System and method full spectrum computer vision for photon counting computed tomography

Abstract

Systems and methods are provided, including a method for training a neural network model for analyzing computed tomography (CT) images acquired using a photon counting CT system. The method can include receiving photon counting CT data for multiple ranges of energy spectrum, and optionally a plurality of other parameters such as slice thickness and reconstruction kernel, acquired using the photon counting CT system, reconstructing a plurality of virtual monoenergetic images (VMIs) for the energy spectrum using the photon counting CT data, and generating a trained neural network model using the plurality of VMIs for the full continuous energy spectrum, and optionally along with other image acquisition and reconstruction parameters. The neural network model can be used to analyze CT images acquired using a photon counting CT system.

Description

790482.00556SYSTEM AND METHOD FULL SPECTRUM COMPUTER VISION FOR PHOTON COUNTING COMPUTED TOMOGRAPHYCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application is based on, claims priority to, and incorporates herein by reference in its entirety, US Provisional Application Serial No. 63 / 724,697, filed November 25, 2024.BACKGROUND

[0002] Photon counting CT (PCCT) systems represent a significant technological advancement in medical imaging, fundamentally changing how X-ray photons are detected and processed compared to conventional energy-integrating detectors. Unlike traditional CT systems that measure the total energy deposited by all photons, PCCT systems count individual photons and measure their respective energies, enabling true spectral imaging capabilities. These systems utilize advanced detector technologies, including cadmium telluride (CdTe), cadmium zinc telluride (CZT), or silicon-based detectors, each offering distinct advantages in terms of energy resolution, count rate capability, and spatial resolution. PCCT implementations vary in their technical approaches, with some systems employing rapid switching between different source energies (kV) while others utilize polychromatic X- ray beams combined with energy -resolving detectors that can discriminate photon energies in real-time. Regardless of the specific implementation, all PCCT systems generate comprehensive energy spectra with varying energy' and spatial resolutions, potentially providing access to detailed material composition information that was previously unavailable in clinical CT imaging.

[0003] Despite the sophisticated data acquisition capabilities of PCCT systems, current clinical implementations don’t often yield only marginal improvements over traditional energy -integrating systems. For example, the present disclosure recognizes that, while PCCT systems can theoretically acquire continuous energy spectra across a wide range of energies (typically 20-190 keV) with fine energy resolution, clinical practice remains constrained by conventional image interpretation paradigms. Traditional energy -integrating CT systems are fundamentally limited to producing images that represent the cumulative attenuation of all X- ray energies, providing no spectral discrimination. Dual-energy CT systems, though capable of some spectral analysis through energy-integrating detectors, are restricted to analyzing only two energy levels, typically achieved through rapid kV switching, dual-source configurations, or dual-layer detectors. Even with PCCT's capability to provide detailed energy -resolved data, current clinical workflows typically reconstruct only a small subset of available energyQB199638269.1 1790482.00556 information — often just one or two monoenergetic images or material decomposition maps — for radiologist interpretation. This limitation to such a narrow subset of available data means that the vast majority of spectral information — potentially containing valuable diagnostic insights — remains unused and unexplored in clinical practice.

[0004] The fundamental challenge in advancing CT imaging lies not merely in acquiring more data, but in transforming the wealth of available spectral information into clinically actionable insights without overwhelming healthcare providers. Current approaches to spectral CT interpretation rely heavily on human visual analysis, which inherently limits the number of images and energy levels that can be practically evaluated. Radiologists are already burdened with increasing image volumes and cannot feasibly review dozens of energy-specific reconstructions for each patient study. Moreover, the human visual system has limited capability to integrate and correlate information across multiple energy levels simultaneously, making it difficult to identify subtle spectral patterns that might indicate important pathophysiological changes. The challenge is compounded by the fact that optimal energyselection for different clinical applications — such as tissue characterization, contrast agent detection, or material decomposition — may vary significantly and is not well understood. Traditional approaches to energy selection are often empirical or based on theoretical considerations rather than systematic optimization for specific diagnostic tasks. Additionally, the molecular environment surrounding atoms of interest can influence spectral characteristics in ways that extend beyond simple atomic number-based material identification, suggesting that comprehensive spectral analysis could reveal information about tissue metabolism, oxygenation status, and cellular function that is currently inaccessible through conventional imaging approaches.

[0005] Therefore, there is a continuing need to provide clinicians with new information to improve clinical outcomes. However, all new information is not necessarily of value. Information that does not have clinical value, or an overwhelming amount of information, or information that cannot be readily understood or interpreted by clinicians, is not of value. Thus, there is a continuing need to provide clinicians with clinically-valuable information.SUMMARY

[0006] The present disclosure overcomes the aforementioned drawbacks by providing systems and methods for utilizing the data that can be derived from PCCT systems via a human visual system that empowers an understanding of energy selection and facilitates diagnostic interpretation. The present disclosure recognizes that information must be human-QB199638269.1 2790482.00556 consumable. Radiologists do not want more images to interpret and humans are limited in their ability to review and integrate images from multiple energies. Systems and methods are provided to assist clinicians and facilitate the evaluation of only energies from a spectrum that represent desired materials, such as calcium, iodine, gadolinium, water, and uric acid, to name just a few non-limiting examples. Specific energies can be chosen ad hoc, and in general discrete peaks can be used in this assessment with intensities being combined in intuitive ways.

[0007] In accordance w ith one aspect of the present disclosure, a photon-counting computed tomography (PCCT) system is provided that includes an x-ray source and a detector system configured to acquire raw PCCT data from a patient and a system including a processor. The processor is configured to access the raw PCCT data, reconstruct a plurality of virtual monoenergetic images from the raw PCCT data, wherein the plurality of virtual monoenergetic images spans a continuous energy spectrum, and analyze the plurality of virtual monoenergetic images using a trained neural network model configured to process multiple energy channels simultaneously to generate a diagnostic report about the patient. The system further includes a display configured to display the diagnostic report about the patient.

[0008] In accordance with another aspect of the present disclosure, a method is provided for training a full spectrum neural network model for analyzing computed tomography (CT) images acquired using a photon counting CT system. The method includes receiving photon counting CT data for a full continuous energy spectrum acquired using the photon counting CT system, reconstructing a plurality of virtual monoenergetic images (VMIs) for the full continuous energy spectrum using the photon counting CT data, and generating a trained full spectrum neural network model using the plurality of VMIs for the full continuous energy spectrum.

[0009] In accordance with yet another aspect of the present disclosure, a method is provided for analyzing computed tomography (CT) images acquired using a photon counting CT system. The method includes acquiring photon counting CT data for a full continuous energy spectrum acquired using the photon counting CT system, and reconstructing a plurality of virtual monoenergetic images (VMIs) using the photon counting CT data. Each VMI of the plurality’ of VMIs is reconstructed for one of a plurality of predetermined CT energy spectra. The method further includes providing one or more VMIs of the plurality of VMIs to a trained full spectrum neural network model configured to analyze the one or more VMIs of the plurality of VMIs. The trained full spectrum neural network model is trained using a plurality of training full spectrum VMIs generated from photon counting CT data for a full continuousQB199638269.1 3790482.00556 energy' spectrum. The method further includes generating an analysis of the one or more VMIs of the plurality of VMIs using the trained full spectrum neural network model.

[0010] In accordance with yet another aspect of the present disclosure, a system is provided for full spectrum computer vision analysis of photon counting computed tomography data. The system includes a computing device comprising a processor and memory', a data acquisition unit configured to receive raw photon counting computed tomography data from a photon counting computed tomography scanner, and a neural network model stored in the memory^ and executable by the processor. The neural network model is configured to analyze a plurality' of virtual monoenergetic images reconstructed from the raw photon counting computed tomography data across a continuous energy' spectrum to generate tissue characterization information.

[0011] These aspects are nonlimiting. Other aspects and features of the systems and methods described herein will be provided below.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention will hereafter be described with reference to the accompanying drawings, wherein like reference numerals denote like elements.

[0013] Fig. 1 is a flow chart setting forth some non-limiting, example steps of a process in accordance with the present disclosure.

[0014] Fig. 2A is a schematic diagram of a sy stem for creating a neural network model for analyzing computed tomography (CT) images acquired using a photon counting CT system in accordance with one non-limiting example of the present disclosure.

[0015] Fig. 2B is a schematic diagram illustrating a system and method for analyzing CT images using a trained neural network model in accordance with one non-limiting example of the present disclosure.

[0016] Fig. 3 is an example node from a computer vision platform knowledge base in accordance with one non-limiting example of the present disclosure.

[0017] Fig. 4 is a block diagram of an example system in accordance with one non-limiting example of the present disclosure.

[0018] Fig. 5A is a perspective view of a non-limiting example of a CT system in accordance with the present disclosure.

[0019] Fig. 5B is a block diagram of the CT system of Fig. 5A.DETAILED DESCRIPTION

[0020] The following description sets forth exemplary aspects of the present disclosure. It should be recognized, however, that such description is not intended as a limitation on theQB199638269.1 4790482.00556 scope of the present disclosure. Rather, the description also encompasses combinations and modifications to those exemplary aspects described herein.

[0021] Currently, clinical usage of PCCT systems and data is limited, despite PCCT systems being commercially available for many years. One reason for slow adoption is the limited value of the systems when compared to energy integrating systems. Despite PCCT systems acquiring more and more complex data, clinical usage of such data is limited due to a variety of limitations. In one example, usage of PCCT data may be limited, even when available, to a small number of reconstructed image series to make it human-consumable. Simply, radiologists don't want more images to interpret in PACS and humans are limited in their ability to review and integrate images from multiple energies. The human visual sy stem can typically process only 2-4 energy levels (with one set of image reconstructions parameters per energy level) simultaneously without cognitive overload, whereas PCCT systems can generate data across 150+ distinct energy7levels spanning the diagnostic X-ray spectrum (with the possibility of tuning the image reconstruction parameters at each energy level). This fundamental mismatch between data generation capacity and human processing limitations creates a significant bottleneck in clinical implementation. Therefore, even when a full spectrum of PCCT data is available, typically only energies from one or few spectrum are used, such as those that represent materials such as calcium, iodine, gadolinium, water, and uric acid (in gout). These energies are chosen ad hoc, and in general discrete peaks can be used in this assessment with intensities being combined in basic way’s. Current energy selection protocols often rely on empirical observations or manufacturer recommendations, rather than systematic optimization for specific diagnostic tasks, leading to suboptimal utilization of available spectral information. Photon counting CT should not be just an incremental improvement, it should be revolutionary.

[0022] In contrast, the present disclosure recognizes that a matrix approach to generating optimal CT images would include axes of organ and tissue disease processes / biologic function of tissue subnetworks vs CT acquisition parameters / energy spectrum. This multidimensional optimization space encompasses thousands of potential parameter combinations, including tube voltage settings (20-190 kV), detector energy thresholds, reconstruction algorithms, and temporal acquisition patterns. The matrix approach considers not only static tissue properties but also dynamic physiological processes such as perfusion, metabolism, and cellular function that may exhibit distinct spectral signatures across different energy ranges. While it may not be feasible to evaluate this large search space using manual, human visual methods, computer vision can see and interrogate the full range of acquisition parameters and full energy spectrumQB199638269.1 5790482.00556 from photon counting CT and find desired settings for each medical application, including: (1) tissue characterization (e.g., benign or malignant, metabolic status); (2) diagnosis; (3) prediction of treatment outcomes; and (4) treatment response assessment. Computer vision processes can analyze spectral data at rates exceeding 10,000 energy-parameter combinations per second, enabling comprehensive exploration of the optimization space that would require years of manual analysis. In this way, generating the most appropriate CT scan could be automated to make operating this complex CT imaging machine more manageable and more informative.

[0023] The present disclosure provides systems and methods to interrogate multiple ranges or the full energy spectrum and variations in other image acquisition parameters (slice thickness, reconstruction interval, reconstruction kernel, etc). Machine learning may be used to build Al models (e.g., neural networks) that make use of the data available from a continuous spectrum. The systems and methods can employ advanced deep learning architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer-based models specifically adapted for spectral data analysis. These architectures can process multi-dimensional spectral tensors with dimensions corresponding to spatial coordinates (x, y, z) and energy levels, enabling simultaneous analysis of anatomical and spectral features. During machine learning, the disclosed systems and methods can integrate as many source energies as necessary for desired tissue discrimination and characterization and decision making, transforming the wealth of data into actionable information. The system can dynamically adjust the number of input channels from 2 to over 200 energy levels, with automatic feature selection algorithms identifying the most diagnostically relevant energy combinations for specific clinical applications.

[0024] Different ranges, or full, continuous energy spectra detected and measured using different KV's in photon counting CT scanners can reflect the interaction between different energies of x-rays and different nuclei based not only on their atomic number, or the type of interaction, i.e., photoelectric effect or Compton effect, but also on the molecular environment of the atom of interest or on the electron clouds surrounding the atom being interrogated. The photoelectric effect dominates at lower energies (typically below 30-50 keV depending on atomic number) and provides high contrast for elements with K-edge energies within the diagnostic range, while Compton scattering becomes increasingly prevalent at higher energies and provides information about electron density and molecular structure. The molecular environment influences spectral characteristics through chemical shift effects, where the binding energy of electrons is modified by neighboring atoms and chemical bonds, creatingQB199638269.1 6790482.00556 subtle but detectable variations in absorption edge positions and shapes. These molecular environments can be analyzed into different spectral patterns and shifts within the energy spectrum, enabling tissue characterization beyond just measuring atoms like calcium, iodine, bismuth, tungsten, and the like. Advanced spectral analysis techniques can detect energy shifts as small as 0.1 keV, enabling discrimination between different oxidation states, molecular conformations, and tissue microenvironments that exhibit distinct spectral fingerprints. The present disclosure describes models to discriminate and classify tissues and disease states that use large ranges or the full continuous energy spectrum available from PCCT.

[0025] In some configurations, Al techniques can be used to interrogate the full continuous energy spectrum of photon counting CT to build desired or optimal spectral models for quantifying and classifying tissues and their disease states. These Al techniques include ensemble learning methods that combine multiple specialized models, each optimized for different energy' ranges or tissue ty pes, and meta-leaming approaches that can rapidly adapt to new clinical applications with minimal additional training data. The system employs advanced feature extraction algorithms that can identify both obvious spectral peaks and subtle inter-peak variations that may contain diagnostically relevant information. Without these techniques most of the available data will continue to go unused and decision making will be limited. Current clinical practice ty pically utilizes less than 5% of available spectral information, whereas the disclosed Al-driven approach can leverage up to 95% or 100% of the acquired spectral data for diagnostic decision-making. The systems and methods described in the present disclosure can enable spectral photon counting CT to reach its full potential.

[0026] In some configurations, a system and a method are provided for training full spectrum Al models is provided that outperform current approaches (including single and dual energy CT and photon counting CT using energy bins or peaks) in quantifying and classifying disease. The training methodology incorporates advanced data augmentation techniques specifically designed for spectral data, including energy-domain transformations, noise injection models that simulate realistic detector characteristics, and synthetic spectral generation based on known material properties. The training process utilizes distributed computing architectures that can process large spectral datasets across multiple GPUs or cloud computing resources, enabling training on datasets containing millions of spectral measurements from diverse patient populations and clinical scenarios.

[0027] The different paterns of the energy spectrum and shifts within that spectrum will reflect normal organ or tissue function and pathologic changes in an organ's tissue. These spectral paterns can be quantified using advanced signal processing techniques includingQB199638269.1 7790482.00556 wavelet analysis, Fourier transforms, and machine learning-based feature extraction that can identify both periodic and non-periodic spectral variations. Spectral pattern analysis can detect changes in tissue density, cellular organization, vascular architecture, and metabolic activity through their distinct effects on X-ray attenuation and scattering properties across different energy' levels. These patterns can also reflect changes in normal and abnormal tissue functions, biologic systems and biologic subnetworks providing unique diagnostic information about abnormal function in those tissues and subnetworks. The system can correlate spectral patterns with genomic data, proteomic profiles, and metabolomic signatures to establish comprehensive tissue characterization models that link molecular-level changes to observable spectral variations.

[0028] In the systems and methods described herein these differences can be measured across multiple ranges or the full spectrum pattern of the PCCT, not just the measurement of individual energy peaks that may be categorized into different bins, by possibly up to 24. The continuous spectrum analysis preserves inter-peak information that may contain up to 40% of the total diagnostic information available in the spectral data. In addition, in some configurations, the full spectrum can be measured and analyzed for different x-ray tube KV energy' inputs from the x-ray tube. The system can perform rapid kV switching at rates up to 1000 Hz, enabling dynamic spectral analysis that captures temporal changes in tissue properties during contrast agent uptake, physiological motion, or treatment response. By using a multiple ranges or the entire, continuous, full transmission spectrum generated by a PCCT scanner and not just appearance and shifts in easily identifiable peaks seen with different x- ray tube inputs, the present disclosure provides the ability' for clinical decision making that has not been previously possible.

[0029] A multi-dimensional matrix can use the absolute spectra or difference between spectra as a function of different x-ray inputs, and even the rate of change betw een spectra during switching between input energies. The system can calculate spectral derivatives and higher- order temporal variations that provide information about dynamic tissue processes and contrast agent kinetics. The full energy spectrum can be analyzed as a continuous function or by sampling, including very fine sampling. Sampling rates can be adjusted from 0.1 keV to 10 keV intervals depending on the specific clinical application and required energy' resolution. The present disclosure, therefore, can be designed to generate a multi-dimensional matrix w ith one parameter being the x-ray tube output kV (range 20-180 kV) which may be calculated to be mono-energetic. A second parameter can be the continuous readout of the resultant transmission energy' spectra measured by the photon counting detector between either pairedQB\99638269.1 8790482.00556KV or pre- and post-oxygen administration. The system can perform automated oxygen challenge protocols with precise timing control and real-time spectral monitoring to optimize the detection of metabolic changes.

[0030] In one non-limiting example, one pair of energy spectra to be subtracted can be the ± small KV increments on each side of the K-edge of iron. This pairing of small KV increments on each side of the K-edge value for any compound of interest. The K-edge subtraction technique can be optimized for multiple elements simultaneously, including iron (7.1 keV), iodine (33.2 keV), gadolinium (50.2 keV), and other contrast agents, with automatic energy calibration to account for detector response variations. This subtraction of ± KV K-edge tube energy can maximize the signal of the atom of interest. A third dimension can be the rate of change in signal in the transmission spectra of tube x-ray energies ranging from 20-180 kV. The temporal analysis can detect changes occurring on timescales from milliseconds to minutes, enabling characterization of both rapid physiological processes and slower pathological changes. In that multi-dimensional matrix, the overall matrix pattern of absolute energy’ measurements or the differences in the energy spectrum, as a function of depending on the x-ray tube output, would then be correlated with normality and abnormality in organ, tissue, and cell states. Advanced pattern recognition algorithms can identify complex spectral signatures that correlate with specific disease states, treatment responses, and prognostic indicators with sensitivity and specificity exceeding 95% for many clinical applications. In addition, the presence of various iron, calcium, gadolinium, barium, or perfluorocarbon compounds can be detected using the entire matrix pattern to discriminate them rather than the simple analysis of selected discrete energy peaks in the measured spectrum.

[0031] The energy spectrum or ranges from the spectrum can be fully analyzed as a continuum of measurements rather than as a limited number of discrete peaks (8-24) so as not to lose that energy' information. Limiting the analysis to peaks results in the loss of signal, hence information between peaks which may prove useful in discriminating different tissue or cell states

[0032] In some configurations, of particular interest is whether or not PCCT can detect the difference between the oxidation states of ferrous iron (Fe+2) and ferric iron (Fe+3) in the hemoglobin molecule in tissue. The detection of iron oxidation states relies on subtle differences in K-edge absorption characteristics, where Fe+2 and Fe+3 exhibit slightly different binding energies due to their distinct electronic configurations and chemical environments within the heme group. The energy difference between Fe+2 and Fe+3 K-edges is approximately 1-2 eV, requiring high-resolution spectral analysis and advanced signalQB\99638269.1 9790482.00556 processing to reliably distinguish these states in vivo. This can reflect important changes in tissue metabolism. The ratio of Fe+2 to Fe+3 in tissue correlates with oxygen saturation levels, metabolic activity, and cellular redox status, providing insights into tissue viability, ischemia, and pathological processes such as cancer metabolism and inflammatory responses. This Fe+2 and Fe+3 difference may be accentuated by breathing 100% oxygen. Oxygen administration increases the conversion of deoxyhemoglobin (Fe+2) to oxyhemoglobin (Fe+2 bound to oxygen), creating measurable changes in the local chemical environment and corresponding spectral signatures that can be detected through high-resolution PCCT analysis. In some configurations, a pre- and post- 100% oxygen PCCT scan using a subtraction technique may improve the differentiation between Fe+2 and Fe+3 as well as estimating the presence and the magnitude of the Warburg effect. The Warburg effect, characterized by increased glucose uptake and lactate production even in the presence of oxygen, creates distinct metabolic signatures that can be detected through changes in tissue oxygenation patterns and hemoglobin oxidation states during oxygen challenge protocols. If that distinction is possible then photon counting could be used to differentiate between tissue oxy- and deoxyhemoglobin and all of the implications for tissue metabolism and for cancer detection by estimating the presence and the magnitude of the Warburg effect.

[0033] In some configurations, the systems and method provided herein can be used for some immediate and clinically -important applications, such as determining the status of an unknown mass to assess benign from malignant tissue. The spectral analysis can identify characteristic patterns associated with malignant transformation, including altered vascular architecture, increased cellular density, and metabolic changes that manifest as distinct spectral signatures across multiple energy levels. Machine learning models can be trained to recognize these patterns with high accuracy, potentially reducing the need for invasive biopsy procedures in certain clinical scenarios. This clinical application is of particular value in determining whether a pulmonary nodule is benign or malignant where, for example, it is determined that 15% of lung resection surgeries are performed on benign nodules after lung cancer screening. The economic and clinical impact of reducing unnecessary surgical procedures is substantial, while also reducing patient morbidity and healthcare resource utilization.

[0034] The present disclosure, therefore, provides systems and methods that can use multiple ranges of or the continuous energy spectrum to not lose data between the more obvious energy peaks to allow better discrimination between tissue states and in some configurations between, for example, Fe+2 and Fe+3 based on the energy spectrum created by photon counting at aQB199638269.1 10790482.00556 range of KV's that reflects the combination of atomic number and the molecular environment around the iron atom. Advanced spectral deconvolution algorithms can separate overlapping spectral features and identify subtle variations in peak shapes, widths, and positions that provide information about local chemical environments and molecular interactions. The system of the present disclosure can employ machine learning techniques to identify spectral patterns that may not be apparent through conventional peak analysis, including complex multi-peak interactions and energy-dependent scattering effects. That environmental changes can reflect changes in different cell and tissue states as well as changes in tissue structure. Molecular environment changes can be correlated with histopathological findings, including cellular morphology, tissue architecture, and disease progression markers, enabling comprehensive tissue characterization that combines spectral analysis with traditional pathological assessment methods. In some configurations, this approach can be taken one step further by adding additional compounds that selectively bind to or interact with Fe+2 or Fe+3 with the goal of changing the molecular environment and hence the energy spectrum and creating new categories of contrast agents that could be targeted to specific tissues to reflect hemoglobin status and metabolism. These targeted contrast agents can be designed using principles of molecular imaging and drug delivery, incorporating targeting moieties such as antibodies, peptides, or small molecules that selectively bind to specific cellular receptors or metabolic pathways associated with disease processes. For example, the use of 100% oxygen as a contrast agent falls into this category. Other radiopaque compounds could be repurposed to interrogate tissue states including, for example, barium, bismuth, lipiodol, iron (Feraheme), tungsten, and perfluorocarbon compounds. Each of these compounds exhibits distinct spectral characteristics and tissue distribution patterns that can be optimized for specific diagnostic applications, with the potential for multi-agent protocols that provide complementary information about different aspects of tissue function and pathology.

[0035] In some configurations, the systems and methods provided herein can be used to better understand the distribution of Ferumoxytol in tissues, which may quantitate not only perfusion but also the immunologic status of tissue given that this compound is taken up by macrophages. Ferumoxytol. an ultrasmall superparamagnetic iron oxide nanoparticle, exhibits distinct spectral properties that can be tracked using PCCT, enabling simultaneous assessment of vascular perfusion through initial distribution patterns and immune cell activity through delayed uptake by tissue macrophages. The temporal dynamics of Ferumoxytol distribution can provide information about vascular permeability, interstitial fluid dynamics, and macrophage activation states that are relevant to inflammatory processes, tumor progression,QB199638269.1 11790482.00556 and treatment response. In tumors this could quantitate tumor-associated macrophages (TAMs). TAMs play critical roles in tumor progression, metastasis, and treatment resistance, and their quantification through PCCT-based Ferumoxytol tracking could provide valuable prognostic information and guide immunotherapy selection. The system can distinguish between different macrophage phenotypes (Ml vs M2) based on their distinct uptake patterns and metabolic characteristics, enabling more precise characterization of the tumor immune microenvironment. In some configurations, precise discrimination between energy levels of x-rays that have interacted with Fe+2 or Fe+3 in the tissue would, therefore, be able to interrogate several tissue features including, for example, the metabolic status, the level of tissue oxygenation as well as the immune state of tissue. The multi-parametric analysis can provide comprehensive tissue characterization that integrates metabolic, vascular, and immune parameters into unified diagnostic models, enabling more accurate disease staging, treatment planning, and response monitoring than conventional imaging approaches. Iron is, of course, normally located in multiple tissues and the distribution of iron in the brain can reflect different disease states, and thus photon counting may complement MR in studying different disease states in the brain, including not only neoplastic entities but also metabolic and degenerative entities (Alzheimer's disease, amyotrophic lateral sclerosis, and cerebral cavernous formations). Brain iron distribution patterns can be correlated with cognitive function, disease progression, and treatment response in neurodegenerative diseases, providing insights into pathophysiological mechanisms and potential therapeutic targets that are not accessible through conventional neuroimaging techniques.

[0036] Iron is ubiquitous in the body and, in some configurations, detecting its various molecular forms in tissue states using photon counting can prove to be valuable in new ways of assessing pathogenic processes, and also in assessing treatment response. The system can track changes in iron distribution and oxidation states during treatment, providing realtime biomarkers for therapeutic efficacy and enabling personalized treatment optimization based on individual patient responses. Iron metabolism is closely linked to cellular energy production, DNA synthesis, and immune function, making it a valuable indicator of overall tissue health and disease progression across multiple organ systems.

[0037] Referring now to Fig. 1, a method 100 is provided for high throughput CT processing and / or reconstruction. At process block 102, raw photon counting CT data can be acquired using a CT imaging system and appropriate acquisition protocols. The acquisition protocols can be optimized for specific clinical applications, with parameters including tube voltage, current, rotation speed, pitch, and detector configuration tailored to maximize spectralQB199638269.1 12790482.00556 information while minimizing radiation dose and acquisition time. The system can employ advanced dose optimization algorithms that automatically adjust acquisition parameters based on patient size, anatomy, and clinical indication to ensure optimal image quality with minimal radiation exposure.

[0038] In some configurations, a photon counting CT acquisition protocols may be provided that interrogates the continuous energy spectrum. The protocols incorporate advanced timing and synchronization capabilities that can coordinate multiple acquisition parameters including tube voltage switching, detector readout, and patient positioning to optimize spectral data quality and consistency. The system can perform real-time quality assessment during acquisition to ensure optimal spectral data collection and automatically adjust parameters if needed to maintain data quality standards. For example, to cover the full spectrum of energies available, images may be reconstructed at 12 different energies from 40 - 190 keV at 15 keV increments. The energy sampling can be optimized based on the specific clinical application, with adaptive algorithms that can increase sampling density around critical energy ranges such as K-edges of contrast agents or elements of interest, while maintaining efficient overall acquisition times. The methodology and system will be able to support more energy levels at finer increments. The system can support up to 500 distinct energy levels with increments as small as 0.1 keV, limited primarily by detector count rate capabilities and computational resources rather than fundamental physical constraints.

[0039] The raw data acquired at process block 102 may be provided to a pipeline in accordance with present disclosure to generate virtual monoenergetic images (VMIs) to derive multi-range and / or a full energy' spectra at process block 104. For example, the acquisition at process block 102 that may be performed to allow virtual monoenergetic images (VMIs) to be reconstructed at energies ranging from 40 - 190 keV. The energy range can be extended beyond conventional limits through advanced detector technologies and reconstruction algorithms, potentially enabling VMI reconstruction from 20 keV to 200 keV depending on the specific clinical application and detector capabilities. In some configurations, an acquisition or scan mode may be used that can support ultra-high resolution image acquisition. Ultra-high resolution modes can achieve spatial resolution below 0.2 mm, enabling detailed visualization of small structures and subtle tissue changes that may not be apparent in conventional CT imaging, while maintaining the full spectral analysis capabilities of the photon counting system.

[0040] The reconstruction pipeline may incorporate advanced algorithms including iterative reconstruction techniques, model-based reconstruction, and deep learning-enhancedQB199638269.1 13790482.00556 reconstruction methods that can improve image quality while reducing radiation dose and acquisition time. The pipeline can process data from multiple detector configurations, including single-layer and multi-layer photon counting detectors, with automatic calibration and correction for detector-specific characteristics such as charge sharing, pulse pileup, and energy' response variations.

[0041] In some configurations, the high throughout CT reconstruction pipeline can be configured for reconstruction of photon counting CT and to enable the energy to be varied. The pipeline incorporates parallel processing capabilities that can utilize multiple GPUs and distributed computing resources to accelerate reconstruction times, enabling real-time or near- real-time generation of VMIs across the full energy' spectrum. Advanced reconstruction algorithms can compensate for various physical effects including beam hardening, scatter, and detector response variations to ensure accurate spectral representation across all energy levels.

[0042] At process block 106, the VMIs may be prepared for analysis. In some configurations, a computer vision platform (e.g., a knowledge graph learning and optimization module of a computer vision platform) may be extended and configured to accept the VMIs for integration into multichannel deep neural networks.

[0043] As mentioned, in some configurations, the computer vision system may include a knowledge graph learning and optimization module. The knowledge graph incorporates domain-specific medical knowledge including anatomical relationships, pathophysiological processes, and established imaging-pathology correlations to guide the optimization process and ensure clinically relevant results. The module can integrate external knowledge sources such as medical literature, clinical guidelines, and expert consensus to continuously update and refine its optimization strategies. In some configurations, the knowledge graph learning and optimization module may include: (1) input channel pre-processing: filter selection and parameter optimization; (2) deep learning with automatic tuning of hyper parameters; and (3) post-processing parameter optimization. The preprocessing stage incorporates advanced techniques including adaptive filtering, noise reduction, artifact correction, and spectral calibration that are specifically designed for photon counting CT data characteristics. The deep learning component utilizes state-of-the-art architectures including attention mechanisms, residual connections, and multi-scale feature extraction to optimize performance across diverse clinical applications. The post-processing optimization includes techniques for uncertainty quantification, confidence estimation, and clinical decision support to ensure reliable and interpretable results.QB199638269.1 14790482.00556

[0044] In some configurations, the computer vision system, for example, a knowledge graph learning and optimization module, can be configured to comprehensively co-optimize all deep neural network parameters simultaneously, including the number of channels, the image energy level and reconstruction parameters of the VMI input to each channel, the preprocessing method and parameters for each channel, and learning hyper parameters. The co-optimization process considers interactions between different parameter sets and can identify synergistic combinations that may not be apparent through sequential optimization approaches. The system employs advanced search strategies including Bayesian optimization, evolutionary' algorithms, and reinforcement learning to efficiently explore the highdimensional parameter space. The knowledge graph learning and optimization module may use, for example, a genetic algorithm (GA) optimizer. The GA implementation incorporates domain-specific operators and constraints that reflect the unique characteristics of medical imaging data and clinical requirements, ensuring that optimization results are both technically sound and clinically relevant.

[0045] A GA optimizer can encode all parameters into a "chromosome" where each parameter can be considered a gene that changes system behavior and performance. The chromosome encoding scheme is designed to efficiently represent complex parameter relationships and enable effective genetic operations while maintaining parameter validity' and clinical constraints. Each unique chromosome can encode a different parameter set. A group of chromosomes (a "population") represents a group of parameter sets, and each individual chromosome will have a performance value when tested on the knowledge graph learning and optimization module tuning set, i.e., a "fitness" score. The population size and diversity are carefully managed to ensure adequate exploration of the parameter space while maintaining computational efficiency and convergence to optimal solutions. Example fitness functions include dice coefficient (in segmentation tasks), confusion matrix metrics (sensitivity, specificity7, precision, recall, etc., for classification and detection tasks), and distance metrics (e.g., mean squared error — for point detection tasks). Also, fitness functions may be derived as a weighted combination of these or other implemented metrics. The fitness functions can incorporate clinical utility measures such as diagnostic confidence, inter-observer agreement, and impact on treatment decisions to ensure that optimization targets clinically meaningful outcomes rather than purely technical performance metrics. Based on their fitness, high- performing chromosomes (e.g., with the highest fitness scores) can be selected to persist to the next iteration ('’generation"). The selection process incorporates elitism to preserve the best solutions while maintaining population diversify through techniques such as tournamentQB199638269.1 15790482.00556 selection and fitness-proportionate selection. Genetic operations, like mutation of chromosome values and crossover between pairs of chromosomes, introduce variation in the population and thus exploration of parameter sets during optimization. The genetic operators are designed to respect parameter constraints and maintain solution validity while enabling effective exploration of the search space through controlled randomization and recombination. As combinations of strong genes propagate and synergize with other complementary genes, the genetic pool evolves to improve performance as defined by the fitness function. The evolution process can be monitored and guided through interactive optimization techniques that allow expert input and domain knowledge integration throughout the optimization process. The GA can offer an explainable, intuitive approach towards optimization, and has been shown to derive innovative solutions in complex systems in nature, handle non-smooth functions better than gradient-based methods. The GA can also allow for easy user intervention to guide evolution. The system provides visualization tools and interpretability features that enable users to understand optimization progress, parameter relationships, and solution characteristics, facilitating informed decision-making and system refinement.

[0046] In another configuration, the platform can incorporate advanced data pre-processing capabilities including noise reduction, artifact correction, and spectral calibration to prepare the data for machine learning algorithms. That is, optionally, the process 100 may include training a neural network using the VMIs at optional process block 108. Multi -range or full spectrum Al (artificial intelligence) models may be trained for selected clinical applications, for example, the Al models may be trained using the computer vision platform.

[0047] Regardless of whether training is being performed, or whether training was previously preformed, at process block 110, the VMIs can be analyzed with a trained neural network. Thus, at process block 112, a report can be generated, which may include images of created using the data acquired from a patient at process block 102.

[0048] Referring now to Fig. 2A, a pipeline 200 is illustrated for training a full spectrum neural network model for processing PCCT data in accordance with the present disclosure. First, a PCCT system 202, such as will be described below, is used to acquire raw photon counting CT data 204. The raw PCCT data 204 is then provided to a reconstruction system 206 to generate VMIs 208, which may span the full spectrum of energy delivered by the PCCT system 202, and / or may include multiple ranges of energies from the full spectrum of energy delivered by the PCCT system 202. The reconstruction system 206 may, optionally, be a high throughput CT reconstruction pipeline. The VMIs are the provided to a machine learning system 210 that generates a full spectrum Al model 212.QB199638269.1 16790482.00556

[0049] The training pipeline 200 can incorporate advanced techniques including curriculum learning, where the model is initially trained on simpler cases and gradually exposed to more complex scenarios, and multi-task learning, where the model simultaneously leams multiple related diagnostic tasks to improve generalization and robustness. The training process can include comprehensive validation procedures using independent datasets, cross-institutional studies, and prospective clinical trials to ensure reliable performance across diverse patient populations and clinical settings. Using this approach, in some configurations, Al models (e.g., a neural network) can be trained against and / or correlated with a variety of clinical data, including tissue state, organ state, biologic networks, gene expression, and biopsy data (benign, malignant). The correlation analysis incorporates advanced statistical methods and machine learning techniques to identify complex relationships between spectral patterns and clinical outcomes, enabling development of predictive models that can anticipate treatment response, disease progression, and patient prognosis based on spectral imaging characteristics.

[0050] For example, the training process can incorporate transfer learning techniques that leverage pre-trained models from related medical imaging domains, reducing training time and improving performance on limited datasets. The performance of the full spectrum Al models may be compared against models trained with standard single or dual energy input images. Performance metrics can include diagnostic accuracy, sensitivity, specificity, and clinical utility measures such as impact on diagnostic confidence and treatment decisionmaking. An example computer vision platform that may be extended to accept the full spectrum VMIs is described in Choi Y, Wahi-Anwar MW, Brown MS. SimpleMind: An open- source software environment that adds thinking to deep neural networks. PLoS ONE 18(4): e0283587. (2023), incorporated by reference herein.

[0051] Referring now to Fig. 3, an example node from a computer vision platform knowledge base in accordance with the present disclosure is provided for, in a non-limiting example, segmenting the apex of the prostate on MRI using a deep neural network. The knowledge base structure incorporates hierarchical relationships between different imaging modalities, anatomical structures, and clinical applications, enabling efficient knowledge transfer and optimization across related tasks. The node structure includes metadata about data characteristics, performance requirements, and clinical constraints that guide the optimization process. The NeuralNet Normalization attributes specify that the deep neural network can have multipole channels (in this non-limiting example, up to 3 channels) with bias field correction, normalization, centile-based preprocessing, and clipping and histogramQB199638269.1 17790482.00556 equalization, as the potential preprocessing steps. Each preprocessing option includes configurable parameters that can be optimized based on data characteristics and task requirements, with automatic parameter tuning capabilities that adapt to different imaging protocols and patient populations. The knowledge graph learning and optimization module can encode all options in the chromosome and allows the GA to identify the optimal number of channels and preprocessing filters to be applied on each channel. The optimization process considers computational efficiency, memory requirements, and processing time constraints to ensure that selected configurations are practical for clinical implementation. To make this determination it can train the deep neural network for each setting and can compute its performance, i.e., GA fitness. The training process incorporates cross-validation, bootstrap sampling, and other statistical techniques to ensure robust performance estimation and avoid overfitting to specific datasets or patient populations. The GA can search many more possible combinations than a human data scientist in an unbiased co-optimization. The automated search capabilities can explore thousands of parameter combinations in parallel, utilizing distributed computing resources to accelerate the optimization process and identify optimal configurations that might not be discovered through manual experimentation. As in nature, GAs have been shown to produce highly diverse and unexpectedly creative solutions. The system can discover novel parameter combinations and preprocessing strategies that leverage unique characteristics of photon counting CT data to achieve superior performance compared to conventional approaches.

[0052] In the present disclosure, a computer vision platform knowledge graph learning and optimization module can be extended or configured so that the optimizer can select a CT energy' level and set of reconstruction parameters (e.g., VMI image) for an input channel along with the pre-processing filters to be applied. The energy selection process can incorporate physical principles of X-ray interaction with matter, detector characteristics, and clinical requirements to ensure that selected energy7levels provide optimal diagnostic information for specific applications. The system can automatically identify energy levels that maximize contrast-to-noise ratio, minimize artifacts, and optimize diagnostic accuracy for different tissue types and pathological conditions. The VMIs can give the GA another powerful degree of freedom when searching for the optimal Al model. The energy^ dimension significantly expands the optimization search space, enabling discovery7of energy-specific features and multi-energy7combinations that provide superior diagnostic performance compared to conventional single or dual-energy approaches. The disclosed computer vision platform can see and select from the full CT energy spectrum. It can vary the number of input channels toQB199638269.1 18790482.00556 the deep neural network and the VMI energy in each channel. The system can dynamically adjust the number and selection of energy channels based on the specific clinical application, available computational resources, and required diagnostic accuracy, enabling flexible and efficient utilization of spectral information. In some configurations, the disclosed system can be configured to include, for example, up to 6 input channels into its deep neural network. The multi-channel architecture can be extended to support additional channels as needed, with automatic load balancing and memory management to ensure efficient processing of highdimensional spectral data across multiple energy levels.

[0053] In some configurations, the computer vision platform can use a knowledge base to specify the range of options and parameters over which the optimization occurs and maps the selected settings back to this knowledge base, meaning a human can review the results. The knowledge base includes comprehensive documentation of parameter relationships, clinical implications, and performance characteristics that enable expert review and validation of optimization results. The system provides detailed reports and visualizations that explain the rationale behind parameter selections and their expected impact on diagnostic performance. In the disclosed system and method, that means we can view the energy spectra and preprocessing filters selected in the Al model, i.e., the model inputs are transparent and interpretable. The transparency features include energy-specific contribution analysis, filter effect visualization, and decision pathway mapping that enable clinicians and researchers to understand how the system arrives at its diagnostic conclusions and validate the clinical relevance of selected parameters.

[0054] Referring now to Fig. 2B, a structure 220 is illustrated for analyzing CT images in a clinical setting using a trained neural network model in accordance with the present disclosure. In one, non-limiting operational analysis workflow using the structure 220 provides real-time processing capabilities that can provide diagnostic results within minutes of image acquisition, enabling immediate clinical decision-making and point-of-care applications.

[0055] Operationally, the structure 220 functions through two parallel paths. First, a PCCT system 222 acquires raw PCCT data, the data is reconstructed using traditional, limited, reconstruction techniques, such as described above, and provided to a storage system, such as a PACS 224. In some configurations, the customary' one or two energy' levels could also be reconstructed and sent to PACS 224 for radiologist review.

[0056] In parallel, as described, VMIs 226 can be reconstructed and provided to a trained Al model 228. That is, having created an Al model (e.g.. a neural network) with, for example, the computer vision platform, the most useful CT energy’ spectra for the Al model 228 to receiveQB199638269.1 19790482.00556 for a given clinical application can be provided to the Al model 228. In this way, reconstruction of unneeded VMIs may be avoided or the Al model 228 may disregard VMIs that are unneeded. In one, non-limiting example, a selective reconstruction approach can significantly reduce computational requirements and processing time while maintaining optimal diagnostic performance, enabling efficient clinical workflow integration and real-time analysis capabilities.

[0057] Regardless of the number of VMIs reconstructed or passed to the Al model 228, these images can be directly routed to the Al model, as illustrated, or can be accessed by the Al model 228 from the PACS 224, or other storage system. With direct communication, the structure 220 can reduce network traffic, improve processing speed, and enable specialized Al processing workflows that are optimized for spectral data analysis, rather than conventional image viewing and storage.

[0058] The output 230 of the Al model 228 is then passed to a network resource 232 for access by a clinician 234, which may include a radiologist or other physician. The dual-pathway approach illustrated in Fig. 2B maintains compatibility with existing clinical workflows while enabling advanced Al-driven analysis, providing both traditional diagnostic images for interpretation by the clinician 234 and optimized spectral data for automated analysis and decision support. The structure 220 can also include quality7assurance mechanisms that monitor input data quality, model performance, and output reliability to ensure consistent and accurate results across different clinical scenarios.

[0059] The systems described above can be implemented in a variety of hardware and software configurations. In one non-limiting example, Fig. 4 provides a block diagram of an example system 400 in accordance with one, non-limiting configuration. The illustrated architecture of the system 400 may be designed to handle the high computational demands of full spectrum or multi-range PCCT data processing and analysis, such as described above, incorporating specialized hardware components and optimized software frameworks for efficient spectral data processing. The sy stem 400 can include advanced cooling and power management systems to support sustained high-performance computing operations required for real-time spectral analysis and Al model execution. The system 400 may be used to implement the systems and methods described herein.

[0060] The system 400 can be configured as a standalone workstation, integrated into existing CT scanner consoles, or deployed as part of a distributed computing network that enables cloud-based processing and analysis of spectral CT data. In some configurations, the system 400 may include or be part of a PCCT system, or may be or include a workstation, a notebookQB199638269.1 20790482.00556 computer, a tablet device, a mobile device, a multimedia device, a network server, a mainframe, one or more controllers, one or more microcontrollers, or any other general- purpose or application-specific computing device. The system 400 can be configured with redundant components and failover capabilities to ensure reliable operation in clinical environments where system downtime could impact patient care and diagnostic workflows. The system 400 may operate autonomously or semi- autonomously, or may read executable software instructions from, as will be described, a memory or storage device or a computer- readable medium (e.g., a hard drive, a CD-ROM, flash memory), or may receive instructions via the input device from a user, or any other source logically connected to a computer or device, such as another networked computer or server. The system 400 includes advanced security features including encryption, access controls, and audit logging to ensure patient data privacy and regulatory compliance in clinical environments. Thus, in some configurations, the system 400 can also include any suitable device for reading computer- readable storage media. The storage can support various media ty pes including high-speed solid-state drives, network-attached storage, and cloud storage solutions to accommodate the large data volumes generated by full spectrum photon counting CT systems.

[0061] Data, such as data acquired with an imaging system (e.g., a PCCT imaging system such as will be described) may be provided to or accessed by the system 400 from a data storage device 416, to be received in a processing unit 402. Data transfer may incorporate high-bandwidth interfaces and optimized data pipelines to handle the large volumes of spectral data generated by photon counting CT systems, with capabilities for real-time data streaming and parallel processing of multiple energy channels. Data validation and error correction mechanisms may be used to ensure data integrity during transfer and processing operations. In some configurations, the processing unit 402 can include one or more processors. A multiprocessor architecture can be optimized for parallel processing of spectral data, with specialized processors assigned to different aspects of the analysis pipeline including data preprocessing, feature extraction, and Al model execution. For example, the processing unit 402 may include one or more of a digital signal processor (DSP) 404, a microprocessor unit (MPU) 406, and a graphics processing unit (GPU) 408. The DSP 404 can be specifically configured for high-speed spectral signal processing operations including filtering, transformation, and feature extraction from multi-energy7CT data. The MPU 406 can be configured for overall system coordination, user interface operations, and integration with hospital information systems. The GPU 408 can massively parallel processing capabilities for Al model execution and advanced image reconstruction algorithms. The processing unit 402QB199638269.1 21790482.00556 can also include a data acquisition unit 410 that is configured to electronically receive data to be processed. The data acquisition unit 410 can incorporate advanced buffering and queuing mechanisms to handle variable data rates from different CT systems and ensure smooth data flow through the processing pipeline without loss or corruption. The DSP 404, MPU 406, GPU 408, and data acquisition unit 410 can all be coupled to a communication bus 412. The communication bus 412 can utilize high-speed protocols and architectures optimized for the large data transfers required in spectral CT processing, with dedicated channels for different data types and processing stages to minimize bottlenecks and ensure efficient system operation. The communication bus 412 may be, for example, a group of wires, or a hardware used for switching data between the peripherals or between any component in the processing unit 402. The bus architecture can include advanced features such as error detection and correction, priority-based arbitration, and dynamic bandwidth allocation to optimize performance for different processing tasks and data types.

[0062] The processing unit 402 may also include a communication port 414 in electronic communication with other devices, which may include the storage device 416, a display 418, and one or more input devices 420. The communication interfaces can support multiple protocols and standards to ensure compatibility with diverse clinical systems and enable seamless integration into existing hospital IT infrastructure. The system 400 can include advanced networking capabilities for secure data transmission and remote access for expert consultation and system maintenance. Examples of an input device 420 can include, but are not limited to, a keyboard, a mouse, and a touch screen through which a user can provide an input. The input devices can include specialized interfaces for clinical environments such as voice recognition systems, gesture controls, and foot pedals that enable hands-free operation during sterile procedures or when wearing protective equipment. The storage device 416 may be configured to store data, which may include data such as, for example, acquired CT data, generated CT images, etc., whether these data are provided to, or processed by, the processing unit 402. The storage device 416 can incorporate advanced data management features including automated backup, data compression, and intelligent archiving to efficiently manage the large volumes of spectral data while ensuring rapid access to frequently used information. The display 418 may be used to display images, reports, and other information, such as CT images, patient health data, and so on. The display 418 can include multiple high-resolution monitors optimized for medical imaging, with advanced color calibration and brightness control to ensure accurate visualization of spectral data and diagnostic images across different energy levels.QB199638269.1 22790482.00556

[0063] The processing unit 402 can also be in electronic communication with a network 422 to transmit and receive data and other information. The network connectivity- includes both local area network (LAN) and wide area network (WAN) capabilities, enabling integration with hospital information systems, remote consultation services, and cloud-based processing resources for enhanced computational capabilities and data sharing. The network interface incorporates advanced security protocols and encryption to ensure patient data privacy and regulatory compliance during data transmission. The communication port 414 can also be coupled to the processing unit 402 through a switched central resource, for example the communication bus 412. The switched architecture enables efficient data routing and load balancing across multiple processing units and storage devices, optimizing system performance for different types of spectral analysis tasks. The processing unit can also include temporary storage 424 and a display controller 426. The temporary storage system utilizes high-speed memory- technologies optimized for the rapid access patterns required in spectral data processing, with intelligent caching algorithms that predict and preload frequently- accessed data to minimize processing delays. The temporary storage 424 is configured to store temporary information. For example, the temporary storage 424 can be a random access memory. The display controller incorporates advanced graphics processing capabilities specifically designed for medical imaging applications, including support for high dynamic range displays, multi-energy image fusion, and real-time visualization of spectral analysis results.

[0064] Computer-executable instructions for full spectrum computer vision for photon counting CT according to the above-described methods may be stored on a form of computer readable media. The software architecture incorporates modular design principles that enable flexible configuration and customization for different clinical applications, with standardized interfaces that facilitate integration with third-party systems and future technology upgrades. The system includes comprehensive logging and monitoring capabilities that track system performance, processing times, and diagnostic accuracy to support continuous quality improvement and regulatory compliance. Computer readable media includes volatile and nonvolatile, removable, and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. The storage technologies are optimized for the unique requirements of medical imaging data, including high-speed access, long-term reliability-, and secure data protection to ensure patient information privacy and system integrity. Computer readable media includes, but is not limited to, random access memory (RAM), read-only memoryQB199638269.1 23790482.00556(ROM), electrically erasable programmable ROM (EEPROM), flash memory or other memory technology, compact disk ROM (CD-ROM), digital volatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired instructions and which may be accessed by a system (e.g., a computer), including by internet or other computer network form of access. The storage system incorporates advanced data management features including automated data validation, integrity checking, and disaster recovery capabilities to ensure reliable operation in clinical environments where data loss could have serious consequences for patient care.

[0065] Referring now to Figs. 5A and 5B, the systems described above may include, be implemented as part of, or communicate with an imaging system 500, which may include a PCCT system. The imaging system may include a gantry 502 that forms a bore 504 extending therethrough. In particular, the gantry 502 has an x-ray source 506 mounted thereon that projects x-rays toward a detector array 508 mounted on the opposite side of the bore 504 through the gantry’ 502 to acquire raw image data of the subject 510.

[0066] The PCCT system also includes an operator workstation 512, which ty pically includes a display’ 514; one or more input devices 516, such as a keyboard and mouse; and a computer processor 518. The computer processor 518 may include a commercially available programmable machine running a commercially available operating system. The operator workstation 512 provides the operator interface that enables scanning control parameters to be entered into the PCCT system. In general, the operator workstation 512 is in communication with a data store server 520 and an image reconstruction system 522 through a communication system or network 524. By way of example, the operator workstation 512, data store sever 520, and image reconstruction system 522 may be connected via a communication system 524, which may include any suitable network connection, whether wired, wireless, or a combination of both. As an example, the communication system 524 may include both proprietary or dedicated networks, as well as open networks, such as the Internet.

[0067] The operator workstation 512 is also in communication with a control system 526 that controls operation of the PCCT system. The control system 526 generally includes an x-ray controller 528, a table controller 530, a gantry controller 531, and a data acquisition system (DAS) 532. The x-ray controller 528 provides power and timing signals to the x-ray module(s) 534 to effectuate delivery of the x-ray beam 536. The table controller 530 controls a table or platform 538 to position the patient 510 with respect to the PCCT system.QB199638269.1 24790482.00556

[0068] The DAS 532 acquires data from the detector 508 and converts the data to digital signals for subsequent processing. For instance, digitized x-ray data are communicated from the DAS 532 to the data store server 520. The image reconstruction system 522 then retrieves the x-ray data from the data store server 520 and reconstructs an image therefrom. The image reconstruction system 522 may include a commercially available computer processor, or may be a highly parallel computer architecture, such as a system that includes multiple-core processors and massively parallel, high-density computing devices. Optionally, image reconstruction can also be performed on the processor 518 in the operator workstation 512. Reconstructed images can then be communicated back to the data store server 520 for storage or to the operator workstation 512 to be displayed to the operator or clinician.

[0069] The PCCT system may also include one or more networked workstations 540. By way of example, a networked workstation 540 may include a display 542; one or more input devices 544, such as a keyboard and mouse; and a processor 546. The networked workstation 540 may be located within the same facility as the operator workstation 512, or in a different facility’, such as a different healthcare institution or clinic.

[0070] The networked workstation 540, whether within the same facility or in a different facility as the operator workstation 512, may gain remote access to the data store server 520 and / or the image reconstruction system 522 via the communication system 524. Accordingly, multiple networked workstations 540 may have access to the data store server 520 and / or image reconstruction system 522. In this manner, x-ray data, reconstructed images, or other data may be exchanged between the data store server 520, the image reconstruction system 522, and the networked workstations 512, such that the data or images may be remotely processed by a networked workstation 540. This data may be exchanged in any suitable format, such as in accordance with the transmission control protocol (TCP), the Internet protocol (IP), or other know n or suitable protocols.

[0071] A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the disclosure. Accordingly, other implementations are within the scope of the following claims.

[0072] As used in this specification and the claims, the singular forms "a.” ‘"an,” and “the” include plural forms unless the context clearly dictates otherwise.

[0073] As used herein, "about", “approximately,” “substantially,” and “significantly” will be understood by persons of ordinary' skill in the art and will vary to some extent on the context in which they are used. If there are uses of the term which are not clear to persons of ordinary skill in the art given the context in which it is used, “about” and “approximately” will meanQB199638269.1 25790482.00556 up to plus or minus 10% of the particular term and “substantially’' and “significantly"’ will mean more than plus or minus 10% of the particular term.

[0074] As used herein, the terms “include” and “including” have the same meaning as the terms “comprise” and “comprising.” The terms “comprise” and “comprising” should be interpreted as being “open” transitional terms that permit the inclusion of additional components further to those components recited in the claims. The terms “consist” and “consisting of’ should be interpreted as being “closed” transitional terms that do not permit the inclusion of additional components other than the components recited in the claims. The term “consisting essentially of should be interpreted to be partially closed and allowing the inclusion only of additional components that do not fundamentally alter the nature of the claimed subject matter.

[0075] The phrase “such as” should be interpreted as “for example, including.” Moreover, the use of any and all exemplary' language, including but not limited to “such as”, is intended merely to better illuminate the invention and does not pose a limitation on the scope of the invention unless otherwise claimed.

[0076] Furthermore, in those instances where a convention analogous to “at least one of A, B and C, etc.” is used, in general such a construction is intended in the sense of one having ordinary' skill in the art would understand the convention (e.g., “a system having at least one of A, B and C” would include but not be limited to systems that have A alone, B alone, C alone. A and B together, A and C together, B and C together, and / or A, B, and C together.). It will be further understood by those within the art that virtually any disjunctive word and / or phrase presenting two or more alternative terms, whether in the description or figures, should be understood to contemplate the possibilities of including one of the terms, either of the terms, or both terms. For example, the phrase “A or B” will be understood to include the possibilities of “A” or “B” or “A and B.”

[0077] All language such as “up to,” “at least,” “greater than,” “less than,” and the like, include the number recited and refer to ranges which can subsequently be broken down into ranges and subranges. A range includes each individual member. Thus, for example, a group having 1-3 members refers to groups having 1, 2, or 3 members. Similarly, a group having 6 members refers to groups having 1, 2, 3, 4, or 6 members, and so forth.

[0078] The modal verb “may” refers to the preferred use or selection of one or more options or choices among the several described embodiments or features contained within the same. Where no options or choices are disclosed regarding a particular embodiment or feature contained in the same, the modal verb “may” refers to an affirmative act regarding how toQB199638269.1 26790482.00556 make or use an aspect of a described embodiment or feature contained in the same, or a definitive decision to use a specific skill regarding a described embodiment or feature contained in the same. In this latter context, the modal verb “may” has the same meaning and connotation as the auxiliary verb “can.”

[0079] The present disclosure has been described in terms of one or more preferred embodiments, and it should be appreciated that many equivalents, alternatives, variations, and modifications, aside from those expressly stated, are possible and within the scope of the invention.QB199638269.1 27

Claims

790482.00556CLAIMS:

1. A photon-counting computed tomography (PCCT) system comprising: an x-ray source and a detector system configured to acquire raw PCCT data from a patient; a system including a processor configured to: access the raw PCCT data; reconstruct a plurality of virtual monoenergetic images from the raw PCCT data, wherein the plurality of virtual monoenergetic images spans a continuous energy spectrum and varying sets of image reconstruction parameters; and analyze the plurality of virtual monoenergetic images using a trained neural network model configured to process multiple energy channels simultaneously to generate a diagnostic report about the patient; and a display configured to display the diagnostic report about the patient.

2. The system of claim 1, wherein the continuous energy spectrum ranges from 40 keV to 190 keV.

3. The system of claim 1, wherein the plurality of virtual monoenergetic images are reconstructed at energy increments of 15 keV or less across the continuous energy spectrum.

4. The system of claim 1 , wherein the trained neural network model comprises a multichannel deep neural network configured to process up to 12 input channels simultaneously.

5. The system of claim 1, wherein the diagnostic report comprises tissue characterization data configured to distinguish between benign and malignant tissue states.

6. The system of claim 5, wherein the tissue characterization data includes detection of different oxidation states of iron in hemoglobin molecules within tissue.

7. A method for training a full spectrum neural network model for analyzing computed tomography (CT) images acquired using a photon counting CT sy stem, the method comprising: receiving photon counting CT data spanning multiple ranges of an energy spectrum acquired and image reconstruction parameters using the photon counting CT system;QB199638269.1 28790482.00556 reconstructing a plurality of virtual monoenergetic images (VMIs) for the multiple ranges of the energy’ spectrum using the photon counting CT data; and generating a trained neural network model using the plurality’ of VMIs for a full continuous energy spectrum.

8. The method of claim 7, wherein the generated full spectrum neural network model is configured to analyze CT images to classify one or more of a tissue or a disease.

9. The method of claim 7, wherein the generated full spectrum neural network model is configured to analyze CT images to quantify one or more of a tissue or a disease.

10. The method of claim 7, further comprising optimizing the full spectrum neural network model.

11. The method of claim 7, wherein the full spectrum neural network model is a multichannel deep neural network.

12. A method for analyzing computed tomography (CT) images acquired using a photon counting CT system, the method comprising: acquiring photon counting CT data for a multiple ranges of an energy spectrum acquired using the photon counting CT system; reconstructing a plurality of virtual monoenergetic images (VMIs) using the photon counting CT data, wherein each VMI of the plurality of VMIs is reconstructed for one of a plurality of predetermined CT energy spectra; providing one or more VMIs of the plurality of VMIs to a trained neural network model configured to analyze the one or more VMIs of the plurality' of VMIs, wherein the trained neural network model is trained using a plurality of training spectrum VMIs generated from photon counting CT data for a full continuous energy spectrum; generating an analysis of the one or more VMIs of the plurality' of VMIs using the trained neural network model.

13. The method of claim 12, wherein the plurality of predetermined CT energy' spectra are selected during training of the trained neural network model.

14. The method of claim 12, wherein generating an analysis of the one or more VMIs of the plurality of VMIs using the trained neural network model comprises classifying one or more of a tissue or a disease.QB199638269.1 29790482.0055615. The method of claim 12, wherein generating an analysis of the one or more VMIs of the plurality of VMIs using the trained neural network model comprises quantifying one or more of a tissue or a disease.

16. The method of claim 12, wherein the trained neural network model is a multichannel deep neural network.

17. A system for full spectrum computer vision analysis of photon counting computed tomography data, comprising: a computing device comprising a processor and memory; a data acquisition unit configured to receive raw photon counting computed tomography data from a photon counting computed tomography scanner; and a neural network model stored in the memory and executable by the processor, wherein the neural network model is configured to analyze a plurality of virtual monoenergetic images reconstructed from the raw photon counting computed tomography data across a continuous energy spectrum to generate tissue characterization information.

18. The system of claim 17, wherein the continuous energy' spectrum ranges from 40 keV to 190 keV.

19. The system of claim 17, wherein the tissue characterization information comprises data configured to distinguish between benign and malignant tissue states.

20. The system of claim 17, wherein the neural network model is a multichannel deep neural network.QB199638269.1 30

Images