neural network;providing the features extracted from the time series of the PET scans to a second neural network system, from which provides an estimated arterial input function is obtained, wherein the second neural network system comprises a deep regression network, wherein the estimated arterial input function obtained from the second neural network is metabolite-corrected; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

[0014] Combinations, (including multiple dependent combinations) of the above-described elements and those within the specification have been contemplated by the inventors and may be made, except where otherwise indicated or where contradictory.Brief Description of the Drawings

[0015] Various features of the examples can be more fully appreciated, as the same become better understood with reference to the following detailed description of the examples when considered in connection with the accompanying figures, in which:

[0016] Fig. 1 illustrates usage of an AIF to model a pharmacokinetic parameter of a radiotracer, according to various embodiments;

[0017] Fig. 2 illustrates a compartment model of a radiotracer, according to various embodiments;

[0018] Fig. 3 illustrates a relationship between a total plasma concentration, a parent fraction, and the metabolite-corrected arterial input function, according to various embodiments;

[0019] Fig. 4 is a schematic diagram of a machine learning system for determining a pharmacokinetic parameter of a radiotracer in a patient, according to various embodiments;

[0020] Fig. 5 illustrates example reconstructions produced by an implementation of an example non-limiting embodiment;

[0021] Fig. 6 illustrates examples of estimated metabolite-corrected AIFs for six patients, juxtaposed with the ground truth AIFs, as produced by the implementation of an embodiment; and

[0022] Fig. 7 is a flow diagram of a method of determining a pharmacokinetic parameter of a radiotracer in a patient, according to various embodiments.Description of the Examples

[0023] Reference will now be made in detail to example implementations, illustrated in the accompanying drawings. Wherever convenient, the same reference numbers will be used throughout the drawings to refer to the same or like parts. In the following description, reference is made to the accompanying drawings that form a part thereof, and in which is shown by way of illustration specific exemplary examples in which the invention may be practiced. These examples are described in sufficient detail to enable those skilled in the art to practice the invention and it is to be understood that other examples may be utilized and that changes may be made without departing from the scope of the invention. The following description is, therefore, merely exemplary.

[0024] Some embodiments provide a deep learning-based method to directly estimate an AIF from dynamic PET image data without human assistance. This provides a reliable and robust method for estimating the AIF to avoid the arterial bloodsampling that is otherwise required for dynamic PET pharmacokinetics with existing and emerging radiotracers.

[0025] Some embodiments use or include a deep regression framework for direct metabolite-corrected AIF estimation from dynamic PET, obviating the need for image segmentation and mitigating potential errors in imaging-based AIF estimation. Some embodiments first employ a variational autoencoder in a self-supervised learning task to reconstruct the input dynamic PET. Some embodiments then extract features from the bottleneck layers, which serve as compact latent feature representations of the dynamic PET images. According to some embodiments, these latent representations are then combined along the temporal axis and input into a trained deep regression network, which provides an estimate of the AIF.

[0026] Furthermore, and to summarize various embodiments, estimating AIFs directly from 4D dynamic PET scans can be computationally demanding because of the large size of sequential volumetric images. Some embodiments first compress each 3D dynamic PET volume into more compact latent representations. Some embodiments then train a regression model to predict AIF using a series of these compressed representations. Some embodiments incorporate self-supervised learning to compress the PET volumes. According to some embodiments, this is achieved by employing an hourglass shaped autoencoder to first learn to reconstruct the input PET volumes. According to some embodiments, the feature representations are then extracted from the bottleneck layers, serving as the compact latent representations of the input images. These representations are concatenated for each patient based on temporal information and fed into a regression model to estimate metabolite-corrected AIF for the patient, according to some embodiments.Embodiments are highly efficient and can be trained and tested on most modern GPUs.

[0027] Some embodiments provide at least three advantages over the prior art. A first disadvantage of the prior art is that it relies on precise blood pool segmentation in dynamic PET images, and existing segmentation techniques are not sufficiently accurate due to the partial volume effect and noise. Some embodiments overcome this first disadvantage of the prior art through the use of a self-supervised learning neural network, such as a variational autoencoder, which obviates the need for segmentation by providing extracted features as compact latent representations of the arterial space. A second disadvantage of the prior art is that the high noise and limited spatial resolution of PET can contribute to observed variability in an AIF calculated therefrom, and such variability reduces the accuracy of the ultimate kinetic modeling results. Some embodiments overcome this second disadvantage of the prior art through the use of both a self-supervised learning neural network and a deep regression network operating together in combination, which reduces the negative effects of the high noise and limited spatial resolution of PET. A third disadvantage of the prior art is that image-derived AIF are not typically metabolite-corrected, and applying a separate metabolite correction step can decrease accuracy. Some embodiments overcome this third disadvantage of the prior art through the use of a deep regression network that provides a metabolite-corrected AIF estimation derived directly from the extracted features, without using a dedicated metabolite-correction step.

[0028] A reduction to practice was applied to clinical data and its performance was assessed using a weighted relative error (WRE) weighted by the sensitivity of kinetic modeling. The reduction to practice achieved a small WRE of 1.2%±0.56,highlighting the effectiveness of the disclosed technique for AIF estimation and its potential for clinical application.

[0029] These and other features and advantages are shown and described herein in reference to the accompanying figures.

[0030] Fig. 1 illustrates usage 100 of an AIF 102 to model 104 a pharmacokinetic parameter of a radiotracer, according to various embodiments. According to various embodiments, the radiotracer may be any radioactive substance, including, but not limited to, radioactive: water, glucose, neurotransmitter, a protein, or any of a variety of ligands, whether originally endogenous or exogenous. By way of non-limiting example, Fig. 1 depicts a model of tissue concentration 106 of the radiotracer over time as the pharmacokinetic parameter. Embodiments may obtain an estimated AIF 102 using a machine learning system, e.g., as shown and described herein in the Appendix.

[0031] In general, an AIF 102 represents a radiotracer concentration input to a volume of interest as a function of time. For example, an AIF 102 may represent a radiotracer concentration in blood plasma in an artery at each of a plurality of timesteps, e.g., each point in time, within a time interval. An AIF 102 may be represented in non-transitory computer readable form; for example, and AIF 102 may be implemented as a signal according to various embodiments. An AIF 102 may relate a concentration to time by way of a plurality of pairs (x / , yz) for / < M for a value of M, e.g., M = 128. The concentration of an AIF 102 may be represented in a.u., Hounsfield units, or kilobecquerel per unit of volume (e.g., cubic centimeter), by way of nonlimiting examples. The time in an AIF 102 may be represented in minutes or seconds, by way of non-limiting example.

[0032] Embodiments may use the estimated AIF 102 to model 104 any of a variety of pharmacokinetic parameters, by way of non-limiting examples: concentration 106 or amount of the parent radiotracer (or “authentic radiotracer,” i.e., excluding radiotracer metabolites) in a target tissue, concentration or amount of the parent radiotracer in a fluid (e.g., blood plasma), concentration or amount of free radiotracer in a target tissue (e.g., including metabolites), concentration or amount of the parent radiotracer in a target tissue that is bound to a specific receptor site, concentration or amount of the parent radiotracer in a target tissue that is non-specifically bound to cell components other than a targeted receptor site, etc. The modeling 104 may be used to estimate any of these parameters at a specific point in time, for example.

[0033] Note that, according to some embodiments, a radiotracer model 104 may be used to infer information about an endogenous or exogenous molecule that is not radioactive, e.g., based on a known competition relationship between the radiotracer and the molecule for a specific binding site.

[0034] Fig. 2 illustrates a compartment model 200 of a radiotracer, according to various embodiments. By way of non-limiting example, Fig. 2 illustrates a two-tissue compartment model. For the model of Fig. 2, the term Cprepresents arterial radiotracer amount or concentration, the term Ci represents radiotracer amount or concentration in a first compartment (non-displaceable binding, e.g. free radiotracer in tissue), the term C2 represents radiotracer amount or concentration in a second compartment (displaceable binding, e.g., specifically bound radiotracer), and the terms Ki, k2, ks, and k4 represent transfer coefficients. In general, a compartment model such as the compartment mode 200 may use any of a variety of compartments to respectively represent any of a variety of pharmacokinetic parameters, e.g., as explicitly disclosed herein or otherwise. Each compartment may represent an amountor concentration of a radiotracer or other molecule in any of a variety of locations (or “states”) over time. Further, in general, a compartment model may be associated with a set of one or more differential equations that describe the pharmacokinetic parameters of each of the compartments over time. An embodiment, such as shown and described herein, with measured or estimated AIF Cp, may be used to estimate some quantities, such as one or more transfer coefficients from dynamic PET scans, thereby permitting the estimates of other biological quantities, e.g., binding potential. An embodiment, such as shown and described herein, may be used to obtain an estimated AIF, which may be used to estimate some quantities in a compartment model, such as Cpand / or one or more transfer coefficients, thereby permitting the compartment model to provide estimates of other quantities, e.g., with other quantities derived from a time sequence of PET scans.

[0035] Fig. 3 illustrates a relationship 300 between a total plasma concentration (square markers in the AIF chart 302), a parent fraction (curve in the parent fraction chart 304), and the metabolite-corrected AIF (curve on the AIF chart 302) according to various embodiments. In general, a PET scan may be used to measure an amount or concentration of a radiotracer in a given compartment. However, the raw radioactivity measurement may include both the radiotracer and any metabolites thereof. This is referred to as the measured “total plasma.” However, typically only the original radiotracer, or “parent” is of interest, and the parent concentration is typically lower than the total plasma, as shown in Fig. 3. For example, the original radiotracer is typically used for the AIF. Accordingly, techniques that measure radioactivity of a radiotracer typically account for the presence of both the original parent radiotracer and any metabolites thereof in the total plasma. In the prior art, such accounting may utilize known or measured mathematical relationships betweena parent and metabolite(s) and calculate an AIF from the total plasma using such relationships in an additional step.

[0036] Some embodiments advantageously estimate an AIF without requiring such additional calculations or measurements. For example, some embodiments, e.g., as shown and described in reference to Figs. 4-7, use machine learning to estimate an AIF ab initio, without requiring further processing of a total plasma amount or concentration to account for metabolites.

[0037] Fig. 4 is a schematic diagram of a machine learning system 400 for determining a pharmacokinetic parameter of a radiotracer in a patient, according to various embodiments. A detailed description of the machine learning system 400 follows. The following establishes notation (rather than computation) used to describe various embodiments in reference to Fig. 4. A 3D dynamic PET image captured at a discrete time point t is shown in Fig. 4 as xt. Some embodiments transform the 4D sequential dynamic PET scans of the patient, shown in Fig. 4 as x = {xt}=0, into a 1 D AIF, shown in Fig. 4 as a. Due to the considerable size of the 4D sequence, directly training a DNN to learn this mapping may be computationally intractable, such that it cannot be reasonably performed by a computer. To alleviate this computational burden, some embodiments divide the regression tasks into a two-stage framework. As illustrated in Fig. 4, in the first stage 402, a network learns to represent xtusing a more compact latent representation, which is shown in Fig. 4 as xt, while in the second stage 404, a regression network takes in the concatenated latent representation, which is shown in Fig. 4 as x = {x t=0, and outputs a. The details of each stage are described presently.

[0038] In the first stage 402 of the example non-limiting embodiment, a selfsupervised learning neural network obtains a latent representation. An input image(xt) is transformed into its extracted features (xt) by training an hourglass-shaped network on a task to reconstruct the input image. After training, the features extracted from the “bottleneck” layers serve as the latent representation (xt). Although a conventional autoencoder could compress the spatial dimensions of the input, they typically require an expansion of the feature dimension (i.e. , the channel dimension) to achieve a satisfactory reconstruction of the input. However, even with an increased feature dimension, conventional autoencoders still struggle to produce high-quality reconstruction. This implies that information is lost during the compression stage with conventional autoencoders, resulting in feature representations that do not adequately represent the input. To address this issue, some embodiments leverage a vector- quantized variational autoencoder 2 (VQ-VAE-2), which compresses both feature and spatial dimensions while reconstructing high-fidelity images. A component of the VQ- VAE-2 is the quantization layers, which effectively compresses varying hierarchical features into the hierarchical latent code-books of smaller discrete representations. For each given input, the latent representations are generated by sampling the nearest elements from the latent codebooks. Some embodiments adopt the two-latent codebook setting, with the top latent codebook encoding global information, and a bottom latent codebook responsible for representing local details. According to some embodiments, the dimension of each latent representation may be empirically determined to be 16, meaning that for an input image (xt) of size H x w x D, its corresponding latent representation (xt) has the size 16 x H / 8 x w / 8 x D / 8 , which is 1 / 32 of the original image size.

[0039] The second stage 404 includes a deep regression network, which transforms input features into a metabolite-corrected AIF. Once a latent representation (xt) is obtained for each time step, they are concatenated along thetemporal dimension and sent to a regression model to estimate the corresponding AIF.The regression model may take into account both the temporal relationships between the extracted features over time (the xt’s) and the quantitative information provided by the extracted features at a point in time (an xt). A conventional fully connected network faces a large computational burden, while a conventional convolutional neural network struggles to capture temporal information due to its limited ability to model long-range spatial or temporal correspondences. While transformers excel at handling spatial-temporal information, their implementation is non-trivial. Some embodiments adopt the MLP-Mixer, which offers the advantage of capturing long-range correspondences while featuring a straightforward architecture to simplify implementation. The MLP-Mixer according to some embodiments may include two types of MLP layers: one applied to the spatial dimension to learn from each xtwithin a single point in time, and another applied to the temporal dimension to learn the relationships between xt’s across different points in time. Some embodiments may utilize the default setting of MLP-Mixer, setting the number of blocks to 12 and the expansion factor to 4.

[0040] Figs. 5 and 6 illustrate results provided by an implementation of an example non-limiting embodiment. The implementation utilized the system shown in Fig. 4. The implementation was applied to a dataset representing 54 healthy subjects and included 90-minute dynamic brain PET images with 30 discrete time points (i.e., T = 29) after a bolus injection of11C-PDA-713 radiotracer obtained using a Siemens HRRT PET system. Each brain PET was converted to SUV and skull-stripped using publicly available software. The dataset was divided into a ratio of 16-1 -3 for training, validation, and testing. The system of the implementation was trained for 200 epochs on an NVIDIA A6000 GPU. The learning rates for training the VQ-VAE-2 and MLP-Mixer were set to 3e-4 and 1e-3, respectively. The mean squared error (MSE) was employed as the loss function for the self-supervised reconstruction task, while the L1 loss was used for training the deep regression network. The regression performance was evaluated using the weighted relative error (WRE), which is defined as WRE = St=owtwhere atand atdenote, respectively, the predicted and ground truthAIFs at time point t, and wtis a weighting factor denoting the sensitivity of the kinetic modeling to the AIF.

[0041] In the initial stage of the MLP-Mixer, a linear embedding layer was employed to map each latent representation to a vector. In the implementation, three different embedding configurations were evaluated to determine if a larger or smaller embedding dimension is more favorable. The evaluated dimensions were 2048, 512, and 128. The quantitative results for each configuration are shown in the Table. The results show that the configuration of 512 embeddings achieved a smaller mean WRE of 1.2%, suggesting that this embedding dimension is more suitable. Note that all configurations used only approximately 5 GiB of GPU memory, indicating that the implementation of an embodiment was lightweight and may be practically employed on most modem GPUs.Table: Quantitative Results of the Implementation

[0042] Fig. 5 illustrates example reconstructions 500 produced by the implementation of an embodiment. In particular, the reconstructions were produced by the VQ-VAE-2 of the example implementation. The reconstructions show that theVQ-VAE-2 can accurately reconstruct input dynamic PET images, with only minor blurring. The blurring shown in the reconstructions 500 may be a result of using MSE as the reconstruction loss function, which is commonly thought to cause averaging among samples and is therefore correctable.

[0043] Fig. 6 illustrates examples of estimated metabolite-corrected AIFs 600 for six patients, juxtaposed with the ground truth AIFs, as produced by the implementation of an embodiment. As is clearly shown in Fig. 6, the AIFs produced by the example implementation are nearly identical to the corresponding ground truths.

[0044] The quantitative and qualitative results shown and described in reference to Figs. 5 and 6 both indicate that embodiments can effectively predict AIFs. Adjustments to hyperparameters, loss functions, and inputs (e.g., integrating MRI scans) may further increase accuracy.

[0045] Fig. 7 is a flow diagram of a method 700 of determining a pharmacokinetic parameter of a radiotracer in a patient, according to various embodiments. The method 700 may be implemented using a system such as is shown and described in reference to Fig. 4. The method 700 may utilize a machine learning system to obtain an estimated metabolite-corrected AIF from a time sequence of 3D PET scans.

[0046] At 702, the method 700 includes obtaining a time sequence of 3D PET scans, e.g., a 4D scan of a target tissue of the patient. The time sequence of PET scans may be obtained directly from a PET scanner, or may be retrieved from memory, e.g., persistent electronic memory in a healthcare database. The time sequence of PET scans may be obtained over a network, for example. The time series of PET scans may be of the target tissue once the patient has been administered aradiotracer, e.g., orally or by injection. Thus, the time sequence of PET scans includes images of the radiotracer in the target tissue, e.g., brain tissue.

[0047] Any radiotracer may be used, including, but not limited to, examples explicitly disclosed herein. Non-limiting examples of suitable radiotracers include radiotracers targeting dopaminergic brain receptors, glutamatergic brain receptors, serotonergic brain receptors, or cholinergic brain receptors. For example, suitable radiotracers include those targeting TSPO (e.g., [18F]DPA-714, [11C]DPA-713, [18F]FEPPA, or [18F]LW223), those targeting MAO-B (e.g., [18F]SMBT1 ), those targeting COX2 (e.g., [11C]MC-1 ), those targeting COX1 (e.g., [11C]PS-13), those targeting amino acid decarboxylase (e.g., [18F]FDOPA), or those targeting mGluR5 (e.g., [18F]FPEB). Other suitable radiotracers may include radiotracers targeting cancer (e.g., FDG).

[0048] At 704, the method 700 includes providing the time series of the PET scans to a first neural network system, which provides features extracted from the time series of the PET scans. The first neural network system may be as shown and described herein in reference to Fig. 4 (e.g., the first stage 402). By way of non-limiting example, the first neural network system may include a self-supervised learning neural network, such as, by way of non-limiting example, an autoencoder. According to some embodiments, and by way of non-limiting example, a vector-quantized variational autoencoder may be used. The features extracted from the time series of the PET scans may be obtained from bottleneck layers of the variational autoencoder. The features extracted from the time series of the PET scans may include compact latent representations of the time series of PET scans.

[0049] At 706, the method 700 includes providing the features extracted from the time series of the PET scans to a second neural network system, which providesan estimated arterial input function. By way of non-limiting example, the second neural network system may be as shown and described herein in reference to Fig. 4 (e.g., the second stage 404). The second neural network system may include a deep regression network according to some embodiments. According to some embodiments, a multi-layer perceptron may be used, such as an MLP-Mixer neural network. According to some embodiments, the deep regression network includes: a multi-layer perceptron layer applied to a spatial dimension to learn from features extracted from the time series of PET scans within a single point in time, as well as a multi-layer perceptron layer applied to a temporal dimension to learn relationships between the features extracted from the time series of PET scans across different points in time. However, embodiments are not limited to such neural networks. For example, according to some embodiments, a transformer may be used. The estimated arterial input function obtained from the second neural network may be metabolite-corrected, e.g., as shown and described in reference to Fig. 3. The arterial input function may be for a parent fraction of the radiotracer.

[0050] At 708, the method 700 includes modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans. By way of non-limiting example, the actions of 708 may include actions as shown and described herein in reference to one or both of Figs. 1 and 2. By way of non-limiting examples, the pharmacokinetic parameter of the radiotracer may be an amount of the radiotracer that is bound to a target receptor site or an amount of the radiotracer in the target tissue. The pharmacokinetic parameter may be any pharmacokinetic parameter, not limited to the examples explicitly recited herein. The modeling may use a compartment model, e.g., one or more differential equations representing a compartment model. The modeling may estimate an amountor concentration of the pharmacokinetic parameter in one or more particular states (e.g., compartments) over time.

[0051] At 710, the method 700 includes outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling. For example, the estimate of the pharmacokinetic parameter may be an amount or concentration of the radiotracer in a particular state (e.g., compartment) at a particular time. The outputting may be performed by display on a monitor, over a network, and / or by transfer to a different process or system, such as a medical records system.

[0052] At 712, the method 700 includes adjusting a dosing of a therapeutic for the patient based at least in part on the estimate of the pharmacokinetic parameter of the radiotracer. The actions of 712 may adjust a standardized dosing upwards or downwards based on the pharmacokinetic modeling. For example, if the modeling of 710 indicates that a compound tracked by the radiotracer will be in a greater concentration in a particular compartment (e.g., location or state) than accounted for according to the standardized dosing at a future time, then the dosing may be adjusted downward. According to some embodiments, the therapeutic may be treatment for neurodegenerative / neurovegetative diseases, psychiatric disorders, or tumors. Parkinson’s disease may be treated by administering L-DOPA. Alzheimer’s disease may be treated by administering lecanemab or Kisunla (donanemab-azbt). Tumors may be treated by administering a radiopharmaceutical (e.g., Lu177-DOTATHERA, Lu177-PSMA, or Ac225-PSMA), an alkylating agent (e.g., altretamine, busulfan, carboplatin, carmustine, cisplatin, cyclophosphamide, dacarbazine, ifosfamide, lomustine, melphalan, temozolomide, or trabectedin), an antimetabolite (e.g., 5- fluorouracil, 6-mercaptopurine, azacitidine, capecitabine, clofarabine, cytarabine, floxuridine, fludarabine, gemcitabine, methotrexate, pemetrexed, pentostatin,pralatrexate, or trifluridine / tipiracil), an alkaloid (e.g., vincristine, vinblastine, vinorelbine, paclitaxel, docetaxel, etoposide, teniposide, irinotecan, or topotecan), or an antibiotic (e.g., daunorubicin, doxorubicin, doxorubicin liposomal, epirubicin, idarubicin, or valrubicin).

[0053] At 714, the method 700 includes administering the therapeutic to the patient. The administrating may include injection, providing medication for oral ingestion, or any other kind of therapeutic administration.

[0054] Certain examples can be performed using a computer program or set of programs. The computer programs can exist in a variety of forms both active and inactive. For example, the computer programs can exist as software program(s) comprised of program instructions in source code, object code, executable code or other formats; firmware program(s), or hardware description language (HDL) files. Any of the above can be embodied on a transitory or non-transitory computer readable medium, which include storage devices and signals, in compressed or uncompressed form. Exemplary computer readable storage devices include conventional computer system RAM (random access memory), ROM (read-only memory), EPROM (erasable, programmable ROM), EEPROM (electrically erasable, programmable ROM), flash memory, and magnetic or optical disks or tapes.

[0055] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented using computer readable program instructions that are executed by an electronic processor.

[0056] These computer readable program instructions may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the electronic processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0057] In embodiments, the computer readable program instructions may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, statesetting data, configuration data for integrated circuitry, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++, or the like, and procedural programming languages, such as the C programming language or similar programming languages. The computer readable program instructions may execute entirely on a user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.

[0058] As used herein, the terms “A or B” and “A and / or B” are intended to encompass A, B, or {A and B}. Further, the terms “A, B, or C” and “A, B, and / or C” areintended to encompass single items, pairs of items, or all items, that is, all of: A, B, C, {A and B}, {A and C}, {B and C}, and {A and B and C}. The term “or” as used herein means “and / or.”

[0059] As used herein, language such as “at least one of X, Y, and Z,” “at least one of X, Y, or Z,” “at least one or more of X, Y, and Z,” “at least one or more of X, Y, or Z,” “at least one or more of X, Y, and / or Z,” or “at least one of X, Y, and / or Z,” is intended to be inclusive of both a single item (e.g., just X, or just Y, or just Z) and multiple items (e.g., {X and Y}, {X and Z}, {Y and Z}, or {X, Y, and Z}). The phrase “at least one of” and similar phrases are not intended to convey a requirement that each possible item must be present, although each possible item may be present.

[0060] The techniques presented and claimed herein are referenced and applied to material objects and concrete examples of a practical nature that demonstrably improve the present technical field and, as such, are not abstract, intangible or purely theoretical. Further, if any claims appended to the end of this specification contain one or more elements designated as “means for [perform]ing [a function]...” or “step for [performing [a function]...”, it is intended that such elements are to be interpreted under 35 U.S.C. § 112(f). However, for any claims containing elements designated in any other manner, it is intended that such elements are not to be interpreted under 35 U.S.C. § 112(f).

[0061] While the invention has been described with reference to the exemplary examples thereof, those skilled in the art will be able to make various modifications to the described examples without departing from the true spirit and scope. The terms and descriptions used herein are set forth by way of illustration only and are not meant as limitations. In particular, although the method has been described by examples, the steps of the method can be performed in a different order than illustrated orsimultaneously. Those skilled in the art will recognize that these and other variations are possible within the spirit and scope as defined in the following claims and their equivalents.

Claims

What is claimed is:1 . A machine learning method of determining a pharmacokinetic parameter of a radiotracer in a patient, the method comprising: obtaining a time sequence of positron emission tomography (PET) scans of a target tissue of the patient, wherein the time sequence of PET scans comprises images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, from which features extracted from the time series of the PET scans are obtained, wherein the first neural network system comprises a self-supervised learning neural network; providing the features extracted from the time series of the PET scans to a second neural network system, from which provides an estimated arterial input function is obtained, wherein the second neural network system comprises a deep regression network, wherein the estimated arterial input function obtained from the second neural network is metabolite-corrected; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

2. The method of claim 1 , further comprising: adjusting a dosing of a therapeutic for the patient based at least in part on the estimate of the pharmacokinetic parameter of the radiotracer; and administering the therapeutic to the patient.

3. The method of claim 1 , wherein the pharmacokinetic parameter of the radiotracer comprises an amount of the radiotracer that is bound to a target receptor site.

4. The method of claim 1 , wherein the pharmacokinetic parameter of the radiotracer comprises an amount of the radiotracer in the target tissue.

5. The method of claim 1 , wherein the self-supervised learning neural setwork system comprises a variational autoencoder.

6. The method of claim 5, wherein the variational autoencoder comprises a vector-quantized variational autoencoder.

7. The method of claim 5, wherein the features extracted from the time series of the PET scans are obtained from bottleneck layers of the variational autoencoder.

8. The method of claim 5, wherein the features extracted from the time series of the PET scans comprise compact latent representations of the time series of PET scans.

9. The method of claim 1 , wherein the deep regression network comprises:a multi-layer perceptron layer applied to a spatial dimension to learn from features extracted from the time series of PET scans within a single point in time; and a multi-layer perceptron layer applied to a temporal dimension to learn relationships between the features extracted from the time series of PET scans across different points in time.

10. The method of claim 1 , wherein the estimated arterial input function is for a parent fraction.

11. A non-transitory computer-readable medium comprising instructions that, when executed by an electronic processor, configure the electronic processor to determine a pharmacokinetic parameter of a radiotracer in a patient by performing actions comprising: obtaining a time sequence of positron emission tomography (PET) scans of a target tissue of the patient, wherein the time sequence of PET scans comprises images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, from which features extracted from the time series of the PET scans are obtained, wherein the first neural network system comprises a self-supervised learning neural network; providing the features extracted from the time series of the PET scans to a second neural network system, from which provides an estimated arterial input function is obtained, wherein the second neural network system comprises a deepregression network, wherein the estimated arterial input function obtained from the second neural network is metabolite-corrected; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

12. The non-transitory computer-readable medium of claim 11 , wherein the actions further comprise: adjusting a dosing of a therapeutic for the patient based at least in part on the estimate of the pharmacokinetic parameter of the radiotracer; and administering the therapeutic to the patient.

13. The non-transitory computer-readable medium of claim 11 , wherein the pharmacokinetic parameter of the radiotracer comprises an amount of the radiotracer that is bound to a target receptor site.

14. The non-transitory computer-readable medium of claim 11 , wherein the pharmacokinetic parameter of the radiotracer comprises an amount of the radiotracer in the target tissue.

15. The non-transitory computer-readable medium of claim 11 , wherein the self-supervised learning neural setwork system comprises a variational autoencoder.

16. The non-transitory computer-readable medium of claim 15, wherein the variational autoencoder comprises a vector-quantized variational autoencoder.

17. The non-transitory computer-readable medium of claim 15, wherein the features extracted from the time series of the PET scans are obtained from bottleneck layers of the variational autoencoder.

18. The non-transitory computer-readable medium of claim 15, wherein the features extracted from the time series of the PET scans comprise compact latent representations of the time series of PET scans.

19. The non-transitory computer-readable medium of claim 11 , wherein the deep regression network comprises: a multi-layer perceptron layer applied to a spatial dimension to learn from features extracted from the time series of PET scans within a single point in time; and a multi-layer perceptron layer applied to a temporal dimension to learn relationships between the features extracted from the time series of PET scans across different points in time.

20. A system for determining a pharmacokinetic parameter of a radiotracer in a patient, the system comprising an electronic processor and a non-transitory computer-readable medium comprising instructions that, when executed by the electronic processor, configure the electronic processor to determining apharmacokinetic parameter of a radiotracer in a patient by performing actions comprising: obtaining a time sequence of positron emission tomography (PET) scans of a target tissue of the patient, wherein the time sequence of PET scans comprises images of a radiotracer in the target tissue; providing the time series of the PET scans to a first neural network system, from which features extracted from the time series of the PET scans are obtained, wherein the first neural network system comprises a self-supervised learning neural network; providing the features extracted from the time series of the PET scans to a second neural network system, from which provides an estimated arterial input function is obtained, wherein the second neural network system comprises a deep regression network, wherein the estimated arterial input function obtained from the second neural network is metabolite-corrected; modeling the pharmacokinetic parameter of the radiotracer based on the estimated arterial input function and the time sequence of PET scans; and outputting an estimate of the pharmacokinetic parameter of the radiotracer based on the modeling.

Images