Fast volumetric and perfusion cone beam computed tomography imaging for radiotherapy using diffusion models

Abstract

Presented herein are systems and methods of reconstructing tomograms from projection datasets obtained during radiotherapy. A computing system may receive a projection dataset corresponding to a portion of an arc about a first structure of interest (SOI) in a first volume of a subject acquired in accordance with a cone beam computed tomography (CBCT). The computing system may generate, based on applying the projection dataset to a machine learning (ML) model, a tomogram (e.g., a perfusion image)corresponding to the first volume of the subject. The computing system may store, using one or more data structures, an association between the subject and the tomogram.

Description

FAST VOLUMETRIC AND PERFUSION CONE BEAM COMPUTED TOMOGRAPHY IMAGING FOR RADIOTHERAPY USING DIFFUSIONMODELSCROSS REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims the benefit of priority to U.S. Provisional Patent Application No. 63 / 673,508, filed July 19, 2024, which is incorporated herein by reference in its entirety.BACKGROUND

[0002] A computer system may apply a machine learning model on an input dataset to generate an output.SUMMARY

[0003] Aspects of the present disclosure are directed to systems and methods of reconstructing tomograms from projection datasets. One or more processors may receive a projection dataset corresponding to a portion of an arc about a first structure of interest (SOI) in a first volume of a subject. The one or more processors may provide the projection dataset as input to a machine learning (ML) model. The ML model may be established using a plurality of examples. Each of the plurality of examples may include a respective tomogram of a second SOI in a second volume of the subject. The one or more processors may generate, based on providing the projection dataset as input to the ML model, a tomogram of the first SOI in the first volume of the subject. The one or more processors may store, using one or more data structures, an association between the subject and the tomogram.

[0004] In some embodiments, the one or more processors may acquire, via a tomograph, the projection dataset, at least partially concurrent to administration of radiotherapy to the first SOI of the subject. In some embodiments, the one or more processors may generate, using userinput, an output based on the first tomogram and the administration of the radiotherapy. In some embodiments, the one or more processors may generate the output including at least one of: (i) a treatment assurance indicating one of failure or success at administering the radiotherapy, (ii) a dosimetry analysis indicating a dosage of the radiotherapy administered or (iii) a treatment adjustment to identify modifications to the administration of the radiotherapy.

[0005] In some embodiments, the one or more processors may provide, for presentation via a user interface, the output based on the tomogram and the administration of the radiotherapy. The administration of the radiotherapy to the first SOI is modified, subsequent to provision of the output. In some embodiments, the subject may be administered with radiotherapy at least partially concurrent with acquisition of the projection dataset. The radiotherapy may include at least one of an intensity-modulated radiation therapy (IMRT), stereotactic body radiation therapy (SBRT), image-guided radiation therapy (IGRT), or brachytherapy.

[0006] In some embodiments, the one or more processors may iteratively apply the ML model. The ML model may include an optimization function to generate an output tomogram using the projection dataset and a reference tomogram from at least one of the plurality of examples. The ML model may include a diffusion model to generate an intermediate tomogram using the output tomogram from the optimization function. In some embodiments, the optimization function may include a sparsifier function to generate a sparse representation of the projection dataset and the reference tomogram.

[0007] In some embodiments, the one or more processors may determine that a number of iterations of applying the ML model is greater than or equal to a maximum number. The one or more processors may identify the tomogram as output for the ML model, responsive to determining that the number of iterations is greater than or equal to the maximum number. In some embodiments, the one or more processors may receive the projection dataset corresponding to the portion of the arc about the first SOI acquired via perfusion cone beam computed tomography (CBCT) imaging. The first SOI may include a blood vessel in the first volume of the subject, theblood vessel injected with a contrast agent. The one or more processors may generate the tomogram including a perfusion map identifying at least one characteristic associated with the blood vessel in the first volume of the subject. The at least one characteristic may include at least one of a blood flow rate, a blood volume, a mean transmit time (MTT), or a time-to-peak (TTP).|0008| In some embodiments, the subject may be at risk of or diagnosed with cancer. The cancer may include at least one of: brain cancer, breast cancer, head and neck cancer, lung cancer, prostate cancer, colorectal cancer, skin cancer, bladder cancer, pancreatic cancer, liver cancer, renal cancer, ovarian cancer, or cervical cancer. In some embodiments, the projection dataset is acquired in accordance with cone beam computed tomography (CBCT). In some embodiments, the first SOI may correspond to at least one organ associated with the cancer in the subject administered with the radiotherapy. The portion may correspond to between 15-90 degrees of the first arc about the subject.(0009] Aspects of the present disclosure are directed to systems and methods of training models to reconstruct tomograms from projection datasets. One or more processors may identify training data identifying a first tomogram of a structure of interest (SOI) in a volume of a subject. The one or more processors may add noise to the first tomogram to generate a second tomogram. The one or more processors may provide the second tomogram as input to a machine learning (ML) model comprising a plurality of parameters to generate a third tomogram of the SOI in the volume of the subject. The one or more processors may determine a loss metric based on the first tomogram from the training data and the third tomogram from the ML model. The one or more processors may update at least one of the plurality of parameters of the ML model using the loss metric. The one or more processors may store the plurality of parameters of the ML model to process a projection dataset corresponding to a portion of an arc about the SOI in the volume of the subject.]0010| In some embodiments, the one or more processors may sample at least a portion of the first tomogram for generation of the second tomogram. In some embodiments, the one or more processors may sample at least a portion of an output in a first iteration of applying the MLmodel to applying in a second iteration of applying the ML model. In some embodiments, the one or more processors may iteratively apply the ML model. The ML model may include a diffusion model to generate an intermediate tomogram using the second tomogram. In some embodiments, an optimization function to be used with the diffusion model may include a sparsifier function to generate a sparse representation of the reference tomogram. In some embodiments, the one or more processors may determine the loss metric to indicate a degree of deviation of the third tomogram relative to the first tomogram.

[0011] In some embodiments, the one or more processors may determine that a number of iterations of applying the ML model is greater than or equal to a maximum number. The one or more processors may identify the tomogram as output for the ML model, responsive to determining that the number of iterations is greater than or equal to the maximum number. In some embodiments, the one or more processors may determine that a number of iterations of applying the ML model is less than a maximum number. The one or more processors may continue applying the tomogram to the ML model, responsive to determining that the number of iterations is less than the maximum number.

[0012] In some embodiments, the tomogram includes a perfusion map identifying at least one characteristic associated with a blood vessel in the volume of the subject. The at least one characteristic may include at least one of a blood flow rate, a blood volume, a mean transmit time (MTT), or time-to-peak (TTP). In some embodiments, the first tomogram may be acquired in accordance with cone beam computed tomography (CBCT). The portion for the projection dataset may correspond to between 15-90 degrees of the first arc about the subject.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The foregoing and other objects, aspects, features, and advantages of the disclosure will become more apparent and better understood by referring to the following description taken in conjunction with the accompanying drawings, in which:(0014) FIG. 1 : The proposed diffusion sampling with CG-guided initialization and denoising. 15 diffusion steps were utilized during the diffusion sampling for image reconstruction.10015] FIG. 2: First to third columns: FDK, CGLS and the reconstruction. Last column: FDK reconstruction with all 897 projections.10016] FIG. 3 : The reconstructed images were fused with the ground truth (GT) to demonstrate the hallucination artifacts introduced during image reconstruction. CGLS reconstructions were provided for the comparison between diffusion-based method and traditional method.

[0017] FIG. 4: kV X-ray projections was proposed to be acquired during radiation treatment using the LINAC on-board imaging system and then reconstructed into real-time or near real-time 3D CBCT volumetric images using a diffusion model-based reconstruction algorithm. These volumetric images can be used to validate or guide radiation treatment and for radiotherapy data analysis.(0018] FIG. 5: For diffusion model (DM)-based reconstruction, the following algorithm was proposed utilizing any pre-trained denoising diffusion model of N steps. The iterative optimizer minimizes the cost function: + (1 —m) || (x) ||i, where(. ) can be any sparsifying function, (3 and a> are empirical weighting factors

[0019] FIG. 6 depicts a block diagram of a system of reconstructing tomograms from projection datasets in accordance with an illustrative embodiment.10020] FIG. 7 depicts a block diagram of a process for training models to reconstruct tomograms from projection datasets in accordance with an illustrative embodiment.[00211 FIG. 8 depicts a block diagram of a process for applying models to reconstruct tomograms from projection datasets in accordance with an illustrative embodiment.

[0022] FIG. 9 depicts a block diagram of a process for generating output based on tomograms reconstructed from projection datasets in accordance with an illustrative embodiment.|0O23| FIG. 10 depicts a flow diagram of a method of reconstructing tomograms from projection datasets in accordance with an illustrative embodiment.

[0024] FIG. 11 depicts a flow diagram of a method of training models to reconstruct tomograms from projection datasets in accordance with an illustrative embodiment.

[0025] FIG. 12 depicts a block diagram of a server system and a client computer system, in accordance with one or more implementations.DETAILED DESCRIPTION

[0026] Following below are more detailed descriptions of various concepts related to, and embodiments of, systems and methods for reconstructing tomograms from projection datasets. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in any of numerous ways, as the disclosed concepts are not limited to any particular manner of implementation. Examples of specific implementations and applications are provided primarily for illustrative purposes.

[0027] Section A describes diffusion-based cone-beam computed tomography (CBCT) image reconstruction in radiotherapy.

[0028] Section B describes accelerated conjugate gradient-guided diffusion sampling for cone-beam computed tomography (CBCT) image reconstruction using patient-specific denoising diffusion probabilistic model.

[0029] Section C describes ultrafast short-arc diffusion-based cone beam computed tomography (CT) image reconstruction.

[0030] Section D describes a system and method for reconstructing tomograms from projection datasets using machine learning.

[0031] Section E describes a network environment and computing environment, which may be useful for practicing various computing related embodiments described herein.A. Diffusion-Based CBCT Image Reconstruction in Radiotherapy|0032| Purpose: Diffusion-based image prior offers promise for CT reconstruction, but the application is hindered by slow image reconstruction and hallucination artifacts. This study aims to investigate the potential of diffusion-based CBCT image reconstruction in a radiotherapy clinical workflow.

[0033] Methods: A clinical patient who was enrolled underwent four-dimensional CT, free-breathing CT, and 3 breath-hold CT scans during radiotherapy simulation. All those axial CT images were used with random shifts to train a patient-specific Denoising Diffusion Probabilistic Model (DDPM). Training images were downsampled to 256x256 for computational efficiency. Model training took 2 days on dual Nvidia Quadro RTX 8000 GPUs. Utilizing the trained model as prior knowledge, the model was applied in sparse-view CBCT reconstructions with 20 and 90 equiangular projections on treatment day. Employing a conjugate-gradient (CG)-guided diffusing sampling scheme, the diffusion-based CBCT was initiated in the reconstruction process from a middle step using noise-blended CG reconstructed images. The patient-specific model aims to minimize hallucination, while the CG-guided backward diffusion sampling scheme expedites the reconstruction process. The reconstructed image quality was assessed using peak signal-to-noise ratio (PSNR) and Structure Similarity (SSIM) metrics, comparing against the conventional Feldkamp-Davis-Kress (FDK) and Conjugate Gradient least squares (CGLS) methods.

[0034] Results: The approach showed significant enhancements across all metrics as compared to the traditional FDK method and CGLS method. For 20-proj ection CBCT reconstruction, PSNR rose from 15.7 to 19.5 and SSIM from 0.55 to 0.66. Similarly, for 90- projection CBCT reconstruction, PSNR increased from 19.9 to 23.8 and SSIM from 0.73 to 0.77. Leveraging 15 diffusion steps, the method accomplished CBCT reconstruction (matrix size: 256x256x87, voxel size: 2x2x2 mm) within 5 minutes.

[0035] Conclusion: The feasibility of diffusion-based CBCT image reconstruction in clinical settings was demonstrated. Employing patient-specific diffusion models alongside tailored sampling schemes was shown to effectively reduces hallucinations and the number of diffusion steps required.

[0036] Diffusion models have emerged as a promising tool in medical imaging for tackling inverse problems. Typically, these models are trained on population data to capture anatomical intricacies, serving as valuable priors for sparse-view or limited-angle CT reconstruction in new patients. However, individual patients possess unique lesion characteristics not fully represented by such models, leading to potential hallucination artifacts — a significant challenge in utilizing image priors for reconstruction. Additionally, diffusion-based reconstruction often demands numerous diffusion steps for competitive image quality, posing a time constraint in busy clinical settings.

[0037] Thus, despite the potential, diffusion-based image reconstruction remains relatively unexplored in clinical practice. This study aims to assess the practicality in CBCT reconstruction under sparse-view imaging geometry, using real clinic-CBCT projection images. The objectives include achieving reconstruction within minutes and ensuring data fidelity.

[0038] To accomplish this, a novel approach: training a patient-specific Denoising Diffusion Probabilistic Model (DDPM) and integrating the model with a conjugate gradient (CG)-guided sampling mechanism was proposed. This combined method not only aims toenhance data consistency but also expedite CBCT image reconstruction, paving the way for the clinical implementation.

[0039] The training was pioneered of patient-specific image priors using simulation CT to enhance data consistency in CBCT image reconstruction during treatment. A novel CG- guided sampling method was utilized, enhancing data consistency and expediting reconstruction. The study represents one of the pioneering endeavors exploring diffusion-based image reconstruction using authentic clinical projection data. An iterative image was implemented normalization technique to enhance numerical compatibility between the diffusion model and real projection data, optimizing reconstruction accuracy.

[0040] An improved DDPM was trained using a patient’s axial images from 4DCT, free- breathing (FB) CT and three breath-hold CT scans acquired during the simulation. During training, the images were augmented by image shifting in the anterior-posterior and lateral directions. To expedite the model training process, the images were downsampled to 256x256 resolution. Training the model required two days on two Nvidia Quadro RTX 8000 GPUs with 48GB of memory each. A linear noise schedule was employed with 1000 diffusion steps in the DDPM, and for the deep learning model, two ResNet blocks and one attention head was utilized with a resolution of 16. Notably, the diffusion noise schedule was pre-determined and not learned by the model.

[0041] On treatment day, the patient was scanned with half-fan geometry with a total of 897 projections acquired across a 360-degree circular trajectory. For sparse view CBCT image reconstruction testing using the proposed method, 20 and 90 equiangular projections were randomly selected from the 897 projections. Leveraging the denoising diffusion implicit model sampling, a CG-guided diffusion sampling mechanism shown in FIG. 1 was proposed. Strided diffusion steps of 30 was utilized instead of the conventional 1000 steps. CG iterative reconstruction was selected for the rapid convergence, which complements the strided denoising diffusion steps effectively. To reduce hallucination and expedite the sampling process, theproposed sampling scheme was initialized from step 15 with noise-blended reconstruction images.

[0042] FDK for CBCT reconstruction was utilized using all available projections, establishing the resulting images as ground truth for the evaluation. Assessing CBCT reconstruction quality, SSIM and PSNR metrics were employed. Results are shown in FIG. 2 reconstructions with 20 and 90 projections.|0043| Results: Table 1 offers a comprehensive evaluation of image quality metrics, including PSNR and SSIM specifically calculated within the body mask. Notably, the results remain unaltered by any post-processing noise-removal techniques. The remarkable performance and reduced noise levels of the method can be attributed to the diffusion model’s training using CT images, rather than CBCT images. Furthermore, the proposed sampling scheme, employing N=15, surpasses both FDK and CGLS iterative reconstructions with 20 iterations across all metrics, underscoring efficacy and superiority.

[0044] Table 1. PSNR and SSIM comparison between the CGLS, FDK, and the method.B. Accelerated Conjugate Gradient-Guided Diffusion Sampling for CBCT Image Reconstruction Using Patient-specific Denoising Diffusion Probabilistic Model

[0045] Purpose: A novel approach that leverages conjugate gradient (CG) guidance was proposed within conditional diffusion sampling to expedite diffusion-based Cone Beam CT (CBCT) image reconstruction. The goals are to enhance data consistency, mitigate hallucination artifacts, and accelerate the reconstruction process.

[0046] Methods: A patient-specific Denoising Diffusion Probabilistic Model (DDPM) was trained to be incorporated for CBCT image reconstruction on treatment days. The diffusion model was trained using 2D axial slices from CT simulations, augmented by random shifting in anterior-posterior and lateral directions, and resized to 256x256 to expedite training. Employing the trained model as the patient’s prior knowledge, sparse-view and limited-angle CBCT image reconstructions were utilized. The diffusion model, integrated with CG optimization, iteratively reconstructed CBCT images. To mitigate hallucination, the diffusion sampling process was hijacked from a middle step, using noise-blended CG reconstructed images, rather than from the beginning step with a pure Gaussian noise image. Evaluations were conducted for both limitedangle and sparse-view CBCT reconstructions using PSNR, SSIM, and MAE. Compared to the sampling scheme initiated from pure Gaussian noise, the proposed sampling scheme demonstrated significant improvements in PSNR, SSIM, and MAE with only 5 diffusion steps.

[0047] Results: For 90 equidistant sparse-view CBCT reconstruction, PSNR improved from 25.7 to 34.6, SSIM from 0.87 to 0.96, and MAE decreased from 20.7 to 7.2. For limitedangle CBCT reconstruction with 90 projections, PSNR improved from 22.8 to 26.6, SSIM from 0.83 to 0.89, and MAE decreased from 28.3 to 18.3. The results indicate a substantial reduction in the required reconstruction steps, significantly expediting the process. With the proposed method, the CBCT reconstruction of image size 256x256x110 with voxel size of 2x2x1.25 mm took approximately 3 minutes using two Nvidia Quadro Rtx 8000 GPUs with 48GB of memory.[0048| Conclusion: The proposed CG-guided diffusion sampling mechanism offers a promising solution for diffusion-based CBCT image reconstruction, effectively promoting data consistency, suppressing hallucination artifacts, and streamlining the reconstruction process.

[0049] Denoising diffusion probabilistic models (DDPMs) have emerged as powerful tools for accurately modeling underlying data distributions by maximizing the likelihood of training image samples. Diffusion models’ performance has often surpassed that of Generative Adversarial Networks (GANs) in various image-generation tasks. Given the ability to effectively model data distributions, diffusion models are increasingly applied to inverseproblems like image inpainting, super-resolution, and denoising. Notably, one key advantage of diffusion models lies in the forward process agnosticism, allowing a single model to address multiple inverse problems that share the same underlying data distribution.

[0050] Diffusion models have gained traction in medical imaging to address inverse problems. Typically, a population-based diffusion model is trained to capture the intricacies of human anatomy, subsequently serving as an image prior for sparse-view or limited-angle CT reconstruction in new patients. Each new patient is a unique individual characterized by distinct pathological characteristics, which the diffusion model may not fully represent. Consequently, a notable challenge in utilizing image priors for reconstruction arises from the potential introduction of hallucination artifacts.

[0051] As generative models, DDPMs are not immune to introducing hallucinations when utilized for image reconstruction. Hence, there is a pressing need to devise novel methods that enhance measured data consistency, while effectively suppressing hallucination artifacts induced by the diffusion model. To train a patient-specific DDPM and integrate the DDPM with image reconstruction by employing a novel conjugate gradient (CG)-guided sampling mechanism was proposed. This combined method aims to improve data consistency, while simultaneously addressing hallucination in Cone Beam Computed Tomography (CBCT) image reconstruction.

[0052] The concept of training a patient-specific image prior was introduced using patient’s simulation CT, tailored for CBCT image reconstruction of the same patient on treatment days. A novel conjugate gradient-guided sampling mechanism was devised to improve data consistency, while minimizing hallucinations. An accelerated backward sampling scheme was developed with 5 steps, although the DDPM used 1000 forward steps. The proposed method offers versatility by enabling sparse-view and limited angle-based CBCT image reconstruction without the need for specifying the CBCT imaging geometry in advance.

[0053] An enhanced version of the DDPM, provided by OpenAI, is now publicly available on GitHub. A patient-specific DDPM model was trained using the patient’s 4DCT axial images, augmented by image shifting in the anterior-posterior and lateral directions. To expedite the model training process, the images were downsampled to 256x256 resolution. Training the model required two days on two Nvidia Quadro Rtx 8000 GPUs with 48GB of memory each. A linear noise schedule was employed with 1000 diffusion steps in the DDPM, and for the deep learning model, two ResNet blocks and one attention head with a resolution of 16 were utilized. Notably, the diffusion noise schedule was pre-determined and not learned by the model.

[0054] The trained DDPM model served as the image prior for reconstructing FB CT images simulated under CBCT geometry. The CBCT inverse problem was formulated as an ill- posed Ax + e — y problem, with x representing the CT image to be reconstructed and y representing the measured data. Leveraging the denoising diffusion implicit model (DDIM) sampling, two CBCT image reconstruction sampling mechanisms: sampling coupled with CG iterative reconstruction (Algorithm 1), and CG-guided sampling coupled with CG iterative reconstruction (Algorithm 2) were proposed. To expedite the reconstruction process, strided diffusion steps of 20 were utilized, as opposed to the conventional 1000 steps. CG iterative reconstruction was selected for the rapid convergence, which complements the strided denoising diffusion steps effectively.|0055| For CBCT reconstruction simulations, two geometries: equidistant sparse-view with 20 and 90 projections and limited angle with 45 and 90 projections with 1 -degree intervals were experimented. GPU-enabled CG iterative reconstruction using the TIGRE image reconstruction packages was implemented. Forward CBCT projection images were simulated using the FB CT under the studied CBCT geometries. Employing the two denoising diffusion sampling schemes coupled with CG iterative reconstruction (Algorithm 1 and Algorithm 2), CBCT images were successfully reconstructed. The reconstruction steps for both algorithms are illustrated in the following two tables. To evaluate CBCT image reconstruction quality, SSIM and PSNR metrics were calculated. Additionally, to quantitatively assess the amount ofhallucination introduced by the diffusion image prior, Mean Absolute Error (MAE) was computed.

[0056] Results: Table 2 shows the quantitative evaluation of image age quality using PSNR, SSIM and MAE. Algorithm 2 with N=5 demonstrated superior performance across all evaluation metrics compared to Algorithm 1, irrespective of the parameter settings (N=5 or N=20). These findings underscore the effectiveness of Algorithm 2 in achieving enhanced image quality. With the proposed method, the CBCT reconstruction of image size 256x256x110 with voxel size of 2x2x1 ,25mm took approximately 3 mins using two GPU. To demonstrate the superiority of diffusion based methods, the traditional CGLS iterative reconstruction results were also provided.[00571 FIG. 3 complements the quantitative evaluation by visually depicting the degree of “hallucination” introduced during the image reconstruction process. This term refers to discrepancies or artifacts present in the reconstructed images compared to the ground truth counterparts. The differences between the reconstructed and ground truth images were computed and presented for each of the three sampling schemes examined in Table 2.

[0058] Table 2. Simulation of limited angle and sparse-view CBCT image reconstruction using Algorithm 1 and Algorithm 2C. Ultrafast Short-Arc Diffusion-based Cone Beam CT Image Reconstruction

[0059] Purpose: For breath-hold Cone Beam CT (CBCT) imaging on a conventional linear accelerator (LINAC), multiple breath holds are typically required to acquire complete projections for faithful image reconstruction. The long acquisition not only causes motion artifacts on the CBCT images but also reduces patient comfort during the treatment. This study aims to introduce a new diffusion-based image reconstruction method for ultra-fast short-arc CBCT acquisition.

[0060] Methods: To facilitate short-arc image reconstruction, a patient-specific Denoising Diffusion Probabilistic Model (DDPM) model was trained using the patient’s 4DCT axial images, augmented by image shifting in the anterior-posterior and lateral directions. To expedite the model training process, the images were downsampled to 256x256 resolution.Training the model required two days on two Nvidia Quadro RTX 8000 GPUs with 48GB of memory each. A linear noise schedule with 1,000 diffusion steps in the DDPM was employed. Utilizing the trained model as prior knowledge, short-arc CBCT reconstructions with 90-degree and 45-degree limited angle data acquisition was applied. A conjugate-gradient (CG)-guided diffusing sampling scheme was proposed to initiate the diffusion-based CBCT reconstruction process from a middle step using noise-blended CG reconstructed images. The patient-specific model aims to minimize hallucination, while the CG-guided backward diffusion sampling scheme expedites the reconstruction process. For validation, a simulation study to reconstruct the patient’s free breathing CT under cone beam geometry was performed, aided by a patientspecific diffusion model trained by the patient’s 4DCT images. The reconstructed image quality using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics was assessed, comparing against the conventional iterative CG least squares (CGLS) methods.|0061[ Results: The CBCT acquisition takes only 15 seconds for the 90-degree short-arc and 7.5 seconds for the 45-degree short-arc on a conventional LINAC, which could be completed within one breath-hold. Compared to the traditional iterative CGLS reconstruction, PSNR improved from 22.1 to 26.6, SSIM from 0.87 to 0.89, and MAE decreased from 35.0 to 18.3 for 90-degree reconstruction and PSNR improved from 18.2 to 22.7, SSIM from 0.83 to 0.85, and MAE decreased from 53.5 to 31.1 for 45-degree reconstruction. With the proposed method, the CBCT reconstruction of image size 256x256x110 with a voxel size of 2x2x1.25 mm took approximately 3 minutes using two Nvidia Quadro RTX 8000 GPUs with 48GB of memory.

[0062] Conclusion: Leveraging patient-specific prior knowledge through diffusion models enables ultra-fast short-arc CBCT image acquisition, enhancing image quality, reducing motion artifacts, and increasing patient comfort during treatment.

[0063] Referring now to FIG. 4, depicted is a block diagram of a system for reconstructing volumetric images from projection data. kV X-ray projections was proposed to be acquired during radiation treatment using the LINAC on-board imaging system and then reconstructed into real-time or near real-time 3D CBCT volumetric images using a diffusionmodel-based reconstruction algorithm. These volumetric images can be used to validate or guide radiation treatment and for radiotherapy data analysis.

[0064] Referring now to FIG. 5, depicted is a block diagram of a system for diffusionmodel based image reconstruction. For diffusion model (DM)-based reconstruction, the following algorithm was proposed utilizing any pre-trained denoising diffusion model of N steps. The iterative optimizer minimizes the cost function: / (x) = — / 3\\y — / lx||2 + m||¥ / 1(x — xt_ i) ||i + (1 — m) || *F2(x) ||i, where(I (. ), ¥ / 2(. ) can be any sparsifying function, / ? and a> are empirical weighting factors.D. Systems and Methods for Reconstructing Volumetric Image from Short-Arc Projection Dataset

[0065] Referring now to FIG. 6, depicted is a block diagram of a system 100 of reconstructing volumetric images from projection datasets. In overview, the system 100 may include at least one image processing system 105, at least one imaging device 110, and at least one administrative device 115, communicatively coupled with one another via at least one network 120. The image processing system 105 may include at least one projection handler 125, at least one model trainer 130, at least one model applier 135, at least one output evaluator 140, at least one image reconstruction model 145, and at least one database 150, among others. The image reconstruction model 145 may include at least one noise injector 155, at least one diffusion model 160, and at least one optimization function 165 (sometimes herein referred to as an optimizer), among others. Each of the components in the system 100 as detailed herein may be implemented using hardware (e.g., one or more processors coupled with memory), or a combination of hardware and software as detailed herein in Section E.

[0066] In further detail, the image processing system 105 may be any computing device including one or more processors coupled with memory and software and capable of performing the various processes and tasks described herein. The image processing system 105 can be in communication with the imaging device 110, the administrative device 115, the database 150,and other devices, via the network 120. The image processing system 105 may be situated, located, or otherwise associated with at least one server group. The server group may correspond to a data center, a branch office, or a site at which one or more servers corresponding to the image processing system 105 is situated.100671 The image processing system 105 may include one or more modules, components, or subsystems to perform the various processes and tasks described herein. The projection handler 125 may communicate with the imaging device 110 to receive projection data from a scan of a volume in a subject. The model trainer 130 may train and establish the image reconstruction model 145 using a trained diffusion model 160 to reconstruct tomograms from projection data. The model applier 135 may apply acquired projection data to the image reconstruction model 145 to reconstruct the tomograms. The output evaluator 140 may use the reconstructed tomograms to generate information for presentation. The image processing system 105 can perform or implement any of the functionalities detailed herein in conjunction with Sections A-C.|0068| The image reconstruction model 145 (sometimes herein referred to as image reconstructor or machine learning model) may include an integration of any type of generative machine learning (ML) architecture or artificial intelligence (Al) algorithm with any type of objective optimization algorithm to reconstruct tomograms from projection data (sometimes herein referred to as projections, sinograms, or raw projection data). The image reconstruction model 145 can include, for example, a deep learning artificial neural network (ANN) (e.g., a U- net architecture with encoders and decoders, residual networks, or transformer-based architecture), a Markov chain, a support vector machine (SVM), a clustering algorithm, a Bayesian classifier, or a decision tree, among others. In some embodiments, the image reconstruction model 145 may be structured as a denoising diffusion probabilistic model (DDPM). In general ,the image reconstruction model 145 may include at least one input (e.g., the projection dataset) and at least one output (e.g., the tomogram). The image reconstruction model 145 may include a set of parameters (or weights) relating the input with the output.(0069) Within the image reconstruction model 145, the noise injector 155 may add noise (e.g., Gaussian noise) to an input (e.g., tomogram or volumetric image). The diffusion model 160 may be implemented with any ML architecture (e.g., Markov chain, U-net, or residual network), and may denoise the input from the noise injector 155 to generate the denoised image. The optimization function 165 may be implemented using an image reconstruction algorithm, such as conjugate gradient (CG) iterative reconstruction, filtered back projection (FBP), algebraic reconstruction technique (ART), total variation (TV) minimization, maximum likelihood expectation maximization (MLEM), iterative close point (ICP) algorithm, or radon transform, among others. The optimization function 165 may apply the algorithm to the input projection dataset or the ML-generated tomogram to generate an updated tomogram, which maximize the consistency of the output with the input projection and with the ML-generated image. The image reconstruction model 145 may be iteratively applied any number of times to generate the tomogram. In some embodiments, the image reconstruction model 145 can be maintained on the image processing system 105. In some embodiments, the image reconstruction model 145 can be maintained on a computing device or server separate from the imaging processing system 105.

[0070] The imaging device 110 (sometimes herein generally referred to as an imaging device, an image acquirer, or a tomograph) may be any device to acquire projection data of a volume of a subject. The projection data may be acquired in accordance with a tomographic imaging technique, such as a magnetic resonance imaging (MRI), a nuclear magnetic resonance (NMR), high-energy electromagnetic radiation (X-ray), computed tomography (CT) (e.g., cone beam computed tomography (CBCT)), an ultrasound imaging, and a positron emission tomography (PET), and a photoacoustic spectroscopy, among others. Although primarily discussed herein in terms of an X-ray and CBCT, other imaging modalities besides those listed above may be supported by the image processing system 105.[00711 The administrative device 115 may (sometimes herein referred to as an operator device or a clinician device) can be any computing device comprising one or more processors coupled with memory and software and capable of providing an output projection image. Theadmini strati ve device 115 can be associated with an entity (e.g., a clinician) examining the subject or tomograms from the subject. The administrative device 115 can be in communication with the image processing system 105 and the imaging device 110 to exchange data. The administrative device 115 can display tomograms acquired from the imaging device 110. The administrative device 115 can present information associated with the subject or the tomograms from the subject.

[0072] Referring now to FIG. 7, depicted is a block diagram of a process 200 for training a generative machine learning model to be used in reconstructing tomograms from projection datasets. The process 200 may include or correspond to operations performed in the system 100 to establish the image reconstruction model 145. Under the process 200, the model trainer 130 executing on the image processing system 105 may retrieve, obtain, or otherwise identify training data 205. In some embodiments, the model trainer 130 may access the database 150 to retrieve the training data 205. The training data 205 may be used to train the image reconstruction model 145. The training data 205 may identify or include a set of examples.Each example of the training data 205 may identify or include at least one tomogram 215A for at least one subject 220. Using the training data 205, the model trainer 130 may initialize, train, and establish the diffusion model 160. When initialized, the set of parameters of the image reconstruction model 145 (or a portion of the image reconstruction model 145, such as the diffusion model 160) may be assigned or set to defined values (e.g., random values).[0073| The tomogram 215A (sometimes herein referred to as a biomedical image or a projection image) may be a scan of at least one volume 225, including at least one structure of interest (SOI) 230 of the subject 220. The tomogram 215A may be derived from or generated using a projection dataset acquired in accordance with an imaging modality (e.g., CT or CBCT imaging). The tomogram 215A may correspond to at least a segment of an arc (e.g., a quarter, half, or full circle) about the volume 225, including the SOI 230 in the subject 220. For example, the tomogram 215A may correspond to at least a scan acquired from a multiple angles along a half circle about the volume 225 of the subject 220. The tomogram 215A may have a set of pixels (or pixels) in a resolution in two dimensions (e.g., in x and y axes) or three dimensions(e.g., in x, y, and z axes). The resolution may range from 512 by 512 pixels to 1024 by 1024 pixels per two-dimensional slice. In some embodiments, the tomogram 215A may include a set of frames forming a video scan. Each frame can correspond to a respective sample time of a sequence of time samples. For instance, each frame can be a tomogram and can be derived from or generated using time-series projection data.

[0074] The tomogram 215A may be acquired in accordance with an imaging modality (e.g., x-ray, SPECT, or CBCT imaging). For example, the tomogram 215A may be obtained, derived, or otherwise acquired of the volume 225 containing the SOI 230 as a beam of energy (e g., x-ray) passing through the volume 225 of the subject 220. The beam of energy may be attenuated by the objects within the volume 225 of the subject 220, such as tissues, bone, fat, and organs, among others. In some embodiments, the tomogram 215A may be acquired or derived in accordance with perfusion imaging (e.g., CBCT perfusion imaging). The tomogram 215A may correspond to the volume 225, including the SOI 230 corresponding to at least one blood vessel (e.g., within the brain or another organ). To prepare for perfusion imaging, the blood vessel of the subject 220 may be injected with a contrast agent (e.g., an iodinated contrast agent, such as iohexel, iopamidol, iversol, or iodixanol). The contrast agent may be to increase the visibility or contrast of the blood vessels within the resultant imaging. With the injection, the imaging device 110 may perform imaging (e.g., CBCT or CT imaging) of the volume 225 to generate the tomogram 215A.

[0075] The tomogram 215 A may identify or include a perfusion map (sometimes herein referred to as a perfusion image). The perfusion map may define or identify at least one characteristic associated with the blood vessel in the volume 225 of the subject 220. The characteristic may include, for example, one or more of: a blood flow rate (e.g., volume of blood passing through the organ or the volume in a given time), a blood volume (e.g., amount of blood present in the organ or the volume), a mean transmit time (MTT) (e.g., average time for blood to pass through the organ or volume), or a time-to-peak (TTP), among others. For instance, the perfusion map can include a set of pixels (or voxels). Each pixel can define or include a value(e.g., an intensity value or a red-green-blue (RGB) value) corresponding to the degree of the characteristic.

[0076] The subject 220 may be a human or animal subject, among others. The subject 220 may have, may be at risk of, or may be diagnosed with at least one condition. The condition may include, for example, brain cancer, breast cancer, head and neck cancer, lung cancer, prostate cancer, colorectal cancer, skin cancer, bladder cancer, pancreatic cancer, liver cancer, renal cancer, ovarian cancer, cervical cancer, and the like. The subject 220 may have been or may be concurrently administered with a radiotherapy to treat the cancer. The radiotherapy can include, for example, intensity-modulated radiation therapy (IMRT), stereotactic body radiation therapy (SBRT), image-guided radiation therapy (IGRT), or brachytherapy, among others. The volume 225 may correspond to a section, a portion, a region, or an area of the subject 220. The volume 225 may include, for example, a head, a neck, a torso, a back, a pelvis, arms, legs, or any organ of the subject 220. The volume 225 can include the SOI 230 within the subject 220. In some embodiments, the volume 225 may correspond to a region on the subject 220 in which cancer is present. The SOI 230 may be at least one organ to be administered for radiotherapy for cancer within the corresponding volume 225 of the subject 220. The organ can be any organ within the subject 220, such as, brain, breast, lung, liver, bone, lymph node, prostate, bladder, colon, rectum, prostate, pancreas, kidney, ovaries, uterus, cervix gastrointestinal tract, pelvis, spine, or a tissue, among others. For example, if the volume 225 is a two-dimensional slice or three-dimensional region about the head of the subject 220, the SOI 230 can be the brain or a specific portion of the brain. In some embodiments, the SOI 230 may correspond to or may include at least one blood vessel in the organ (e.g., blood vessels in the brain).

[0077] The model applier 135 executing on the image processing system 105 may input, feed, or otherwise apply the tomogram 215A to the noise injector 155 of the image reconstruction model 145. In some embodiments, the model applier 135 may sample at least a portion of the tomogram 215A to input the ML model 145 for generating additional tomograms. For instance, the model applier 135 may down-sample the tomogram 215A to reduce the resolution from 1024 by 1024 pixels to 256 by 256 pixels to use as input to the imagereconstruction model 145. In processing, the model applier 135 may input, feed, or apply the tomogram 215A to the noise injector 155. Upon applying, the noise injector 155 may inject, include, or otherwise add the noise to the tomogram 215A. The noise may be generated in accordance with any number of noise functions (e.g., probability distribution functions), such as Gaussian noise, salt-and-pepper noise, speckle noise, Poisson noise, or uniform noise, among others. Using the noise, the noise injector 155 may produce, create, or otherwise generate at least one tomogram 215B. The tomogram 215B may be a noised version or instance of the tomogram 215A.

[0078] With the addition of noise to the tomogram 215A, the model applier 135 may input, feed, or apply the tomogram 215B to the diffusion model 160. In some embodiments, the model applier 135 may divide or partition the tomogram 215B to generate a set of tiles. Each tile may correspond to a respective portion of the tomogram 215B. With the generation, the model applier 135 can provide each tile as input to the diffusion model 160. In feeding, the model applier 135 may process the tomogram 215B (or the set of tiles) in accordance with the set of weights of the diffusion model 160 to generate an intermediate tomogram 215C (or an intermediate set of tiles). The diffusion model 160 may perform de-noising of the tomogram 215B to remove the noise added by the noise injector 155 with an objective to recover the original tomogram 215A without the noise. For example, the model applier 135 may process the tomogram 215B in accordance with DDPM architecture of the diffusion model 160 in outputting the intermediate tomogram 215C. As the input traverses through the set of weights of the diffusion model 160, additional noise may be removed from the tomogram 215B. In some embodiments, when providing the set of tiles, the model applier 135 can form the output tomogram 215C using the intermediate set of tiles from the diffusion model 160. In some embodiments, the tomogram 215C can include the perfusion map.

[0079] In some embodiments, the model applier 135 may identify or determine whether to continue with another iteration of applying the image reconstruction model 145. The model applier 135 may maintain or keep track a number of iterations correspond to a number of times the noise injector 155, the diffusion model 160, and the optimization function 165 through theimage reconstruction model 145 are applied. Upon the completion of each iteration, the model applier 135 may compare the number of iterations with a maximum number of iterations (or threshold number). The maximum number of iterations may be fixed or set for the image reconstruction model 145, and may range from 10 to 50 iterations. If the number of iterations is greater than or equal to the maximum, the model applier 135 may refrain from performing another iteration through the image reconstruction model 145. The model applier 135 may also identify or select the output tomogram 215C as the final output from the image reconstruction model 145. On the other hand, if the number of iterations is less than the maximum, the model applier 135 may perform another iteration through the image reconstruction model 145 (e g., diffusion model 160). In some embodiments, the sampling may be in accordance with a denoising diffusion implicit model (DDIM). In some embodiments, the sampling is performed subsequent to the addition of the noise by the noise injector 155.

[0080] In conjunction, the model trainer 130 may calculate, generate, or otherwise determine at least one loss metric 235 based on the tomogram 215A from the training data 205 and the output tomogram 215C from the diffusion model 160. In some embodiments, the model trainer 130 may compare the tomogram 215A from the training data 205 and the output tomogram 215C from the diffusion model 160 to determine the loss metric 235. The loss metric 235 may indicate or a degree of deviation of the output tomogram 215C relative to the initial tomogram 215A. The loss metric 235 may be generated in accordance with any number of loss functions, such as any number of loss functions, such as a norm loss (e.g., LI or L2), mean absolute error (MAE), mean squared error (MSE), a quadratic loss, a cross-entropy loss, and a Huber loss, among others. In some embodiments, the model trainer 130 may determine the loss metric 235 as a similarity metric. The similarity metric may be used as the loss metric 235, and the function used to calculate the similarity metric may include, for example, a Dice co-efficient metric, a Jaccard index, or an overlap coefficient, among others.

[0081] Using the loss metric 235, the model trainer 130 may modify, change, or otherwise update at least one parameters of the diffusion model 160. The updating of the weights may be in accordance with backpropagation algorithm and may include dropping unitsand connections within the diffusion model 160. The updating of parameters of the diffusion model 160 may be in accordance with objective function. The objective function may define one or more rates at which the parameters of the diffusion model 160 are to be updated. The objective function may be in accordance with stochastic gradient descent, and may include, for example, an adaptive moment estimation (Adam), implicit update (ISGD), and adaptive gradient algorithm (AdaGrad), among others. The updating of the parameters of the diffusion model 160 may be repeated until convergence. By updating the parameters, the diffusion model 160 may learn features particular to the subject 220. Upon completion of training, the model trainer 130 may store and maintain the set of parameters of the diffusion model 160 on the database 150 to be used to reconstruct tomograms from scan for the subject 220. The diffusion model 160 may be particular or specific to the subject 220.(0082] Referring now to FIG. 8, depicted is a block diagram of a process 300 for applying models to reconstruct tomograms from projection datasets. The process 300 may include or correspond to operation performed in the system 100 to apply the image reconstruction model 145 to generate tomograms from projection datasets. Under the process 300, the projection handler 125 on the image processing system 105 may retrieve, identify, or otherwise receive at least one projection dataset 310 for at least one subject 320. In some embodiments, the projection handler 125 may obtain, fetch, or otherwise acquire the projection dataset 310 via the imaging device 110. The projection dataset 310 may contain or include raw data acquired by the imaging device 110 in accordance with an imaging modality (e g., x-ray, SPECT, or CBCT imaging). Upon acquisition, the imaging device 110 may send, transmit, or otherwise provide the projection dataset 310 to the image processing system 105.

[0083] The projection dataset 310 may correspond to at least a segment 340 of at least one arc 335 about a volume 325, including a structure of interest (SOI) 330 in the subject 320. For example, the projection dataset 310 may correspond to at least a scan acquired from different angles along arc or a portion of the arc (e g., ranging between 15-90 degrees) about the volume 325 of the subject 320. The projection dataset 310 may have any number of projections, ranging between 20 to 90 projections. The scan may be taken as a beam of energy (e.g., x-ray) passingthrough the volume 325 of the subject 320. The beam of energy may be attenuated by the objects within the volume 325 of the subject 320, such as tissues, bone, fat, and organs, among others. The arc 335 may correspond to a circular region or a half-circular region about the volume 325 of the subject 320. For example, the arc 335 may include a cylindrical slice (e.g., 360 degrees, + / - 5%), a half-cylindrical slice (e.g., 180 degrees, + / - 5%), or a quarter cylindrical slice about (e.g., 90 degrees, + / - 5%) the volume 325 of the subject 320, or a smaller angle cylindrical slice, among others. The segment 340 may correspond to a portion of the arc 335. The segment 340 may range between 15-90 degrees of the overall arc 335 about the volume 325 of the subject 320. The segment 340 may be, for example, at least one of 15 degrees, 17.5 degrees, 20 degrees, 22.5 degrees, 25 degrees, 30 degrees, 35 degrees, 40 degrees, 45 degrees, 50 degrees, 55 degrees, 60 degrees, 65 degrees, 70 degrees, 75 degrees, 80 degrees, 85 degrees, or 90 degrees, among others, of the arc 335 about the volume 325 of the subject 320. In some embodiments, the projection dataset 310 may include a set of frames forming a video of raw data from which a video of tomograms is to be reconstructed. Each frame can correspond to a respective sample time in time-series data forming the projection dataset 310. With the establishment of the image reconstruction model 145 for the subject, tomograms may be reconstructed from projection datasets derived from just a portion (e.g., the segment 340) of the arc about the volume of the subject.

[0084] The subject 320 may be a human or animal subject, among others. The subject 320 may the same as the subject 220. The subject 320 may have, may be at risk of, or may be diagnosed with at least one condition. The condition may include, for example, brain cancer, breast cancer, head and neck cancer, lung cancer, prostate cancer, colorectal cancer, skin cancer, bladder cancer, pancreatic cancer, liver cancer, renal cancer, ovarian cancer, cervical cancer, and the like. The subject 320 may have been or may be concurrently administered with a radiotherapy 332 to treat the cancer. The radiotherapy 332 can include, for example external beam radiation therapy (EBRT), brachytherapy, proton therapy, radioimmunotherapy, radiosurgery, among others. The radiotherapy 332 may be delivered, provided, or otherwise administered via a radiotherapy device, such as a linear accessor (LINAC).[0(185] The volume 325 may correspond to a section, a portion, a region, or an area of the subject 320. The volume 325 may include, for example, a head, a neck, a torso, a back, a pelvis, arms, legs, or any organ of the subject 320. The volume 325 can include the SOI 330 within the subject 320. In some embodiments, the volume 325 may correspond to a region on the subject 320 in which cancer is present. The SOI 330 may be at least one organ to be administered for radiotherapy 332 for cancer within the corresponding volume 325 of the subject 320. The organ can be any organ within the subject 320 such as, brain, breast, lung, liver, bone, lymph node, prostate, bladder, colon, rectum, prostate, pancreas, kidney, ovaries, uterus, cervix gastrointestinal tract, pelvis, spine, or a tissue, among others, among others. For example, if the volume 325 is a two-dimensional slice or three-dimensional region about the head of the subject 320, the SOI 330 can be the brain or a specific portion of the brain. In some embodiments, the SOI 230 may correspond to or may include at least one blood vessel in the organ (e.g., blood vessels in the brain).

[0086] In some embodiments, the projection dataset 310 may be acquired in accordance with perfusion imaging (e.g., CBCT perfusion imaging). The projection dataset 310 may correspond to the volume 225, including the SOI 230 corresponding to at least one blood vessel (e.g., within the brain or another organ). To prepare for perfusion imaging, the blood vessel of the subject 320 may be injected with a contrast agent (e.g., an iodinated contrast agent, such as iohexel, iopamidol, iversol, or iodixanol). The contrast agent may be to increase the visibility or contrast of the blood vessels within the resultant imaging. With the injection, the imaging device 110 may perform imaging (e.g., CBCT or CT imaging) of the volume 325 to generate the projection dataset 310.

[0087] With the receipt, the model applier 135 may input, feed, or otherwise apply the projection dataset 310 to the image reconstruction model 145. The model applier 135 may input, feed, or apply the projection dataset 310 to the optimization function 165. In processing, the model applier 135 may execute the image reconstruction algorithm of the optimization function 165 (e.g., CG or FBR) to output, produce, or otherwise generate an initial tomogram 315A. In some embodiments, the optimization function 165 may include at least one sparsifier function.The sparsifier function may be used to generate a sparser representation of an input. For the initial iteration, the input for the optimization function 165 may include the projection dataset 310 and another tomogram (e.g., derived from a reference tomogram 315B). The initial tomogram 315A may correspond to an initial reconstruction from the raw data in the projection dataset 310. For the subsequent iterations, the input may include output tomogram from the diffusion model 160 from a current iteration and an output tomogram from the diffusion model 160 from a previous iteration. The tomogram 315A may have a set of pixels (or pixels) in a resolution in two dimensions (e.g., in x and y axes) or three dimensions (e.g., in x, y, and z axes). The resolution may range from 512 by 512 pixels to 1024 by 1024 pixels per two-dimensional slice. In some embodiments, the tomogram 315A may include a set of frames forming a video scan. Each frame can correspond to a respective sample time of a sequence of time samples.(0088] The model applier 135 may input, feed, or otherwise apply the initial tomogram 315A to the noise injector 155. In some embodiments, the model applier 135 may sample at least a portion of the tomogram 315A to input the ML model 145 for generating additional tomograms. For instance, the model applier 135 may down-sample the initial tomogram 315A to reduce the resolution from 1024 by 1024 pixels to 356 by 356 pixels to use as input to the image reconstruction model 145. For the first iteration of the image reconstruction model 145, the model applier 135 may apply at least one reference tomogram 315B to the image reconstruction model 145 (e.g., at the noise injector 155). The reference tomogram 315B may correspond to at least one of the examples in the training data 205. For subsequent iterations, the model applier 135 may apply the output tomogram from the optimization function 165 to the noise injector 155.

[0089] In processing, the model applier 135 may input, feed, or apply the initial tomogram 315A (or the reference tomogram 315B) to the noise injector 155. Upon applying, the noise injector 155 may inject, include, or otherwise add the noise to the tomogram 315A (or 315B. The noise may be generated in accordance with any number of noise functions (e g., probability distribution functions), such as Gaussian noise, salt-and-pepper noise, speckle noise, Poisson noise, or uniform noise, among others. Using the noise, the noise injector 155 mayproduce, create, or otherwise generate at least one tomogram 315C. The tomogram 315C may correspond to a noised version or instance of the tomogram 315A or 315B. For the first iteration, the tomogram 315C may be a noised version or instance the reference tomogram 315B. For the subsequent iterations, the tomogram 315C may be a noised version or instance the output tomogram from the optimization function 165.

[0090] With the addition of noise to the tomogram 315A or 315B, the model applier 135 may input, feed, or apply the tomogram 315C to the diffusion model 160. In some embodiments, the model applier 135 may divide or partition the tomogram 315C to generate a set of tiles. Each tile may correspond to a respective portion of the tomogram 315C. With the generation, the model applier 135 can provide each tile as input to the diffusion model 160. In feeding, the model applier 135 may process the input tomogram (or tiles) in accordance with the set of weights of the diffusion model 160 to generate an intermediate tomogram 315D (or an intermediate set of tiles). The diffusion model 160 may perform de-noising of the tomogram 315C to remove the noise added by the noise injector 155 with an objective to recover the original tomogram 315A or 315B without the noise. For example, the model applier 135 may process the tomogram 315C in accordance with DDPM architecture of the diffusion model 160 in outputting the intermediate tomogram 315D. As the input traverses through the set of weights of the diffusion model 160, additional noise may be removed from the tomogram 315C. In some embodiments, when providing the set of tiles, the model applier 135 can form the output tomogram 315C using the intermediate set of tiles from the diffusion model 160.[00911 Continuing on, the model applier 135 may input, feed, or apply the intermediate tomogram 315D to the optimization function 165. In processing, the model applier 135 may execute the image reconstruction algorithm of the optimization function 165 (e.g., CG or FBR) to output, produce, or otherwise generate an output tomogram 315E. In some embodiments, the optimization function 165 may include at least one sparsifier function. The sparsifier function may be used to generate a sparser representation of an input as detailed herein. For the initial iteration, the input for the optimization function 165 may include the projection dataset 310 and another tomogram (e.g., derived from a reference tomogram 315B). For the subsequentiterations, the input may include output tomogram from the diffusion model 160 from a current iteration and an output tomogram from the diffusion model 160 from a previous iteration.(0092] The model applier 135 may identify or determine whether to continue with another iteration of applying the image reconstruction model 145. The model applier 135 may maintain or keep track a number of iterations correspond to a number of times the noise injector 155, the diffusion model 160, and the optimization function 165 through the image reconstruction model 145 are applied. Upon the completion of each iteration, the model applier 135 may compare the number of iterations with a maximum number of iterations (or threshold number). The maximum number of iterations may be fixed or set for the image reconstruction model 145, and may range from 10 to 50 iterations. If the number of iterations is greater than or equal to the maximum, the model applier 135 may refrain from performing another iteration through the image reconstruction model 145. The model applier 135 may also identify or select the output tomogram 315E as the final output from the image reconstruction model 145. On the other hand, if the number of iterations is less than the maximum, the model applier 135 may perform another iteration through the image reconstruction model 145. In some embodiments, the model applier 135 may sample the output tomogram 315E to produce, create, or otherwise generate a sampled tomogram for the next iteration. In some embodiments, the sampling may be in accordance with a denoising diffusion implicit model (DDIM). In some embodiments, the sampling is performed subsequent to the addition of the noise by the noise injector 155. With the generation, the model applier 135 may repeat the process detailed herein using the sampled tomogram.(0093] In some embodiments, the output tomogram 315E may identify or include a perfusion map (sometimes herein referred to as a perfusion image) generated by the image reconstruction model 145 using the projection dataset 310. The perfusion map may define or identify at least one characteristic associated with the blood vessel in the volume 325 of the subject 320. The characteristic may include, for example, one or more of a blood flow rate (e.g., volume of blood passing through the organ or the volume in a given time), a blood volume (e.g., amount of blood present in the organ or the volume), a mean transmit time (MTT) (e.g., averagetime for blood to pass through the organ or volume), or a time-to-peak (TTP), among others. For instance, the perfusion map of the output tomogram 315E can include a set of pixels (or voxels). Each pixel can define or include a value (e.g., an intensity value or a red-green-blue (RGB) value) corresponding to the degree of the characteristic.|0094| Referring now to FIG. 9, depicted is a block diagram of a process 400 for generating output based on tomograms reconstructed from projection dataset. Under the process 400, the output evaluator 140 executing on the image processing system 105 may store and maintain an association between the subject 320 (e.g., using a subject identifier) and the output tomogram 315E on the database 150 using one or more data structures. The association may be a link, a map, or a relation between the subject 320 and the output tomogram 315E. The one or more data structures can include an array, a linked list, a stack, a tree, a hash table, among others. The association may be also with the projection dataset 310 and information related to the subject 320, such as parameters characterizing the radiotherapy 332 administered at least in partial concurrence to the subject 320. The output tomogram 315E may be used for any number of applications, such as image segmentation, classification, or localization. For example, the output tomogram 315E may be provided by the output evaluator 140 to another machine learning model to generate a segmented image or classify the subject 320 based on the SOI 330 or cancer.(0095] The output evaluator 140 may create, produce, or otherwise generate at least one output 405 based on the output tomogram 315E and the administration of the radiotherapy 332 to the subject 320. The output 405 may include information about the administration of the radiotherapy 332. In some embodiments, the output evaluator 140 may retrieve, obtain, or otherwise identify the information, including a set of parameters defining the administration of the radiotherapy 332. The set of parameters can include or identify, for example: a target (e.g., gross tumor volume (GTV)) to which the radiotherapy 332 is delivered; a dosage (e.g., amount of radiation as measured in grays); a dose rate (e.g., a rate at which the dosage is delivered); a beam energy; a beam modulation; beam orientation; and a time duration, among others. With the retrieval, the output evaluator 140 may include the set of parameters defining the administration of the radiotherapy 332 as the output 405, along with the output tomogram 315E.

[0896] With the generation, the output evaluator 140 may send, transmit, or otherwise provide the output 405 to the administrative device 115 for presentation via a user interface 410. The output 405 may include the output tomogram 315E and information associated with the administration of the radiotherapy to the subject 320. Upon receipt, the administrative device 115 may render, display, or otherwise present the output 405, including the output tomogram 315E and the information about the radiotherapy 332 via the user interface 410. The output evaluator 140 may provide the user interface 410 to facilitate the presentation of the output 405. The user interface 410 may include one or more user interface element (e.g., graphical user interface elements) to generate additional information for the radiotherapy 332. A clinician (e g., a physician) examining the subject 320 may use the information presented via the administrative device 115 to adjust to the radiotherapy plan for the subject 320 and other decisions related to therapy. The user interface 410 may accept, obtain, or otherwise receive an input 415 to define the additional information (including adjustments) for the radiotherapy 332.

[0097] Using the input 415, the output evaluator 140 may produce, create, or otherwise output 405’. In some embodiments, the output evaluator 140 may determine or generate the output 405’ to identify or include a treatment assurance, indicating one of failure or success in administering the radiotherapy332 to the subject 320. For example, the treatment assurance may indicate whether the radiotherapy 332 was successful at reducing the size of tumor on the organ corresponding to the SOI 330 in the subject 320. The input 415 may include an indication of the whether the radiotherapy 332 was successful based on a comparison between the target as defined by the set of parameters and the tomogram 315E.(0098] In some embodiments, the output evaluator 140 may determine or generate the output 405’ to identify or include a dosimetry analysis, indicating a dosage of the radiotherapy 332 administered to the subject 320. For instance, the dosimetry analysis may indicate an amount of dosage that the region of the organ corresponding to the SOI 330 received when administered with the radiotherapy 332. The input 415 received via the user input (e g., by a clinician examining the tomogram 315E) may include an adjustment to the dosage as predefinedin the set of parameters for the administration of the radiotherapy 332. The adjustment can be based on a discrepancy between the beam orientation or beam energy and the predefined dosage.

[0099] In some embodiments, the output evaluator 140 may determine or generate the output 405’ to identify or include a treatment adjustment. The treatment adjust may define or identify modifications to the administration of the radiotherapy 332. For example, the treatment adjustment may identify or indicate changes to the dosage of the radiation or orientation of the applicator with respect to the organ corresponding to the SOI 330 in the subject 320. The adjustment may be defined using the input 415 via the administrative device 115. In some embodiments, the output 405’ may identify or include a set of treatment parameters defining the administration of the radiotherapy 332. For instance, the treatment parameters may include or identify a dose parameter (e.g., amount of radiation), fractionation, dose rate, beam energy, beam type, beam geometry, beam angle, beam orientation, subject positioning, or subject orientation, among others.

[0100] With the generation, the output evaluator 140 may send, transmit, or otherwise provide the output 405’ to the administrative device 115. Upon receipt, the administrative device 115 may render, display, or otherwise present the output 405’, including additional information about the radiotherapy 332. For example, the administrative device 115 may present containing the output tomogram 315E and related information, such as the treatment assurance, dosimetry analysis, or the treatment adjustment, via the user interface 410. The clinician (e.g., a physician) examining the subject 320 may use the information presented via the administrative device 115 to adjust to the radiotherapy 332 for the subject 320 and other decisions related to therapy. Subsequent to the presentation of the output 405’, the administration of the radiotherapy 332 to the SOI 330 in the subject 320 may be modified (e.g., by the clinician or technician via the LINAC). The modification may be a change to any of the set of parameters of the administration of the radiotherapy 332. The modification may include, for example, one or more of: a change in the beam orientation; an increase or decrease in dosage; an increase or decrease in dose rate; an increase or decrease in time duration, among others. Upon modification, the SOI 330 may be administered with the radiotherapy 332 (e.g., via the LINAC).[01011 In this manner, with the establishment of the image reconstruction model 145, the image processing system 105 may generate tomograms 315E from projection datasets 310 from just a segment (e.g., 15-90 degrees) of the arc used to take scans of the volume 325 of the subject 320. The acquisition may be completed in less than one breath-hold (e.g., less than 30- 90 seconds) for the subject 320, taking much less time than approaches that rely on the full quarter, half, or full circular scanning. Projection datasets 310 with less than the full arc tend to be noisy when reconstructed as tomograms. The image reconstruction model 145 may be able to de-noise and create tomograms with clear and distinct features (e.g., tumors, organs, and other structures of interest). This may allow the image reconstruction model 145 to reconstruct tomograms in near-real time and permit the clinician examining the subject 320 to adjust radiotherapy using the information provided by the image processing system 105. The quick and faster reconstruction of tomograms from the projection datasets can also allow for perfusion imaging using CBCT, thereby providing clinicians with valuable insight regarding the blood flow and anatomy of the subject. From a computing resource perspective, the image processing system 105 may reduce the amount of processing time and memory used to reconstruct tomograms from projection datasets, relative to approaches that rely on the full arc, thereby freeing up the computing resources for other processes.|0102| Referring now to FIG. 10, depicted is a flow diagram of a method 500 of reconstructing tomograms from projection datasets. The method 500 may be implemented by any components detailed herein, such as the system 100 or 700. Under the method 500, a computing system may receive a projection dataset (505). The computing system may apply an image reconstructor to generate an optimization function (510). The computing system may add noise to the initial tomogram (515). The computing system may apply a diffusion model to the initial tomogram with noise to generate an intermediate tomogram (520). The computing system may determine whether to continue with another iteration (525). If the determination is to continue with another iteration, the computing system may repeat the process from (510).Otherwise, if the determination is to not to continue with another iteration, the computing system may generate a final tomogram (530). The computing system may provide an output (535).

[0103] Referring now to FIG. 11, depicted is a flow diagram of a method 600 of training models to reconstruct tomograms from projection datasets. The method 600 may be implemented by any components detailed herein, such as the system 100 or 700. Under the method 600, a computing system may identify a tomogram from training data (605). The computing system may add noise to the initial tomogram (610). The computing system may apply a diffusion model to the initial tomogram with noise to generate an intermediate tomogram (615). The computing system may determine whether to continue with another iteration (620). If the determination is to continue with another iteration, the computing system may repeat the process from (610). Otherwise, if the determination is to not to continue with another iteration, the computing system may determine a loss metric by comparing the output tomogram with the original tomogram from the training data (625). The computing system may update the model using the loss metric (630).E. Computing and Network Environment

[0104] Various operations described herein can be implemented on computer systems. FIG. 12 shows a simplified block diagram of a representative server system 700, client computing system 714, and network 726 usable to implement certain embodiments of the present disclosure. In various embodiments, server system 700 or similar systems can implement services or servers described herein or portions thereof. Client computing system 714 or similar systems can implement clients described herein. The system 700 described herein can be similar to the server system 700. Server system 700 can have a modular design that incorporates a number of modules 702 (e.g., blades in a blade server embodiment); while two modules 702 are shown, any number can be provided. Each module 702 can include processing unit(s) 704 and local storage 706.

[0105] Processing unit(s) 704 can include a single processor, which can have one or more cores, or multiple processors. In some embodiments, processing unit(s) 704 can include a general-purpose primary processor, as well as one or more special-purpose co-processors, such as graphics processors, digital signal processors, or the like. In some embodiments, some or allprocessing units 704 can be implemented using customized circuits, such as application specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs). In some embodiments, such integrated circuits execute instructions that are stored on the circuit itself. In other embodiments, processing unit(s) 704 can execute instructions stored in local storage 706. Any type of processors in any combination can be included in processing unit(s) 704.

[0106] Local storage 706 can include volatile storage media (e.g., DRAM, SRAM, SDRAM, or the like) and / or non-volatile storage media (e.g., magnetic or optical disk, flash memory, or the like). Storage media incorporated in local storage 706 can be fixed, removable, or upgradeable as desired. Local storage 706 can be physically or logically divided into various subunits, such as a system memory, a read-only memory (ROM), and a permanent storage device. The system memory can be a read-and-write memory device or a volatile read-and-write memory, such as dynamic random-access memory. The system memory can store some or all of the instructions and data that processing unit(s) 704 need at runtime. The ROM can store static data and instructions that are needed by processing unit(s) 704. The permanent storage device can be a non-volatile read-and-write memory device that can store instructions and data even when module 702 is powered down. The term “storage medium” as used herein includes any medium in which data can be stored indefinitely (subject to overwriting, electrical disturbance, power loss, or the like) and does not include carrier waves and transitory electronic signals propagating wirelessly or over wired connections.

[0107] In some embodiments, local storage 706 can store one or more software programs to be executed by processing unit(s) 704, such as an operating system and / or programs implementing various server functions, such as functions of the system 100 of FIG. 1 or any other system described herein, or any other server(s) associated with system 100 or any other system described herein.|0108| “Software” refers generally to sequences of instructions that, when executed by processing unit(s) 704 cause server system 700 (or portions thereof) to perform various operations, thus defining one or more specific machine embodiments that execute and performthe operations of the software programs. The instructions can be stored as firmware residing in read-only memory and / or program code stored in non-volatile storage media that can be read into volatile working memory for execution by processing unit(s) 704. Software can be implemented as a single program or a collection of separate programs or program modules that interact as desired. From local storage 706 (or non-local storage described below), processing unit(s) 704 can retrieve program instructions to execute and data to process in order to execute various operations described above.

[0109] In some server systems 700, multiple modules 702 can be interconnected via a bus or other interconnect 708, forming a local area network that supports communication between modules 702 and other components of server system 700. Interconnect 708 can be implemented using various technologies including server racks, hubs, routers, etc.

[0110] A wide area network (WAN) interface 710 can provide data communication capability between the local area network (interconnect 708) and the network 726, such as the Internet. Technologies can be used, including wired (e.g., Ethernet, IEEE 802.3 standards) and / or wireless technologies (e.g., Wi-Fi, IEEE 802.24 standards).|01H| In some embodiments, local storage 706 is intended to provide working memory for processing unit(s) 704, providing fast access to programs and / or data to be processed, while reducing traffic on interconnect 708. Storage for larger quantities of data can be provided on the local area network by one or more mass storage subsystems 712 that can be connected to interconnect 708. Mass storage subsystem 712 can be based on magnetic, optical, semiconductor, or other data storage media. Direct attached storage, storage area networks, network-attached storage, and the like can be used. Any data stores or other collections of data described herein as being produced, consumed, or maintained by a service or server can be stored in mass storage subsystem 712. In some embodiments, additional data storage resources may be accessible via WAN interface 710 (potentially with increased latency).

[0112] Server system 700 can operate in response to requests received via WAN interface 710. For example, one of modules 702 can implement a supervisory function and assign discrete tasks to other modules 702 in response to received requests. Work allocation techniques can be used. As requests are processed, results can be returned to the requester via WAN interface 710. Such operation can generally be automated. Further, in some embodiments, WAN interface 710 can connect multiple server systems 700 to each other, providing scalable systems capable of managing high volumes of activity. Other techniques for managing server systems and server farms (collections of server systems that cooperate) can be used, including dynamic resource allocation and reallocation.

[0113] Server system 700 can interact with various user-owned or user-operated devices via a wide-area network, such as the Internet. An example of a user-operated device is shown in FIG. 7 as client computing system 714. Client computing system 714 can be implemented, for example, as a consumer device, such as a smartphone, other mobile phone, tablet computer, wearable computing device (e.g., smart watch, eyeglasses), desktop computer, laptop computer, and so on.10114] For example, client computing system 714 can communicate via WAN interface 710. Client computing system 714 can include computer components, such as Processing unit(s) 716, storage device 718, network interface 720, user input device 722, and user output device 724. Client computing system 714 can be a computing device implemented in a variety of form factors, such as a desktop computer, laptop computer, tablet computer, smartphone, other mobile computing device, wearable computing device, or the like.

[0115] Processing unit(s) 716 and storage device 718 can be similar to processing unit(s) 704 and local storage 706 described above. Suitable devices can be selected based on the demands to be placed on client computing system 714; for example, client computing system 714 can be implemented as a “thin” client with limited processing capability or as a high- powered computing device. Client computing system 714 can be provisioned with program code executable by Processing unit(s) 716 to enable various interactions with server system 700.

[0116] Network interface 720 can provide a connection to the network 726, such as a wide area network (e.g., the Internet) to which WAN interface 710 of server system 700 is also connected. In various embodiments, network interface 720 can include a wired interface (e.g., Ethernet) and / or a wireless interface implementing various RF data communication standards, such as Wi-Fi, Bluetooth, or cellular data network standards (e.g., 3G, 4G, LTE, etc.).

[0117] User input device 722 can include any device (or devices) via which a user can provide signals to client computing system 714; client computing system 714 can interpret the signals as indicative of particular user requests or information. In various embodiments, user input device 722 can include any or all of a keyboard, touch pad, touch screen, mouse or other pointing device, scroll wheel, click wheel, dial, button, switch, keypad, microphone, and so on.

[0118] User output device 724 can include any device via which client computing system 714 can provide information to a user. For example, user output device 724 can include a display to present images generated by or delivered to client computing system 714. The display can incorporate various image generation technologies, e.g., a liquid crystal display (LCD), lightemitting diode (LED) including organic light-emitting diodes (OLED), projection system, cathode ray tube (CRT), or the like, together with supporting electronics (e.g., digital-to-analog or analog-to-digital converters, signal processors, or the like). Some embodiments can include a device, such as a touchscreen that function as both input and output device. In some embodiments, other user output devices 724 can be provided in addition to or instead of a display. Examples include indicator lights, speakers, tactile “display” devices, printers, and so on.

[0119] Some embodiments include electronic components, such as microprocessors, storage and memory that store computer program instructions in a computer-readable storage medium. Many of the features described in this specification can be implemented as processes that are specified as a set of program instructions encoded on a computer-readable storage medium. When these program instructions are executed by one or more processing units, they cause the processing unit(s) to perform various operation indicated in the program instructions.Examples of program instructions or computer code include machine code, such as is produced by a compiler, and files including higher-level code that are executed by a computer, an electronic component, or a microprocessor using an interpreter. Through suitable programming, processing unit(s) 704 and 717 can provide various functionality for server system 700 and client computing system 714, including any of the functionality described herein as being performed by a server or client, or other functionality.

[0120] It will be appreciated that server system 700 and client computing system 714 are illustrative and that variations and modifications are possible. Computer systems used in connection with embodiments of the present disclosure can have other capabilities not specifically described here. Further, while server system 700 and client computing system 714 are described with reference to particular blocks, it is to be understood that these blocks are defined for convenience of description and are not intended to imply a particular physical arrangement of component parts. For instance, different blocks can be, but need not be, located in the same facility, in the same server rack, or on the same motherboard. Further, the blocks need not correspond to physically distinct components. Blocks can be configured to perform various operations, e.g., by programming a processor or providing appropriate control circuitry, and various blocks might or might not be reconfigurable depending on how the initial configuration is obtained. Embodiments of the present disclosure can be realized in a variety of apparatus, including electronic devices implemented using any combination of circuitry and software.[01211 While the disclosure has been described with respect to specific embodiments, one skilled in the art will recognize that numerous modifications are possible. Embodiments of the disclosure can be realized using a variety of computer systems and communication technologies, including but not limited to, the specific examples described herein. Embodiments of the present disclosure can be realized using any combination of dedicated components and / or programmable processors and / or other programmable devices. The various processes described herein can be implemented on the same processor or different processors in any combination. Where components are described as being configured to perform certain operations, suchconfiguration can be accomplished, e.g., by designing electronic circuits to perform the operation, by programming programmable electronic circuits (such as microprocessors) to perform the operation, or any combination thereof. Further, while the embodiments described above may make reference to specific hardware and software components, those skilled in the art will appreciate that different combinations of hardware and / or software components may also be used and that particular operations described as being implemented in hardware might also be implemented in software or vice versa.

[0122] Computer programs incorporating various features of the present disclosure may be encoded and stored on various computer-readable storage media; suitable media include magnetic disk or tape, optical storage media, such as compact disk (CD) or DVD (digital versatile disk), flash memory, and other non-transitory media. Computer-readable media encoded with the program code may be packaged with a compatible electronic device, or the program code may be provided separately from electronic devices (e.g., via Internet download or as a separately packaged computer-readable storage medium).[01231 Thus, although the disclosure has been described with respect to specific embodiments, it will be appreciated that the disclosure is intended to cover all modifications and equivalents within the scope of the following claims.

Claims

WHAT IS CLAIMED IS:

1. A method of reconstructing tomograms from projection datasets, comprising: receiving, by one or more processors, a projection dataset corresponding to a portion of an arc about a first structure of interest (SOI) in a first volume of a subject; providing, by the one or more processors, the projection dataset as input to a machine learning (ML) model, wherein the ML model is established using a plurality of examples, each of the plurality of examples including a respective tomogram of a second SOI in a second volume of the subject; generating, by the one or more processors, based on providing the projection dataset as the input to the ML model, a tomogram of the first SOI in the first volume of the subject; and storing, by the one or more processors, using one or more data structures, an association between the subject and the tomogram.

2. The method of claim 1, wherein receiving the projection dataset further comprises acquiring, via a tomograph, the projection dataset, at least partially concurrent to administration of radiotherapy to the first SOI of the subject, and further comprising: generating, by the one or more processors, an output based on the tomogram and the administration of the radiotherapy.

3. The method of claim 2, wherein generating the output further comprises generating, using user input, the output including at least one of: (i) a treatment assurance indicating one of failure or success at administering the radiotherapy, (ii) a dosimetry analysis indicating a dosage of the radiotherapy administered or (iii) a treatment adjustment to identify modifications to the administration of the radiotherapy.

4. The method of claim 2, further comprising providing, by the one or more processors, for presentation via a user interface, the output based on the tomogram and the administration of the radiotherapy,wherein the administration of the radiotherapy to the first SOI is modified, subsequent to provision of the output.

5. The method of claim 1, wherein the subject is administered with radiotherapy at least partially concurrent with acquisition of the projection dataset, and wherein the radiotherapy comprises at least one of an intensity-modulated radiation therapy (IMRT), stereotactic body radiation therapy (SBRT), image-guided radiation therapy (IGRT), or brachytherapy.

6. The method of claim 1, wherein providing the projection dataset to the ML model further comprises iteratively applying the ML model comprising: an optimization function configured to generate an output tomogram based on the projection dataset and a reference tomogram from at least one of the plurality of examples, and a diffusion model configured to generate an intermediate tomogram using the output tomogram from the optimization function.

7. The method of claim 6, wherein the optimization function comprises a sparsifier function configured to generate a sparse representation of the projection dataset and the reference tomogram.

8. The method of claim 1, further comprising determining, by the one or more processors, that a number of iterations of applying the ML model is greater than or equal to a maximum number, and wherein generating the tomogram further comprises identifying the tomogram as output for the ML model, responsive to determining that the number of iterations is greater than or equal to the maximum number.

9. The method of claim 1, wherein receiving the projection dataset further comprises receiving the projection dataset corresponding to the portion of the arc about the first SOI acquired viaperfusion cone beam computed tomography (CBCT) imaging, the first SOI comprising a blood vessel in the first volume of the subject, the blood vessel injected with a contrast agent, and wherein generating the tomogram further comprises generating the tomogram including a perfusion map identifying at least one characteristic associated with the blood vessel in the first volume of the subject, the at least one characteristic comprising at least one of a blood flow rate, a blood volume, a mean transmit time (MTT), or a time-to-peak (TTP).

10. The method of claim 1, wherein the subject is at risk of or diagnosed with cancer, wherein the cancer comprises at least one of brain cancer, breast cancer, head and neck cancer, lung cancer, prostate cancer, colorectal cancer, skin cancer, bladder cancer, pancreatic cancer, liver cancer, renal cancer, ovarian cancer, or cervical cancer, wherein the projection dataset is acquired in accordance with cone beam computed tomography (CBCT), wherein the first SOI corresponds to at least one organ associated with the cancer in the subject administered with radiotherapy, and wherein the portion corresponds to between 15-90 degrees of the first arc about the subject.

11. A method of training models to reconstruct tomograms from projection datasets, comprising: identifying, by one or more processors, training data identifying a first tomogram of a structure of interest (SOI) in a volume of a subject; adding, by the one or more processors, noise to the first tomogram to generate a second tomogram; providing, by the one or more processors, the second tomogram as input to a machine learning (ML) model comprising a plurality of parameters to generate a third tomogram of the SOI in the volume of the subject; determining, by the one or more processors, a loss metric based on the first tomogram from the training data and the third tomogram from the ML model;updating, by the one or more processors, at least one of the plurality of parameters of the ML model using the loss metric; and storing, by the one or more processors, the plurality of parameters of the ML model to process a projection dataset corresponding to a portion of an arc about the SOI in the volume of the subject.

12. The method of claim 11, further comprising sampling, by the one or more processors, at least a portion of the first tomogram for generation of the second tomogram.

13. The method of claim 11, wherein providing the second tomogram to the ML model further comprises sampling at least a portion of an output in a first iteration of applying the ML model to applying in a second iteration of applying the ML model.

14. The method of claim 11, wherein providing the second tomogram to the ML model further comprises iteratively applying the ML model further comprising a diffusion model configured to generate an intermediate tomogram using the second tomogram.

15. The method of claim 14, wherein an optimization function to be used with the diffusion model comprises a sparsifier function configured to generate a sparse representation of the first tomogram.

16. The method of claim 11, wherein determining the loss metric further comprises determining the loss metric to indicate a degree of deviation of the third tomogram relative to the first tomogram.

17. The method of claim 11, further comprising determining, by the one or more processors, that a number of iterations of applying the ML model is greater than or equal to a maximum number, andwherein generating the tomogram further comprises identifying the tomogram as output for the ML model, responsive to determining that the number of iterations is greater than or equal to the maximum number.

18. The method of claim 11, further comprising determining, by the one or more processors, that a number of iterations of applying the ML model is less than a maximum number, and wherein generating the tomogram further comprises continue applying the tomogram to the ML model, responsive to determining that the number of iterations is less than the maximum number.

19. The method of claim 11, wherein the tomogram includes a perfusion map identifying at least one characteristic associated with a blood vessel in the volume of the subject, the at least one characteristic comprising at least one of a blood flow rate, a blood volume, a mean transmit time (MTT), or time-to-peak (TTP).

20. The method of claim 11, wherein the first tomogram is acquired in accordance with cone beam computed tomography (CBCT), wherein the portion for the projection dataset corresponds to between 15-90 degrees of the first arc about the subject.

21. A system for reconstructing tomograms from projection datasets, comprising: one or more processors coupled with memory, configured to: receive a projection dataset corresponding to a portion of an arc about a first structure of interest (SOI) in a first volume of a subject; provide the projection dataset as input to a machine learning (ML) model, wherein the ML model is established using a plurality of examples, each of the plurality of examples including a respective tomogram of a second SOI in a second volume of the subject; generate, based on applying the projection dataset to the ML model, a tomogram of the first SOI in the first volume of the subject; andstore, using one or more data structures, an association between the subject and the tomogram.

22. The system of claim 21, wherein the one or more processors are further configured to: acquire, via a tomograph, the projection dataset, at least partially concurrent to administration of radiotherapy to the first SOI of the subject, and further comprising: generate an output based on the tomogram and the administration of the radiotherapy.

23. The system of claim 22, wherein the one or more processors are further configured to generate, using user input, the output including at least one of: (i) a treatment assurance indicating one of failure or success at administering the radiotherapy, (ii) a dosimetry analysis indicating a dosage of the radiotherapy administered or (iii) a treatment adjustment to identify modifications to the administration of the radiotherapy.

24. The system of claim 22, wherein the one or more processors are further configured to provide, for presentation via a user interface, the output based on the tomogram and the administration of the radiotherapy, wherein the administration of the radiotherapy to the first SOI is modified, subsequent to provision of the output.

25. The system of claim 21, wherein the subject is administered with radiotherapy at least partially concurrent with acquisition of the projection dataset, and wherein the radiotherapy comprises at least one of an intensity-modulated radiation therapy (IMRT), stereotactic body radiation therapy (SBRT), image-guided radiation therapy (IGRT), or brachytherapy.

26. The system of claim 21, wherein the one or more processors are further configured to iteratively apply the ML model comprising: an optimization function configured to generate an output tomogram using the projection dataset and a reference tomogram from at least one of the plurality of examples, anda diffusion model configured to generate an intermediate tomogram using the output tomogram from the optimization function.

27. The system of claim 26, wherein the optimization function comprises a sparsifier function configured to generate a sparse representation of the projection dataset and the reference tomogram.

28. The system of claim 21, wherein the one or more processors are further configured to: determine that a number of iterations of applying the ML model is greater than or equal to a maximum number, and identify the tomogram as output for the ML model, responsive to determining that the number of iterations is greater than or equal to the maximum number.

29. The system of claim 21, wherein the one or more processors are further configured to: receive the projection dataset corresponding to the portion of the arc about the first SOI acquired via perfusion cone beam computed tomography (CBCT) imaging, the first SOI comprising a blood vessel in the first volume of the subject, the blood vessel injected with a contrast agent; and generate the tomogram including a perfusion map identifying at least one characteristic associated with the blood vessel in the first volume of the subject, the at least one characteristic comprising at least one of a blood flow rate, a blood volume, a mean transmit time (MTT), or time-to-peak (TTP).

30. The system of claim 21, wherein the subject is at risk of or diagnosed with cancer, wherein the cancer comprises at least one of: brain cancer, breast cancer, head and neck cancer, lung cancer, prostate cancer, colorectal cancer, skin cancer, bladder cancer, pancreatic cancer, liver cancer, renal cancer, ovarian cancer, or cervical cancer, wherein the projection dataset is acquired in accordance with cone beam computed tomography (CBCT),wherein the first SOI corresponds to at least one organ associated with the cancer in the subject administered with radiotherapy, and wherein the portion corresponds to between 15-90 degrees of the first arc about the subject.

31. A system for training models to reconstruct tomograms from projection datasets, comprising: one or more processors coupled with memory, configured to: identify training data identifying a first tomogram of a structure of interest (SOI) in a volume of a subject; add noise to the first tomogram to generate a second tomogram; provide the second tomogram as input to a machine learning (ML) model comprising a plurality of parameters to generate a third tomogram of the SOI in the volume of the subject; determine a loss metric based on the first tomogram from the training data and the third tomogram from the ML model; update at least one of the plurality of parameters of the ML model using the loss metric; and store the plurality of parameters of the ML model to process a projection dataset corresponding to a portion of an arc about the SOI in the volume of the subject.

32. The system of claim 31, wherein the one or more processors are further configured to sample at least a portion of the first tomogram for generation of the second tomogram.

33. The system of claim 31, wherein the one or more processors are further configured to sample at least a portion of an output in a first iteration of applying the ML model to applying in a second iteration of applying the ML model.

34. The system of claim 31, wherein the one or more processors are further configured to iteratively applying the ML model further comprising a diffusion model configured to generate an intermediate tomogram using the second tomogram.

35. The system of claim 34, wherein an optimization function to be used with the diffusion model comprises a sparsifier function configured to generate a sparse representation of the first tomogram.

36. The system of claim 31, wherein the one or more processors are further configured to determine the loss metric to indicate a degree of deviation of the third tomogram relative to the first tomogram.

37. The system of claim 31, wherein the one or more processors are further configured to: determine that a number of iterations of applying the ML model is greater than or equal to a maximum number; and identify the tomogram as output for the ML model, responsive to determining that the number of iterations is greater than or equal to the maximum number.

38. The system of claim 31, wherein the one or more processors are further configured to: determine that a number of iterations of applying the ML model is less than a maximum number, and continue applying the tomogram to the ML model, responsive to determining that the number of iterations is less than the maximum number.

39. The system of claim 31, wherein the tomogram includes a perfusion map identifying at least one characteristic associated with a blood vessel in the volume of the subject, the at least one characteristic comprising at least one of a blood flow rate, a blood volume, a mean transmit time (MTT), or time-to-peak (TTP).

40. The system of claim 31, wherein the first tomogram is acquired in accordance with cone beam computed tomography (CBCT), wherein the portion for the projection dataset corresponds to between 15-90 degrees of the first arc about the subject.

Images