Systems and methods for image based condition remediation

Abstract

Systems and methods for imaging include decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono‑energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing.

Description

[0001] SYSTEMS AND METHODS FOR IMAGE BASED CONDITION REMEDIATION GOVERNMENT RIGHTS STATEMENT This invention was made with government support under CA206171 awarded by the National Institutes of Health. The Unite government has certain rights in the invention. CROSS REFERENCE TO RELATED APPLICATIONS This application claims the benefit of priority to United States Provisional Application, serial number 63 / 708,566, filed on October 17, 2024, and titled IMAGE BASED CONDITION REMEDIATION and also claims the benefit of priority to United States Provisional Application, serial number 63 / 763,483, filed on February 26, 2025 and titled IMAGE BASED CONDITION REMEDIATION. The disclosures of such applications are hereby incorporated by reference in their entireties. FIELD OF THE DISCLOSURE Embodiments herein relate to imaging in general and specifically image-based condition remediation. BACKGROUND Computed Tomography (CT), also referred to as a CAT scan, is an advanced medical imaging modality that utilizes X-rays in conjunction with sophisticated computer algorithms to generate detailed cross-sectional images of the body. In contrast to conventional X-rays, which provide only two- dimensional representations, CT scans enable more accurate visualization of internal structures, including organs, bones, soft tissues, and blood vessels, making them indispensable for diagnosing conditions such as tumors, internal injuries, and infections. Magnetic Resonance Imaging (MRI) complements CT by producing high-resolution images of soft tissues, employing powerful magnetic fields and radiofrequency waves rather than ionizing radiation. MRI is especially valuable for imaging the brain, spinal cord, muscles, and joints, allowing for precise evaluation of complex soft tissue abnormalities. Ultrasound, or sonography, utilizes high-frequency sound waves to create images of internal soft tissues and is widely used in obstetrics, cardiac evaluations, and abdominal assessments due to its non-invasive nature. Positron Emission Tomography (PET) is another sophisticated imaging technique, which involves the administration of a radioactive tracer to detect metabolic activity within tissues, making it a crucial tool for identifying malignancies, cardiovascular conditions, and neurological disorders. PET is frequently combined with CT (PET / CT) to enhance diagnostic accuracy. Similarly, Single Photon Emission Computed Tomography (SPECT) uses radioactive tracers, focusing on blood flow and functional activity in organs, with applications in cardiology and neurology. Fluoroscopy, which provides real-time X-ray imaging, is instrumental in allowing clinicians to observe internal structures dynamically, particularly during interventional procedures such as catheter placements, barium studies, and orthopedic surgeries. Mammography, a specialized form of X-ray imaging designed for breast tissue, plays a pivotal role in the early detection of breast cancer. Finally, traditional X-rays remain a foundational imaging technique, particularly suited for visualizing bone structures and detecting fractures, pulmonary conditions, and certain malignancies. Each of these imaging technologies offers distinct advantages, providing healthcare professionals with critical insights necessary for the diagnosis and management of a wide array of medical conditions. Computed Tomography (CT) has various specialized forms tailored for specific diagnostic purposes. Conventional CT provides general cross-sectional imaging, while Helical (Spiral) CT enables faster scans with detailed imaging, especially in trauma cases. High-Resolution CT (HRCT) focuses on detailed lung and chest images, and Multidetector (MDCT) or Multislice CT (MSCT) captures multiple slices rapidly for detailed 3D reconstructions. Dual-Energy CT (DECT) distinguishes different tissues using two X-ray energy levels, and CT Angiography (CTA) visualizes blood vessels with contrast agents. Cardiac CT specializes in heart and coronary imaging, CT Perfusion measures blood flow in tissues, and PET / CT combines functional imaging with detailed CT scans to assess metabolic and structural information together. Cancer ranks as the second leading cause of death among adults. Early detection of specific cancers can enable treatment at earlier stages and has been shown to significantly reduce mortality rate. Currently there are three major imaging-based cancer screening protocols: lung, colorectal and breast. In one of the largest clinical trials ever performed, low dose computed tomography (LdCT) screening for early detection of lung cancer was shown to decrease mortality rate by over 24%. This clinical trial paved the way for nation-wide screening for lung cancer in the United States with other countries following suit. In this clinical trial, the high sensitivity for detection of focal lesions or pulmonary nodules (PNs) is often at the cost of high false positive (FP) rate of over 96%, meaning many screening- detected PNs are non-cancerous or benign. The high FP rate is closely related to high prevalence of PNs and the widely recognized lesion heterogeneity, which reflects the footprint of each individual lesion evolutional and surrounding micro environmental processes. While some FPs can be clarified through medical expert interpretation on the acquired LdCT images and consideration of related patient medical records, a substantial portion remain ambiguous and are categorized as indeterminate lesions (IDLs) by the experts. For example, a clinical lesion risk model, recently developed by medical experts based on the LdCT screening PN images and patient medical records, has reached AUC (area under the curve of receiver operating characteristics) in the range from 0.88 to 0.89 when the PN lesion pathology was used as the ground truth of malignancy. Further increasing AUC is challenging due to the IDLs. Resolving the ambiguity of IDLs to determine their malignancy or benign nature necessitates costly and variably risky follow-up evaluations, e.g., tissue biopsy. While the reduction in mortality rate of over 24% is a significant gain, it comes with substantial costs associated with diagnosing IDLs. Therefore, managing IDLs in current LdCT screening for early lung cancer becomes a pivotal aspect of the screening and underscores the issue of over-detection and under-diagnosis in medical imaging-enabled screening of early diseases. As a representation of artificial intelligence, machine learning (ML) has made significant strides in emulating the expertise of medical professionals in interpreting medical image characteristics to predict lesion malignancy. For example, using the National Lung Screening Trial (NLST) database, where up to four medical experts evaluated each LdCT screening-detected PN image as malignant or benign, current ML algorithms have achieved AUC of 0.97. This indicates an outstanding performance of ML emulating the clinical lesion risk model developed by medical experts. However, when actual PN pathological reports were used as malignant or benign, these ML algorithms reached AUC as high as 0.896, showing performance on par with that of medical experts but not exceeding it. In essence, despite their advancements, current ML algorithms have not improved the efficacy of early lung cancer screenings provided by LdCT and continue to face challenges in diagnosing IDLs like those encountered by medical experts. It is expected that this difficulty arises from the current data-driven ML algorithms' inability to integrate prior knowledge beyond the acquired image data they analyze. BRIEF DESCRIPTION Shortcomings of the prior art are overcome, and additional advantages are provided, through the presently disclosed embodiments. In one embodiment, a method can include, for example: decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing. In another aspect, a computer program product can be provided. The computer program product can include a computer readable storage medium readable by one or more processing circuit and storing instructions for execution by one or more processor for performing a method. The method can include, for example: decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing. In a further aspect, a system can be provided. The system can include, for example a memory. In addition, the system can include one or more processor in communication with the memory. Further, the system can include program instructions executable by the one or more processor via the memory to perform a method. The method can include, for example: decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing. Additional features are realized through the techniques set forth herein. Other embodiments and aspects, including but not limited to methods, computer program product and system, are described in detail herein and are considered a part of the claimed invention. BRIEF DESCRIPTION OF THE DRAWINGS Fig.1 depicts a system having a manger system, imaging devices, user equipment (UE) devices, data sources, and an application program interface (API) service endpoint according to one embodiment. Fig.2A-2B is a flowchart illustrating performance of a method according to one embodiment. Fig.3 depicts energy decomposition and image transformation according to one embodiment Fig.4A depicts a graph of four typical X-ray energy spectra used in current CT devices, with peak tube voltages of 80, 100, 120, and 140 kVp, according to one embodiment. Fig.4B depicts a CT system according to one embodiment. Fig.4C depicts a graph of unfiltered and filtered X-ray spectrum according to one embodiment. Fig.4D depicts CT slice image reconstructions for sagittal and coronal views at low energy, high energy and total energy, according to one embodiment. Fig.5A displays log visualizations of the coefficient curves for four basic tissue types—bone, muscle, fat, and lung—as well as for water, according to one embodiment. Fig.5B presents the coefficients of Fig.5A using HUs, with water set as the reference across all energy levels, according to one embodiment. Fig.5C offers a zoomed-in view of the HU curves from panel (b), excluding bone to highlight the differences among the remaining tissues. Fig.6 depicts a typical histogram of human chest CT DICOM image acquired with X-ray tube voltage of 120 kVp. Fig.7A depicts a generalized predictive model having a plurality of machine learning models associated to different energy levels, according to one embodiment. Fig.7B depicts a random forest based predictive model having a plurality of machine learning models associated to different energy levels, according to one embodiment. Fig.7C depicts an artificial neural network (ANN) based predictive model having a plurality of machine learning models associated to different energy levels, according to one embodiment. Fig.8 is a block diagram of an embodiment of the presented pkMI system for medical imaging- enabled diagnosis of in vivo tissues, where an adaptive learning classification (ALC) algorithm with area under the curve (AUC) as the figure of merit is applied to n sets of TEEFs. Fig.9 depicts training and inferencing of an artificial neural network (ANN) based machine learning model according to one embodiment. Fig.10 depicts an artificial neural network (ANN) based machine learning model according to one embodiment. Fig.11 depicts a computer system according to one embodiment. DETAILED DESCRIPTION OF THE DISCLOSURE System 100 for use in characterizing images such as voxel-based medical images is shown in Fig.1. System 100 can be useful in de-noising images so that images more accurately represent underlying physical attributes being subject to imaging. System 100 can include manager system 110 having an associated data repository 108, imaging devices 130A-130Z, user equipment (UE) devices 140A-140Z, data sources 150A-150Z and one or more application program interface (API) service endpoint 160. Imaging devices 130a-130Z, UE devices 140A-140Z, data sources 150A-150Z and one or more API service endpoint 160 can be connected and in communication with one another via network 190. Network 190 can be a physical network and / or a virtual network. A physical network can be, for example, a physical telecommunications network connecting numerous computing nodes, such as computer servers and computer clients. A virtual network can, for example, combine numerous physical networks or parts thereof into a logical virtual network. In another example, numerous virtual networks can be defined over a single physical network. Data repository 108 can store various data. Data repository 108 and images area 2121 can store images such as voxel-based images. Images 2121 in images area can include medical imaging images such as computed tomography (CT), X-ray, magnetic resonance imaging (MRI), ultrasound, positron emission tomography (PET), fluoroscopy, nuclear medicine imaging, thermography images, and the like. Data repository 108 includes decision data structures area 2122 which can store decision data structures for return of action decision. Decision data structures can include e.g. decision tables and decision trees. Data repository 108 can further include models area 2123 which can store trained predictive models trained by machine learning. After training with sufficient training data, models stored in models area 2123 can be inferenced for return of predictions as to characteristics of underlying image data used for querying. UE devices 140A-140Z can be associated to various administrator users of system 100. Administrator users can include users associated e.g. to manager system 110, imaging devices 130A-130Z, data sources 150A-150Z or users that associated to any one of manager system imaging devices 130A-130Z or data sources 150A-150Z. In one aspect, manager system 110 can present user interfaces on UE devices for return of selection data by administrator users of such UE devices 140A-140Z. UE devices 140A-140Z can be provided, e.g., by smartphones, PCs, laptops, tablets, custom consoles, and the like. Imaging devices 130A-130Z can include imaging devices that capture image data. Imaging devices 130A-130Z can include e.g. include e.g. computed tomography (CT) imaging devices which can combine multiple x-ray images taken from different angles to create cross-sectional images (slices of bones, blood vessels and soft tissues), X-ray imaging devices, magnetic resonance imaging (MRI) imaging devices, ultrasound imaging devices, positron emission tomography (PET) imaging devices, fluoroscopy imaging devices, nuclear medicine imaging devices, thermography imaging devices, and the like. An output CT image can include voxel values at various voxel positions expressed in terms of Hounsfield Units (HUs). Manager system 110 can run various processes. Manager system 110 running user interface (UI) process 111 can include manager system 110 presenting user interface on a UE device to permit an administrator user to label image data, such as voxel-based image data. In one use case, an administrator user can label an historical image with a label according to a diagnosis based on examination of tissue associated to the historical image. Manager system 110 running UI process 111 can include manager system 110 presenting a user interface to a user that permits the user to define configurations in the performance of machine learning processes such as attributes impacting training, inferencing, testing, and the like. In one aspect user interface functionality can permit a user to select, e.g., the number of image energy levels to which a patient image is to be decomposed by decomposition process 112. Manager system 110 running image decomposition process 112 can include manager system 110 decomposing an image into multiple different energy levels. Manager system 110 running transformation process 113 can perform various transformations of respective energy level differentiated images output by decomposition process. Transformation of an image can include e.g., providing a Harris operator G in dependence on a first order derivative producing a second order derivative function H in providing a measurement of in vivo tissue elastic movement. Manager system 110 running transformation process 113 can output in one embodiment a plurality of elastic images at differentiated energy levels. Manager system 110 performing training process 114 can include manager system 110 training a predictive model. Training of a predictive model herein can include applying data resulting from performance of transformation process 113 as training data for training of a predictive model. The predictive model, once trained, can be queried for return of predictions as to attributes of a physical object represented by image data. Manager system 110 performing testing process 115 can include the manager system 110 testing a trained predictive model, such as a model developed in training process 114. Manager system 110 can qualify a trained predictive model for deployment when the testing process 115 establishes that a predictive model is performing satisfactorily according to one or more performance criterion. Manager system 110 performing inferencing process116 can include the manager system 110 querying a trained predictive model once the predictive model has been trained 114, qualified 115, and deployed. Manager system 110 performing inferencing process 116 can include the manager system 110 inferencing a trained predictive model with use of applied inferencing data defined by elastic images at multiple energy levels with the trained predictive model providing insights regarding the applied data. A method for performance by manager system 110 interoperating with data sources 150A-150Z, imaging devices 130A-130Z and UE devices 140A-140Z is set forth in reference to the flowchart of Fig. 2A-2B. At block 1501 data sources 150A-150Z can be sending image data for receipt by manager system 110, which can be stored by manager system 110 on receipt of the image data. The image data can include e.g. voxel-based image data from a medical device imaging device as set forth herein. Data sources 150A-150Z can be trusted data sources storing medical image data. Data sources 150A-150Z can be data sources that provide various repositories of medical images of patients and / or image repositories for educational and / or research purposes. Examples of data sources 150A-150Z can include, e.g., Radiopaedia, which is a large free educational radiology source with images and tech case studies, The Cancer Imaging Archive (TCIA), the National Institute of Health (NIH) Clinical Center, the National Library of Medicine (NLM), Open-i and the like. Image data sent at block 1501 can include associated pathologically-determined documentation, e.g., specifying malignant / benign status of each image. A pathologically determined status associated to an image can preferably be assigned by a medical expert based on comprehensive testing of tissue associated to an image. A pathologically determined status herein can represent the ground truth. Embodiments herein recognize that accuracy of imaging results hereon can be improved with training with the use of labels based on pathologically determined status determined by comprehensive tissue testing, rather than labels determined by human viewing of images alone. At block 1301 imaging devices 130A-130Z can be sending image data, e.g., voxel-based image data representing current patient tissue. In one embodiment, the voxel data can be live real time voxel data. Imaging devices of imaging devices 130A-130Z can include, e.g., computed tomography (CT) imaging devices which can combine multiple x-ray images taken from different angles to create cross-sectional images (slices of bones, blood vessels and soft tissues), X-ray imaging devices, magnetic resonance imaging (MRI) imaging devices, ultrasound imaging devices, positron emission tomography (PET) imaging devices, fluoroscopy imaging devices, nuclear medicine imaging devices, thermography imaging devices, and the like. On receipt of imaging data sent at block 1301, manager system 110 can store received images into image library 2121. At send block 1401, UE devices 140A-140Z can be sending selection data. Selection data defined by an administrator user of one or more UE device of UE devices 140A-140Z can include selection data, e.g., for activating labeling, activating remediation, activating training, and the like. At store block 1101, manager system 110 can store image data sent at block 1501 and / or block 1301 into image library 2121 and can also store sent selection data sent at block 1401 into data repository 108. In some cases, image data can be stored into non-volatile storage memory and / or, in some instances, received image data can be stored, e.g. into volatile working memory, e.g., to facilitate, provide buffering and real-time processing of live current patient image data. At block 1102 manager system 110 can ascertain whether a labeling mode has been activated based on sent selection data sent at block 1401. On determining that image labeling has been activated manager system 110 can proceed to send block 1103. At send block 1103, manager system 110 can send prompting data for display on a user interface of a UE device. The prompting data can prompt the user for label assignment to displayed labels, which may provide further information related to diagnostics, treatment or intervention. With labeling selection data sent at block 1402, the administrator user can set labels to sample images received from data sources 150A-150Z in accordance with the health care expert pathologically determined status of sample images. In another example, the sample images sent at block 1501 can be pre-labeled at data sources 150A-150Z, e.g., with benign or malignant status. Another example, manager system 110 can automatically label received images sent at block 1501, e.g., by use of methods that include subjecting image documentation to natural language processing. At store block 1104, manager system 110 can store any administrator designated labels to associated frames of image data stored in image library 2121. On completion of store block 1104, manager system 110 can proceed to decision block 1105. At decision block 1105, manager system 110 can ascertain whether training of a predictive model has been activated. At decision block 1104, manager system 110 can examine selection data sent at block 1401. On the determination that training has not been activated, manager system 110 can bypass block 1106 to 1108 and can proceed to block 1109. On the determination at block 1105 the training has been activated, manager system 110 can proceed to training block 1106. At training block 1106, manager system 110 can perform training and validating of a predictive model. In one embodiment, a predictive model can be, e.g., random forest-based predictive model or a neural network based predictive model. Training at training block 1106 can include, e.g., training one or more random forest and / or one or more neural network. On completion of training block 1106, manager system 110 can proceed to testing block 1107. At testing block 1107, manager system 110 can perform testing of the trained predictive model at trained block 1106. Testing at block 1107 can include testing with use of holdout data. On completion of testing block 1107, manager system 110 can proceed to store block 1108. At store block 1108, manager system 110 can store into models area 2123 trained predictive models trained and tested at block 1106 and 1107. On completion of store block 1108, manager system 110 can proceed to remediation decision block 1109. At remediation decision block 1109, manager system 110 can ascertain whether remediation has been activated for one or more patient. At block 1109, in one embodiment, manager system 110 can examine selection data sent at block 1401. For performing remediation herein, manager system 110 can process image data, such as CT scanner-based image data and based on the processing of such image data can provide an action decision. On the determination at block 1109 that remediation has not been activated, manager system 110 can bypass blocks 1110 to 1115 and can proceed to return block 1116. On determination at block 1109 that treatment has been activated, manager system 110 as set forth in reference to Fig.3 can proceed to perform decomposition 1110 at (A) to decompose an imageinto multiple different energy decomposed images ^^ఌ^ െ ^^ఌ^. Manager system 110 can then performtransformation processing 1111 at (B) to output multiple elastic images ^^ఌ^ െ ^^ఌ^, there being oneelastic image for each decomposed image. In reference to Fig.3, processing by manager system 110 can include energy decomposition for producing of decomposed images having different energy levels and transformation processing for production of elastic images at different energy levels. Decomposition and transformation processing are described further in reference to decomposition block 1110 and transformation block 1111 as set forth in reference to the flowchart of Figs 2A-2B. On determination at block 1109 that treatment has been activated, manager system 110, in one embodiment, can proceed to decomposition block 1110. At decomposition block 1110, manager system 110 can perform energy decomposition of a received patient image. The received patient image can be a CT image in one embodiment. By performance of energy decomposition of a CT image, manager system 110 can produce representations of the CT image at different energy levels. For example, Fig. 4D illustrates CT image reconstructions of Sagital and Coronal images at low energy, high energy and total energy. Energy decomposition of a CT image refers to the process of breaking down the CT data into different energy levels or spectral components based on the x-ray energy spectrum used during imaging. X-ray beams consist of various energy levels, and different tissues and materials (such as bone, soft tissue, or contrast agents) absorb x-rays differently depending on their energy. In dual-energy CT, images are captured at two distinct energy levels, while spectral CT can involve multiple energy ranges. By using decomposition algorithms, the image data is separated into components corresponding to these energy levels, allowing for more precise material differentiation. For instance, image decomposition by energy level enables, e.g., the distinction between bone and soft tissue, or between iodine contrast agents and calcium. This process enhances material characterization, improves contrast resolution, enables quantitative imaging, and reduces artifacts, especially in the presence of high-density materials like metal implants. Applications of energy decomposition include identifying specific tissues, such as distinguishing kidney stones from other calcifications, enhancing the visibility of tumors or clots, measuring contrast agent concentrations, and mitigating imaging artifacts. The decomposition can occur in two-material forms, such as separating water from iodine, or three-material forms, such as separating water, iodine, and calcium, all of which lead to improved diagnostic accuracy and image quality in clinical settings. Energy decomposition, in one embodiment, is described further in reference to Figs.4A-6. When using CT imaging modality to evaluate pulmonary nodules (PNs) and colorectal polyps (CPs), it's critical to understand the interactions between in vivo human tissues and X-ray energy, the type of tissue information embedded in the acquired data, and how these data are reconstructed into images. Fig.4A is a graph which depicts four typical X-ray energy spectra used in current CT devices, with peak tube voltages of 80, 100, 120, and 140 kVp. These spectra sharply decrease at around 40 keV and drop to zero near 20 keV due to filtration of the emission from the X-ray tube, as X-rays with energies below 40 keV rarely penetrate the human body sufficiently to contribute to data acquisition and image formation. X-rays with energies from about 20 to 120 keV (when X-ray tube is operated at a tube voltage of 120 kVp) interact with body tissues during traversal, creating varying yet reproducible image contrasts essential for different clinical applications. This interaction captures tissue-specific information within the image contrasts, making it critical for accurately diagnosing lesions composed of in vivo tissues. Fig.4A depicts four typical X-ray energy spectra used in current CT imaging devices, operating at X-ray tube peak voltages of 80, 100, 120, and 140 kV, respectively. The x-axis represents X-ray energies in keV, while the y-axis shows the X-ray flux density at each energy value. When an energy spectrum is normalized to 1, the flux density at a specific energy value indicates the probability of an X- ray having that energy value within the flux. Fig.4B illustrates a simplified CT system suitable for acquiring CT image data in accordance with the present methods. Fig.4C is a graph depicting a simulation of filtered and unfiltered X-ray spectrum. The slope line 400 shown in Fig.4C is a simulation of unfiltered or original X-ray tube emission energy spectrum. The shaded areas together show a simulation of filtered energy spectrum. The shaded areas 405, 410, 415, 420, 425 under the spectrum curve depict different energy ranges from the lowest 405 to the highest 425. The energy-dependent image contrast variation across the body can be seen from Fig.4D. The right-most images show image slices of the human body in two views, i.e., sagittal and coronal, using the spectrum of Fig.4A with tube kVp = 120, called Total Energy images. The middle images shows the two slice images when the low energies (< 55 keV) were blocked, and are referred to as High Energy images. The left-most images illustrate the two slice images when the high energies (> 55 keV) were blocked and are labeled Low Energy images. The Low Energy images have high image contrasts but are generally characterized as being noisy. The High Energy images typically suffer from having lower contrasts but are generally less noisy than the corresponding Low Energy images. The Total Energy images typically look like an “average” of the Low and High Energy images thereby taking advantage of the properties of both the Low Energy and High Energy images. The energy-dependent image contrasts can be explained by the X-ray Physics below. The interaction between X-ray energies and tissues is quantified by X-ray attenuation coefficients. Fig.5A depicts the linear attenuation coefficient curves for various tissues such as bone, muscle, fat, and lung, as well as for water, across energies ranging from 1 to 140 keV. Although each tissue type may exhibit different attenuation curves when located in various organs or under different pathological conditions, the attenuation curve for water remains consistent across all scenarios. This consistency is crucial for developing models to describe attenuation differences and image contrasts among tissues. Fig.5B reinterprets these curves using water as a reference and expresses the attenuation coefficients in Hounsfield Units (HUs). Fig.5C provides a zoomed-in view of Fig.5B with the Y-axis limited to the range of -700 to 200 HU. These curves demonstrate that while the energy range from 20 to 40 keV is effective for generating distinct image contrasts among bone, lung, muscle, and fat, the contrast between muscle and fat is limited in the 40 to 120 keV range. Consequently, CT imaging devices that use the spectra shown in Fig.4A and operate above 20 keV often produce minimal contrast between muscle and fat tissues. This is significant because many tissue changes in lesions, such as PNs and CPs, primarily occur around muscle and fat, resulting in reduced diagnostic capability. Additionally, since data acquisition spans from 20 to 120 keV, the reconstructed polychromatic image does not show contrast variation across different energy values, further limiting diagnostic efficacy. Systems and methods described herein derive distinct image contrasts at individual energy values within the full spectrum from 1 to 120 keV range, based on lesion images reconstructed from data obtained using the broad spectra illustrated in Fig.4A. For the task of lesion diagnosis, the present disclosure preferably focuses on the energy range of 1 to 50 keV, where there are significant image contrast variations between fat and muscle tissues. Fig.5A-5C depict various aspects of linear attenuation coefficients for different materials in CT imaging. Fig.5A displays log visualizations of the coefficient curves for four basic tissue types—bone, muscle, fat, and lung—as well as for water. Fig.5B presents these coefficients using HUs, with water set as the reference across all energy levels. Fig.5C offers a zoomed-in view of the HU curves from panel (b), excluding bone to highlight the differences among the remaining tissues. The derivation of image contrasts at individual energy values from 1 to 50 keV is based on prior knowledge of X-ray interactions with the four basic tissue types: bone, muscle, fat, and lung. It is expected that deviations occur in the derived image contrasts when some properties of these tissues change from normal to abnormal. To test this, we identify which tissue properties should be considered to reflect pathological changes from normal to abnormal and how to relate these considered tissue properties to the derived image contrasts. Solving a problem often requires a task-specific approach. In the context of diagnosing lesions, prior research indicates that lesions result from the pathological transformation of in vivo tissue properties from normal to abnormal. Tissue elasticity and growth rate are examples of properties indicative of pathological conditions. This disclosure presents a method to determine the characteristics of tissue elasticity and growth rate in normal tissue using established scientific principles and prior knowledge. The study then extracts the tissue pathological characteristic features (TPCFs) from the derived image contrasts at individual energy values from 1 to 50 keV. By applying this method to patient data, the extracted TPCFs will show deviations from those of normal tissues. These deviations quantitatively indicate pathological changes from normal to abnormal states. Consequently, if tissue pathology is established as the definitive indicator of malignancy, an intelligent machine learning (ML) model should be capable of predicting the malignancy of a lesion based on these deviations. In the following sections, we aim to establish two baselines to differentiate changes in tissue from normal to abnormal. The first baseline focuses on the image contrast variations of normal tissues within the X-ray energy range of 1 to 50 keV. The second baseline examines the dynamic or functional characteristics of normal tissues, specifically tissue elasticity and growth rate. These methods will be applied to patient data to extract TPCFs. A corresponding ML algorithm will be developed to classify the TPCFs using tissue pathology as the ground truth of malignancy. The classification outcomes will provide a quantitative measure of the efficacy of the presented pkMI system. In the following, medical experts have interpreted the patients' original CT DICOM images, which were obtained via clinical LdCT screening protocols. They identified all abnormalities or lesions in the CT images and categorized the indeterminate lesions (IDLs). They further delineated the borders of volumetric IDLs on each image slice in the CT volumetric images. Each of these delineated images ofIDLs is denoted as a function, ^^^^^^,^^, ^^^, where the indices ^^^,^^, ^^^ represent the coordinates of allimage voxels within the volumetric data of an IDL and the value of ^^^^^^,^^, ^^^ indicates the imageintensity at voxel ^^^,^^, ^^^.Fig.6 depicts a typical histogram of human chest CT DICOM image acquired with X-ray tube voltage of 120 kVp. From the histogram, three thresholds can be determined to label all the image voxels as predominantly containing bone, muscle, fat, or lung tissue [17,18]. By this histogram-based thresholding segmentation method, a voxel is labeled as containing a single tissue type, although there are frequently more than one tissue type represented by each voxel. This is called the partial volume (PV) effect. To address the PV effect, this disclosure employs a neighboring system to estimate the percentage of each tissue type in a voxel. From the segmented CT image, a cubic mask of 3×3×3 is applied to analyze the tissue composition within. By counting the voxels corresponding to different tissue types within the mask, the ratio of the number of voxels of a specific tissue type to the total number of voxels (27 in this case) within the mask provides the percentage representation of that tissue type at the central voxel of the mask. This is denoted as ^^^ఛ, representing the percentage of tissue type ^^ in voxel i. The attenuation coefficient of tissue type ^^ in voxel i at a given energy value of ^^ is represented as ^^^ఛ^^^^. Utilizing the tissue-energy relationship depicted in Fig.6, along with the calculated percentages of the four basic tissue types—bone, muscle, fat, and lung—within each voxel, the attenuation coefficients for each tissue type across the energy range of 1 to 50 keV can be computed. To derive an image of a specific energy value, the contributions of all four basic tissue types within each voxel are summed based on their respective attenuation coefficients at that energy value. This sum represents the total attenuation within the voxel, leading to a reconstructed image that reflects the combined tissue characteristics at the specified energy value: ^^ఌ ൌ ∑ସ ^ఛୀ^ μఛ^^ε^ρఛ ^ (1) where τ is the index for the four basic tissue types of bone, lung, muscle, and fat for lung For abdominal CT colonography (CTC) imaging, the tissue types are adapted to reflect the areas of interest: bone is replaced by oral tagging materials and lung is replaced by colon lumen. This adjustment allows for a more accurate representation of the tissue composition relevant to the specific medical imaging or CTC application. The computed image at each energy value ^^, denoted as ^^ఌ^^^,^^, ^^^, is referred to as effectivemono-energetic image (EMI). The EMI serves a similar purpose to the virtual mono-energetic image, conceptually designed to provide enhanced imaging clarity and specificity by simulating an image obtained at a single energy value. These EMIs are instrumental in enhancing visibility on different tissues and particularly providing image contrasts of different tissues at individual energy values in the energy range from 1 to 50keV for lesion diagnosis. If the tissues within the body are normal, their attenuation coefficients will generally align with the curves depicted in Figs.5A and 5B at any specified energy values. These curves serve as the baseline or standard reference for each tissue type at various energy values. When imaging patient lesions, any deviation observed in the attenuation coefficients from these baseline curves indicates a potential pathological change. Such deviations could suggest a transition from normal to abnormal tissue states. While it is possible to generate a set of EMIs at any energy values, it's often practical to select a range of energy values where the contrast differences among lesion tissues are most noticeable. In addition, an optimal number of energy values in the range may be needed to balance the benefits of enhanced image contrasts against the potential drawbacks of data redundancy. The focused energy range and number of samples in the range will help perform the diagnostic task. For this example, a set of ten EMIs were generated at 5, 8, 10, up to 45 in an interval of 5 keV. To analyze each image voxel along the curves or the ten EMIs (energies) for detecting deviations from the baseline for lesion diagnosis, a descriptor of lesion malignancy is required. In embodiments of the present disclosure, constructing a lesion malignancy descriptor is performed using the contrast distribution across the lesion volume at each individual energy value or EMI. This volume-based construction of the lesion malignancy descriptor can be viewed as follows. Each EMI, ^^ఌ^^^,^^, ^^^,can be interpreted as the output from the energy-sensitive sensor shown in Figs.5A and 5B, applied to the in vivo tissues at the specific energy value ^^. Under normal conditions, the contrast distribution in the EMI of the tissues throughout the volume should mirror the properties of healthy tissues. For imaging in vivo tissues, the distribution of contrast variation in each EMI contains tissue dynamic information. Deviations from this norm suggest potential pathological changes. Various tissue properties can indicate such changes, and this study specifically investigates the dynamic properties of tissue elasticity and growth rate as biomarkers of pathology. In the next section, our goal is to establish a baseline for identifying tissue pathological characteristics (TPC) from the contrast distribution in each EMI. Deviations in the pathological characteristics of a patient's LdCT scan from this baseline at certain energy values will indicate a transition in the tissue from normal to abnormal. There is set forth herein, in reference to decomposition block 1110 and Figs.4A-6, decomposing a patient image representing a patient tissue section into a plurality of effective mono- energy images at different energy levels 1 through n, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level. First and second energy levels herein can refer to any first and second energy levels of energy levels 1 through n, e.g., energy levels 1 and 2, energy levels 1 and 5, energy levels 6 and 2, energy levels n and n-3, and the like. On completion of energy decomposition at energy decomposition block 1110, manager system 110 can proceed to transformation block 1111. At transformation block 1111, manager system 110 can perform transformation of energy decomposed images produced at decomposition block 1110 for output of an elastic image at the respective energy levels of the energy decomposed images. Elastic images herein can represent dynamic properties of images. Production of an elastic image herein can include calculating at least a jth order derivative of an input energy decomposed image, wherein j ≥ 1. As previously stated, lesions result from pathological changes in tissue properties, transitioning from normal to abnormal. These changes encompass both anatomical and functional alterations, henceforth referred to as tissue dynamic changes. These dynamic changes in tissues are mirrored by changes in image contrast dynamics. Consequently, an image contrast dynamic model is necessary to accurately describe these pathological changes in tissues. Ideally, capturing the dynamic changes in tissues requires an imaging device with extremely high spatial, temporal, and energy resolutions. However, such a perfect device does not exist in practice. We treat the image contrast distributions across the lesion space as a dynamic process, with the computed EMI at each energy value serving as an observation of this dynamic process. As the displacement between two observations approaches zero, the dynamics can be mathematically describedby the derivatives of the computed EMI function, denoted as ^^^^^,^^, ^^^, where the energy level index ^^ isomitted for simplicity. Our objective is to extract tissue dynamical characteristics from these derivatives, which can reveal various dynamic properties indicative of pathological conditions. This embodiment employs the well-known Affine transform as a model to describe soft tissue movement, specifically referencing its use in expressing soft tissue elasticity as elastic modulus as further described in

[0022] ,

[0023] . The Affine transform, widely utilized in elastic registration of soft tissue images, provides a mathematical framework to correlate two such images. Crucially, based on physical principles, this transformation must remain invariant. The task is to identify an invariant Affine transform using thederivatives of the function ^^^^^,^^, ^^^. The subsequent section details our approach to achieving this task.Let ^^ ൌ ^^^௫,௬,௭ ^ represent volumetric data. The 1st order derivative is expressed as the gradientoperator, denoted as a vector by ∇^^ ൌ ൫^^௫ᇱ , ^^௬ᇱ , ^^௭ᇱ൯, where ^^௫ᇱ , ^^௬ᇱ and ^^௭ᇱ are the 1st order derivatives of thevolumetric data along the x, y, z directions, respectively. The gradient signifies the maximum local shape changes at every point along under some microscopic stressing or pushing on the in vivo tissues. Drawing from the concept of the gradient and its practical uses, the Harris operator, G, is introduced. G is a matrix or tensor extensively utilized to encapsulate the full local deformation along all directions, providing a comprehensive representation of the tissue's structural dynamics

[0024] : ^^௫ᇱ^^௫ᇱ^^௫ᇱ^^௬ᇱ^^௫ᇱ^^௭ᇱ^^ ൌ ^^^^^^் ∙ ^^^^^^ ൌ ^^^௫ᇱ^^௬ᇱ^^௬ᇱ^^௬ᇱ^^௬ᇱ^^௭ᇱ^ (2) matrix (called Hessian matrix)

[0025] : ^^௫ᇱᇱ௫ ^^௫ᇱᇱ௬ ^^௫ᇱᇱ௭^^ ൌ ^^^௫ᇱᇱ௬ ^^௬ᇱᇱ௬ ^^௬ᇱᇱ௭^(3)order derivatives of the volumetricdata. registration of soft tissuemovement, a differential Affine invariant can be derived using the derivatives of function ^^^^^,^^, ^^^ asfollows. Define two hybrid tensors as: ^^^ ൌ ^^ െ ^^ and ^^ଶ ൌ ^^ ^ ^^. In the differential space atnon-critical points (|^^| ് 0), define two scalers:^^ ൌ|^భ| |^^|ு|and ^^ ൌమ|ଶ|ு|(4) (5)(6) ∇^^. Combining Eq. (4) and Eq. (6), the scaler ^^^can be derived, and similarly the scaler ^^ଶcan be obtained: ^|^భ||ு|൫^ି^∇ூ^ுషభ^∇ூ^^^^ൌ ൌ൯ ൌ 1 െ ^∇^^^^^ି^^∇^^^் (7)(8) to how the Affine transform is performed, the expression ^∇^^^^^ି^^∇^^^்remains unaffected by the Affine transformation. Therefore, this expression yields a critically important invariant term involving the 1st and 2nd order derivatives, which is used to describe the elastic movement of in vivo tissues. This invariant provides a reliable basis for quantifying tissue dynamics under different conditions, thereby facilitating accurate assessments of tissue elasticity and other related dynamic properties

[0028] : ^^ ൌ ^∇^^^^^ି^^∇^^^் (9)where the 1st order derivative, ∇^^, represents a move trend at a location ^^^,^^, ^^^, while the 2ndorder ^^^. Together, these dynamic attributesform a simple symmetric formula representing the in vivo tissue elastic movement descriptor ^^ at thesame location ^^^,^^, ^^^.Using the derived formula of Eq. (9) and the expressions of Eqs. (2) and (3), we can compute animage, ^^ఌ^^^,^^, ^^^, from the corresponding effective mono-energetic image or EMI, ^^ఌ^^^, ^^, ^^^, on avoxel-by-voxel manner

[0028] . This process results in ten images, ^^^ఌ^, corresponding to the ten individual energy values indexed by ε. For simplicity, ^^ఌis called elastic image hereafter. Inputting the set of ten volumetric elastic images into a current ML algorithm for predicting lesion malignancy encounters the problem of redundancy among multiple datasets from the same object or lesion

[0029] . While there may be other ways to modify current ML architecture designs to address this redundancy problem, this disclosure applies an adaptive learning and classification (ALC) strategy. This approach can intelligently distinguish and learn from the nuances between the datasets, thereby enhancing the diagnostic process by reducing redundancy and improving accuracy in malignancy classification. More details are presented in the following section. There is set forth herein, in reference to blocks 1110 and 1111, Figs.4A-6, and Eqs.1-9 decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively. Referring to Fig.2B, on completion of transformation block 1111, manager system 110 can proceed to inferencing block 1112. At inferencing block 1112, manager system 110 can perform inferencing of a trained predictive model, such as a random forest based predictive model or a neural network based predictive model. Inferencing at inferencing block 1112 can include inputting one or more image(s) such as depicted in Fig.4D of the derived elastic images derived at decomposition block 1111 as inferencing data into a predictive model such as predictive model 7002 (Fig.7A). On performance of inferencing at inferencing block 1112, manager system 110 can return a diagnosis for patient based on a processed image such as a processed CT scan image. The diagnosis can include, e.g., the status ‘malignant’ or ‘benign’. Referring back to Fig.2B, at inferencing block 1112, manager system 110 can inference a trained predictive model returning result data. Manager system 110 can perform inferencing atinferencing block 1112 with use of elastic images ^^ఌ^ െ ^^ఌ^ returned at transformation block 1111.Manager system 110 inferencing a trained predictive model is illustrated and as described further in reference to Fig.7A. In reference to Fig.7A, inferencing data defined by elastic images ^^ఌ^െ ^^ఌ^can be input into trained predictive model 7002 for return of result data. As shown in Fig.7A, trained predictive model 7002 can include a plurality of separately trained machine learning models 7003, 7004, 7005 MLM1-MLMn. In further reference to Fig.7A, an image I representing tissue of a patient 101 can be output by imaging device 130 and can be subject to energy level decomposition at decomposition block 1110 as explained in reference to decomposition block 1110 of Figs.2A-2B. Forproduction of energy decomposed images ^^ఌ^ െ ^^ఌ^, the energy decomposed images ^^ఌ^ െ ^^ఌ^. can besubject to transformation by transformation block 1111 of Figs.2A-2B for output of elastic images^^ఌ^ െ ^^ఌ^ which elastic images ^^ఌ^ െ ^^ఌ^ defining inferencing data can be input into trainedpredictive model 7002. As indicated by trained predictive model 7002, the respective elastic images^^ఌ^ െ ^^ఌ^ derived by processing a patient image defining inferencing data can be respectively inputinto separate and distinct machine learning models 7003, 70004, 7005 MLM1-MLMn. In a further aspect, predictive model 7002 configured as a random forest based predictive model is depicted in Fig. 7B, and predictive model 7002 configured as an artificial neural network (ANN) based predictive model is depicted in Fig.7C. Separate machine learning models MLM 17003, MLM27004, MLMn 7005 can be previously subject to separate and differentiated training at training block 1106 with use of differentiated training datasets. For the separate training of the separate machine learning models MLM1-MLMn, manager system 110 can process historical images associated to pathologically examined tissue from image library 2121 that have been sent from data sources 150A-150Z at send block 1501. For production of training data for training of the separate machine learning models MLM1-MLMn, manager system 110 can subject historical images to decomposition and transformation processing in the manner of patient images as set forth in reference to decomposition block 1110 and transformation block 1111. As indicated in Fig.7A, manager system 110 for a plurality of historical pathological images I can subject the historical pathological images I to decomposition processing at decomposition block2110 for production of output energy decomposed images ^^ఌ^ െ ^^ఌ^ having different energy levels andcan then subject the respective different energy decomposed images ^^ఌ^ െ ^^ఌ^ to transformation attransformation block 2121 with use of the transformation processes set forth in reference to Eqs.2through Eq. 8 for production of elastic images ^^ఌ^ െ ^^ఌ^, which elastic images ^^ఌ^ െ ^^ఌ^ can be inputinto trained predictive model 7002 as training data. Based on processing by transformation block 2111, manager system 110 can output as trainingdata training data that comprises different elastic images ^^ఌ^ െ ^^ఌ^, each having respectively differentenergy levels. In one aspect, separate machine learning models MLM17003 to MLMn 7005 defining trained predictive model 7002 can each have respectively different energy levels. Trained predictive model 7002 can be trained with elastic image training data of respectivelydifferent energy levels defined by training data elastic images ^^ఌ^ െ ^^ఌ^. MLM17003 can be trainedwith elastic image training data defined by iterations of elastic image ^^ఌ^derived by processing of historical instances of images associated to pathologically examined tissue, MLM27004 can be trained with elastic image training data defined by iterations of elastic image ^^ఌଶ, , and machine learning model MLMn can be trained with elastic image training data defined by elastic image ^^ఌ^. Machine learning models MLM1 to MLMn can assume a variety of different machine learning model types, e.g., random forest, artificial neural network (ANN), support vector machine (SVM) and the like. Referring to Figs.2A and 2B, at inferencing block 1112, manager system 110 can select for inferencing a trained predictive model appropriate for processing of the received patient image received in response to send block 1301. At inferencing block 1112, manager system 110 can qualify and select for inferencing a previously trained and stored predictive model previously trained and tested at training and testing blocks 1106 and 1107 based on one or more criteria having been satisfied. At inferencing block 1112, manager system 110 can perform domain matching so that the received patient image for processing is of a common domain as historical pathological images used for training the selected predictive model 7002 as shown in Fig.7A. For performing domain matching, referring to Fig.1, manager system 110 can perform clustering analysis to compare feature vectors of the received patient image to feature vectors of training data used for training each respective trained predictive model stored in models area 2123. CT scans can be used to visualize a wide range of tissue type domains by detecting differences in density and composition. Bone tissue, which is dense, is clearly visualized, making CT highly effective for identifying fractures, abnormalities, and bone density changes. Soft tissues, including organs such as the liver, spleen, kidneys, and muscles, can be assessed for tumors, cysts, inflammation, and other abnormalities. Lung tissue is commonly examined for infections, nodules, and interstitial lung diseases. Fat tissue is distinguishable on CT scans, useful for identifying lipomas and assessing fat distribution within organs. Blood vessels are visualized through CT angiography, aiding in the detection of aneurysms, stenosis, and blockages. Brain tissue is frequently scanned to assess structural abnormalities related to strokes, hemorrhages, tumors, or trauma. While CT is not as detailed as MRI for soft tissues like cartilage and ligaments, it can still be used to identify injuries, particularly in joints. Additionally, CT is often effective in locating and characterizing tumors or masses, providing crucial information about size, location, and potential metastasis. Contrast agents are often used in CT imaging to enhance the visibility of specific tissues, such as blood vessels or certain organs, for more precise diagnostics. In reference to predictive model 7002 described in Figs.7A-7C, manager system 110, e.g., at testing block 1107, can evaluate performance of machine learning models, MLM1-MLMn defining predictive model 7002. In one aspect, manager system 110 can perform evaluating of the performance of machine learning models MLM1-MLMn defining predictive model 7002, and can adjust weights associated to the various machine learning models MLM17003 to MLMn 7005 in dependence on the evaluating of the performance of machine learning models. In applying evaluation dependent weight adjustments to the various machine learning models MLM1 to MLMn, manager system 110 can, e.g., deactivate one or more machine learning model for economization of computing resources, reduce a volume of training data applied to one or more machine learning model for economization of computing resources, and / or can scale result data of one or more machine learning model in providing a result from inferencing predictive model 7002. In one embodiment, there is set forth herein, in reference to Fig.2B blocks 1107 (testing), 1110 (decomposition) and 1111 (transformation), Figs.4A-7C, and Eqs.1-9 decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively, wherein the predictive model includes a first machine learning model trained with training data having the first energy level, and a second machine learning model trained with training data having the second energy level, wherein the method include evaluating performance of the first machine learning model and the second machine learning model, and adjusting a weight associated to one or more of the first machine learning model or the second machine learning model responsively to the evaluating, wherein the adjusting a weight includes deactivating the first machine learning model. Embodiments herein recognize that the loop of blocks 1101-1106 is iterative, i.e., can repeat over the course of deployment of manager system 110. At an earlier iteration result data from inferencing first and second of machine learning models MLM1- MLMn can contribute to a prediction, and in some cases, at a later iteration responsively to evaluation and testing at block 1107, manager system 110 can deactivate one or more of the first or second machine learning models. Manager system 110 evaluating of machine learning models MLM1-MLMn defining predictive model 7002 can include evaluating one or more performance criterion of MLM1-MLMn. When evaluating machine learning models, manager system 110 can apply AUC (Area Under the Receiver Operating Characteristic (ROC) Curve) as a metric, particularly for classification tasks, as it can measure the model's ability to distinguish between classes across different thresholds. AUC values can range from 0.5 (random guessing) to 1.0 (perfect classification), with a higher AUC indicating stronger performance. AUC can be derived from the ROC curve, which can plot the trade-off between the true positive rate (recall) and false positive rate across various decision thresholds. In addition to AUC, other key classification metrics can include accuracy, which can represent the proportion of correct predictions but can sometimes be misleading for imbalanced datasets, and precision, which can assess how many of the predicted positives are actually correct. Recall (or sensitivity) can measure the model's ability to identify actual positives, and the F1 score can provide a balanced measure between precision and recall, especially useful in scenarios where both false positives and false negatives are of concern. Manager system 110 can apply a confusion matrix to further break down the predictions into true positives, true negatives, false positives, and false negatives, offering more granular insight. For models that output probabilities, logarithmic loss (log loss) can evaluate the accuracy and confidence of predictions, penalizing confident but incorrect predictions more heavily. Manager system 110 can apply metrics like Mean Absolute Error (MAE), which can calculate the average absolute difference between predicted and actual values, and Mean Squared Error (MSE), which can square the errors to give more weight to larger deviations, are commonly used to assess model performance. R-squared (R²) can measure the proportion of variance in the target variable explained by the model, indicating its quality of fit. For inferencing of predictive model 7002 defined by a random forest based predictive model of Fig.7B, manager system 110 can apply as inferencing data into each respective machine learning model of MLMs MLM17003-MLMn 7005 elastic image data defined by an elastic image E having an energy level in common with an energy level of the MLM. The respective MLMs MLM17003-MLMn 7005 can subject the received inferencing data to feature extraction by feature extractors 7201, and the features output by the respective feature extractors 7201 can be applied to the respective random forest models 7202 of respective MLMs MLM1-MLMn. Returning to Fig.2B, on completion of inferencing block 1112, manager system 110 can proceed to action decision block 1113. At action decision block 1113, manager system 110 can return an action decision for remediation of a condition in dependence on a result of the inferencing at inferencing block 1112 as output by predictive model 7002. In one aspect, the action decision can include presenting on a user interface (UI), for remediation of a condition detected responsively to the inferencing at inferencing block 1112, prompting data specifying a result of the inferencing at inferencing block 1112. For example, the prompting data can include a text-based specifier of a returned diagnosis returned by the inferencing at inferencing block 1112. Such prompting data remediates a condition at least by prompting an administrator user, e.g., a medical expert to treat the condition. Treatments for a cancer diagnosis condition can include, e.g., surgery, radiation therapy, and chemotherapy. Surgery aims to remove the tumor, either fully or partially, depending on its size and location, while radiation therapy uses high-energy beams to destroy cancer cells, either externally (external beam radiation) or internally (brachytherapy). Chemotherapy involves administering drugs to target and kill rapidly dividing cancer cells throughout the body, and it can be used before or after surgery or radiation. Other options include targeted therapies, immunotherapy to boost the body’s immune response against cancer, and hormone therapy for hormone-sensitive cancers. Prompting data can include, e.g., a text-based specifier of condition diagnoses, and / or a text-based specifier of one or more treatment of such condition. In some embodiments prompting data can be presented with a visualization of patient image. In another aspect, an action decision can include defining an action for remediation of a condition, wherein the action decision includes an action decision to control live real-time remediation and treatment processes. Manager system 110 at action decision block 1113 can return an action decision to control one or more treatment process enhanced by live CT scanner images that are processed to determine malignancy (1) or benign status (0), enabling highly precise interventions across various medical fields. In adaptive radiation therapy, malignancy-processed CT images can dynamically adjust radiation doses and targeting to ensure that tumors are treated accurately while sparing healthy tissues. During intraoperative tumor resection, real-time CT images can help surgeons differentiate between malignant and benign tissue, allowing more precise excision of cancerous areas and minimizing damage to healthy tissue. Similarly, in CT-guided biopsies and needle placements, real-time CT images processed for malignancy can guide the needle precisely to suspicious areas, facilitating accurate tissue sampling or avoiding unnecessary procedures if benign regions are identified. For tumor ablation procedures, such as radiofrequency or microwave ablation, real-time CT imaging can ensure the precise targeting of malignant areas, reducing damage to surrounding benign tissue. Cryoablation procedures can similarly benefit from malignancy-processed CT images, guiding the freezing process to target only malignant regions and preserving healthy tissue. In stereotactic surgery, malignancy-processed CT images can be used in real time to guide instruments precisely to cancerous tissues, with continuous monitoring and adjustments to avoid benign areas. Brachytherapy can utilize processed CT images to ensure the accurate placement of radioactive seeds within or near malignant tumors, optimizing radiation dosage while minimizing harm to benign tissue. For lung nodules, real-time CT guidance based on malignancy status can direct interventions such as biopsies or laser ablation, targeting malignant nodules with precision and avoiding unnecessary treatment for benign ones. In chemotherapy, malignancy-processed CT images can be compared over time to track tumor shrinkage, allowing for real-time adjustments in treatment plans based on the tumor's response. Finally, in radioembolization, real-time CT images processed for malignancy can guide the delivery of radioactive beads directly to the malignant regions of the liver, ensuring targeted treatment while sparing benign liver tissue. Across these various applications, the ability to process CT images in real time to identify malignancy allows for enhanced precision in treatments, improving outcomes while minimizing damage to healthy tissue, making them advantageous in oncology and other medical interventions. There is set forth herein, in reference to Figs.2 A and 2B blocks 1110-1113, Figs.4A-6, and Eqs.1-9 decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively, and returning one or more action decision in dependence on the inferencing. On completion of action decision block 1113, manager system 110 can proceed to send block 1114. At send block 1114, manager system 110 can send prompting data to UE devices 140A-140Z for display on a display of a displayed user interface of UE devices 140A-140Z. In response to receipt of the prompting data, UE devices 140A-140Z can present the prompting data at present block 1403. The prompting data can be presented with or without a visualization of a patient image with the prompting data specifying a text-based result of the inferencing (e.g. a condition diagnosis) at inferencing block 1112 described in reference to action decision block 1113. Prompting data can include additionally or alternatively a text-based description of one or more treatment of a conditioned determined responsively to inferencing at inferencing block 1112. On completion of send block 1114, manager system 110 can proceed to send block 1115. At send block 1114, manager system 110 can send command data to API service endpoint 160 for controlling a remediation treatment action such as a remediation treatment action set forth in reference to action decision 1113. At action block 1601 the service, e.g., remediation treatment service associated to API service endpoint 160 can perform the action associated to the command data. On completion of send block 1115, manager system 110 can proceed to return block 1116. At return block 1116, manager system 110 can return to a stage preceding store block 1101 to receive and process a next iteration of image data sent at block 1501 and / or block 1301 and selection data sent at block 1401. Manager system 110 can iteratively perform the loop of block 1101-1116 for deployment period of manager system 110. In the performance of iterations of the loop of blocks 1101-1116, manager system 110, e.g., can iteratively train existing or new predictive models, can iteratively optimize such predictive models, e.g., by adjusting weights associated to machine learning models defining such predictive models. In adjusting a weight associated to a machine learning model, manager system 110, e.g., can deactivate machine learning model, e.g., by decommissioning and / or deletion. In adjusting a weight associated to machine learning model, manager system 110 additionally or alternatively can, e.g., scale result data of the various machine learning models so that a prediction result is biased in favor of better performing machine learning model(s). Manager system 110 in adjusting weights associated machine learning models can additionally or alternatively, e.g., adjust the volume of training data applied for training of various machine learning models. For example, manager system 110 can direct and apply a relatively higher volume of training data to relatively higher scored better performing machine learning models than lower scored machine learning models. For adjusting weights associated to machine learning models, manager system 110 can evaluate performance of the various machine learning models. In evaluating the performance of machine learning models, manager system 110 can score the machine learning models based on an evaluation of the models under one or more criterion and rank the machine learning models in dependence on the evaluating. In optimizing a predictive model 7002, manager system 110 can economize computing resources consumed for deployment of the predictive model. For example, in deactivating one or more machine learning model 7003, 70047005, manager system 110 conserves computing resources associated with inferencing and training such machine learning model. Further, in reducing the volume of training data applied for training machine learning model, manager system 110 reduces computing resources associated to deployment of the machine learning model. In reference again to Fig.7B, Fig.7B depicts predictive model 7002 configured as a random forest based predictive model. Referring to Fig.7B, each MLM of MLM1-MLMn can include a feature extractor 7201 in combination with random forest model 7202, and predictive model 7002 can include feature extractor 7302 for extracting training data features from elastic images derived from historical images associated to pathologically examined tissue applied as training data. MLM17003 can be trained with elastic image training data defined by iterations of elastic image ^^ఌ^derived by processing of historical instances of images associated to pathologically examined tissue, MLM27004 can be trained with elastic image training data defined by iterations of elastic image ^^ఌଶ, and machine learning model MLMn 7005 can be trained with elastic image training data defined by elastic image ^^ఌ^. Feature extraction by feature extractor 7201 can include texture feature extraction. In one embodiment, Haralick feature extraction can be employed to quantify the texture of an image by analyzing spatial relationships between pixel intensity values. Feature extraction can include constructing the Gray Level Co- occurrence Matrix (GLCM), which measures how often pairs of pixels with specific intensity values occur at a given spatial relationship (distance and direction) in the image. In one embodiment, the GLCM is calculated for multiple orientations (0°, 45°, 90°, and 135°) and distances between pixel pairs. Once the GLCM is formed, it is normalized into a probability distribution that represents the likelihood of these pixel intensity combinations. From this normalized matrix, a set of statistical features is extracted that describe various textural properties of the image, such as contrast, correlation, energy, and homogeneity. Embodiments herein recognize that feature extraction improves the ability of random forests to identify patterns by transforming raw data into more informative features, which helps capture complex relationships and reduces overfitting. Embodiments herein recognize that feature extraction can be particularly useful in handling high-dimensional data by simplifying the feature set, ensuring that the decision trees focus on the most relevant information. Additionally, embodiments herein recognize that feature extraction enhances the model's ability to determine feature importance and handle complex relationships, ultimately improving the performance and efficiency of the random forest model. In one example, manager system 110 can employ a training dataset with three features F1, F2, and F3, and a binary classification label (1 = malignant, 0 = benign). Table A depicts an example of a portion of the training data matrix. Table A Instance F1 F2 F3 Label In reference to Table A, columns F1, F2, and F3 represent the features, and the “Label” column represents whether the classification is malignant (1) or benign (0). For training of machine learning models MLM1-MLMn, manager system 110 can split the data into a training set (e.g., 80% of the data) and a holdback set (20% of the data) that will be used later for testing. In one embodiment, manager system 110 can perform bootstrapping of the training data. Bootstrapping can include randomly sampling the training data with replacement. For example, if there are 100 instances in a training set, 100 instances can be sampled with replacement, meaning that some instances might appear more than once, and some might not appear at all in each sample. For a specific tree, the bootstrapped dataset can take on characteristics as set forth in Table B. Table B Instance F1 F2 F3 Label 3 0.6 0.2 0.1 1 Here, Inst t, while Instance 4 is not included in this particular bootstrapped sample. During the training of each decision tree, random forests use random feature selection at each decision node. Since we have three features (F1, F2, F3), at each split, a random subset of features is selected. For example, the model might randomly choose two of the features (say, F1 and F2) and then determine the best split based on those features only. This process enhances the diversity among the trees and reduces overfitting. For example, suppose that for a particular node, F1 and F3 are selected as the features to evaluate. The tree will look for the best threshold for either F1 or F3 to split the data at that node. In one embodiment, the random forest can include multiple decision trees, each trained on a different bootstrapped dataset and using different random feature subsets at each split. For example, if there are 100 trees, each tree will be trained on its own bootstrapped sample of the training data. Embodiments herein recognize that the diversity between the trees (from bootstrapping and random feature selection) ensures that the random forest is less likely to overfit to specific noise or patterns in the data. After training the random forest, manager system 110 can employ the holdback set (which was not used during training) to evaluate the model's performance. Each instance in the holdback set can be passed through all the trees in the random forest. Each tree can output a classification (malignant or benign), and the random forest can aggregate these outputs, typically using majority voting. For example, if 60 out of 100 trees classify an instance as malignant, and 40 classify it as benign, the final prediction will be “malignant” (label = 1). Table C illustrates a holdback matrix example. Table C Instance F1 F2 F3 True Label Predicted Label 101 0.6 0.3 0.2 0 0 the true labels in the holdback set. Based on these predictions, manager system 110 can return evaluation metrics, e.g., AUC, accuracy, precision, recall, F1 score, to determine how well the random forest generalizes to unseen data. In performing training and testing of a machine learning model of machine learning models, manager system 110 can (a) Split the dataset into training and holdback sets, (b)Bootstrap the training data to create diverse samples for each decision tree, (c) Perform random feature selection at each node during tree construction, (d) Train multiple decision trees to form a random forest using bootstrapped samples, (e) Use the holdback set to test the random forest’s performance and evaluate with appropriate metrics. Aspects of evaluating and adjusting weights to machine learning models MLM1-MLMn defining predictive model are further described in reference to Fig.8 wherein predictive model 7002 is configured as a random forest based predictive model. Intuitively, redundancy among multiple datasets derived from the same object can be conceptualized as a series of pair-wise overlaps between two datasets. The challenge lies in determining the optimal sequence for arranging these datasets. The arrangement is contingent on the specific task to be performed. The figure of merit (FOM) for the task determines the redundancy between two datasets. For diagnosing lesions, we utilize the AUC as the FOM to rank all elastic images, ^^^ఌ^, in descending order, starting from the ^^ఌwith the highest AUC down to the one with the lowest. The ^^ఌthat contributes the most to the diagnosis task is placed first. To handle redundancy, we consider each subsequent ^^ఌin the ordered sequence by posting a conditional question: “Given the 1st ^^ఌ, does the addition of the 2nd ^^ఌincrease the AUC of the first?” If the answer is no, the 2nd ^^ఌis disregarded. If yes, only those tissue pathological characteristics from the 1st and 2nd ^^ఌs that contribute to the increase in AUC are retained. The characteristics from the 1st and 2nd ^^ఌs that do not contribute to the increase in AUC represent the redundancy between the two elastic images and are therefore eliminated. This process is repeated for the 3rd ^^ఌand subsequent ^^ఌs to determine if they can further improve the AUC. After all the elastic images have been evaluated, the final AUC represents the overall efficacy of the adaptive learning and classification (ALC ) prediction of lesion malignancy. This task-driven elimination of redundancy ensures an increase in AUC from a single dataset to multiple datasets. More details on the implementation of this ALC algorithm will be provided in the following section. Fig.8 depicts a specific method for evaluating machine learning models having different energy levels, 1 to n, where n-10. As set forth herein, manager system 100 can adjust weights associated to machine learning models having different energy levels. Fig.8 illustrates the adaptive learning classification (ALC) algorithm and the pkMI system in greater detail, providing a clearer visual guide to all the processes from the left column to the right column. The 1st column depicts the process of generating a set (n=10) of EMIs and then elastic images, {^^ఌ}, at individual energy values ε from acquired LdCT image. The second column shows the process of extracting a set of TPC features (TPCF) from each ^^ఌ. There are various methods to extract TPCFs. This study employs the gray-level co-occurrence matrix

[0030] to extract a set of tissue energy-specific elastic features (TEEFs) from each ^^ఌ, as shown in the third column. Each of these TEEFs can be analyzed using the R package “randomForest” (version 4.7-1.1)

[0031] to evaluate its relative contribution to the AUC for the corresponding ^^ఌ, as shown in the fourth column. The TEEFs at various energy levels as shown in Fig. 8 can define machine learning models having different energy levels, MLM1-MLMn, as set forth herein. In adjusting a weight associated to a machine learning model, manager system 110, e.g., can deactivate machine learning model, e.g., by decommissioning and / or deletion. In adjusting a weight associated to machine learning model, manager system 110 additionally or alternatively can, e.g., scale result data of the various machine learning models so that a prediction result is biased in favor of better performing machine learning model(s). The ALC algorithm starts in the fourth column by reading the ten AUCs of the corresponding ten elastic images, {^^ఌ}. These ten AUCs are ranked in descending order from the highest to the lowest, as shown in the fifth column. The elimination of redundancy among the ranked ten elastic images is performed on the corresponding ten sets of TEEFs, as shown in the sixth and seventh columns. Given the 1st set of TEEFs and its obtained AUC, the algorithm checks if the 2nd set of TEEFs can increase the 1st AUC. If yes, those TEEFs from the 1st and 2nd sets that contribute to the increase in AUC are retained, while other TEEFs are eliminated. The same R package “randomForest” (version 4.7-1.1) is used to evaluate the contribution of each TEEF to the increase of AUC. Given the retained TEEFs, the algorithm then checks if the 3rd set of TEEFs can further increase the AUC. More details of this pair- wise conditional model are provided in the caption of Fig.8. In evaluating the contribution of each TEEF to the increase in AUC, the lesion samples were randomly divided into training, testing, and validation sets with a ratio of 0.6:0.2:0.2, such that the percentage representation of each lesion class (e.g., the ratio of malignant / benign) was maintained in each split. The reported AUC value for the lesion samples was taken as the average of 100 random dataset splits. In reference to Fig.8 there is depicted an overall framework of the presented pkMI for LdCT-enabled diagnosis of in vivo tissues. The ALC approach utilizes the AUC as the FOM applied to n sets of TEEFs. Initially, all n sets of TEEFs are classified and then ranked in descending order based on their AUCs, for example, A, B, C, etc. The process involves conditional classification to enhance AUC values: starting with the calculation of AUCmax[A@B∣A], which determines how set B can improve A’s AUC. Only the best feature set, A1, is retained from this step, discarding the others. Subsequently, AUCmax[A1@C∣A1] is calculated to identify the next best feature set, A2. This conditional classification process is iteratively applied, considering sets up to AUCmax[A2@D|A2], and continues until all n sets of TEEFs are evaluated. This methodology ensures that the highest AUC is achieved progressively, starting from AUCmax[A]. In one example, LdCT screening datasets were collected from two distinct types of lesions: pulmonary nodules or PNs and colorectal polyps or CPs. These datasets were acquired using two different imaging protocols: (i) a fixed X-ray tube current measured in milliampere-seconds (mAs)—FmAs for all patients, and (ii) a modulated mAs (MmAs) that adjusts based on the individual patient’s body shape. For both protocols, the X-ray tube voltage was consistently maintained at 120 kVp. This dual approach allows for a comparison of imaging efficacy and potential impacts on the diagnostic accuracy between standardized and patient-specific imaging settings. Two datasets of nodules were collected from patients identified with at least one nodule each, which an expert group assessed as indeterminate regarding malignant or benign status. Consequently, these patients were recommended for tissue sampling via biopsy. Accordingly, each nodule is documented with a screening CT image accompanied by a pathological report from the biopsy, providing a comprehensive data set for analysis and validation of diagnostic processes. Nodule Dataset#1(MmAs): Total of 68 nodules from 68 patients (52% males, 48% females, age range from 33 to 91 with mean age of 69 years old) were biopsied (50 malignant and 18 benign). The diameter of these nodules ranges from 0.9 to 13cm (mean size of 3.15cm). Nodule Dataset#2(FmAs): Total of 114 nodules from 114 patients (51 males, 63 females, age range from 18 to 96 with mean age of 69 years old) were biopsied (64 malignant and 50 benign). The diameter of these nodules ranges from 0.5 to 8cm (mean size of 1.54cm). The polyp datasets were acquired by MmAs protocol from patients with polyps identified at CT colonography (CTC) screening and underwent subsequent colonoscopy for polypectomy or biopsy. Thus, each polyp has a screening CTC image and the pathological report after the polyp was resected or surgically removed. Among all the patients, 51% are males and 49% are females with age ranges from 45 to 91 years old (mean age of 66 years old). All polyp CTC images were grouped into two categories, depending on the polyp sizes, (i) large polyps with size range from 1 to 3cm; and (ii) small polyps with size<1cm. Polyp Category#1: Total of 182 large polyps. The neoplastic group includes 158 adenomatous polyps (which can develop into cancer) and the non-neoplastic group includes 24 hyperplastic polyps (which will not develop into cancer). Polyp Category#2: Total of 357 small polyps. The neoplastic group includes 266 adenomatous polyps, and the non-neoplastic group includes 91 hyperplastic polyps. Two categories of existing ML algorithms were implemented as references for comparison purposes, one is the traditional ML (TL) and the other is the deep learning with convolutional neural network (DL). A typical TL algorithm is the Haralick texture feature extraction method

[0030] , which uses a gray level co-occurrence matrix to quantitatively compute the image characteristic or texture features, called Haralick features (HFs). A set of HFs was extracted from the LdCT screening volumetric lesion image,^^^^^^,^^, ^^^, and classified by RF R-package

[0031] , see section 2.4 for more details. This reference iscalled RF-HFs. The best results from the four lesion datasets are shown in the 2nd row of Table D-1, wherein there are depicted AUCs (mean ± standard deviation) of the presented pkMI system, compared to four existing ML algorithms using four IDL datasets. The table includes data sample numbers (total samples categorized as malignant / benign or neoplastic / non-neoplastic). Table D-1 Nodules #1 Nodules #2 Large Polyps* (182: Small Polyps* ML Model (68: 50 / 18) (114: 64 / 50) 158 / 24) (357: 266 / 91) 7 0 6 1 34 neo p as c s. o - eopas c. Table D-2: Accuracy (mean ± standard deviation) of the presented pkMI system, compared to four existing ML algorithms using four IDL datasets. ML Nodules #1 Nodules #2 Large Polyps* Small Polyps* Table D-3: Sensitivity (mean ± standard deviation) of the presented pkMI system, compared to four existing ML algorithms using four IDL datasets. ML Nodules #1 Nodules #2 Large Polyps* Small Polyps* Model (68: 50 / 18) (114: 64 / 50) (182: 158 / 24) (357: 266 / 91) to four existing ML algorithms using four IDL datasets. Nodules #1 Nodules #2 Large Polyps* Small Polyps* ML M d l (68: 50 / 18) (114: 64 / 50) (182: 158 / 24) (357: 266 / 91) T hree DL algorithms were adopted as another category of reference for comparison purposes. The 1st one is recently reported

[0032] and labeled as 3D-DL-Clinic. The 2nd one is publicly available

[0033] and labeled as 3D-DL-Public. The 3rd one is a refinement of the 2nd by replacing batch normalization in the first three layers with instance normalization

[0034] and labeled as 3D-DL-IRIS. A3D rectangular mask was applied to contain the volumetric lesion image, ^^^^^^,^^, ^^^, in the center of themask with zero filling in those voxels outside the lesion. The lesion data in the 3D mask were inputted to the 3D-DL algorithms. The architecture of the 3D-DL model contains 17 layers, generated using Tensorflow v2.10. A stratified shuffle split cross validation method was used, where the data was randomly separated into 80% training set and 20% testing set and repeated 50 times. The model was trained for 1000 epochs with the implementation of an early stopping callback strategy and cross- entropy loss function. The input size of the polyps was set to 50×50×50 voxels, with a batch size of 8 and a learning rate of 0.01. Their best outcome results are shown in the 3rd to 5th rows of Table 1. Results of the presented pkMI are shown in the 6th row of Table D. Ten sets of TEEFs were extracted from the corresponding ten elastic images, respectively, of a lesion volume and classified by the ALC method. To show the statistical significances among the five algorithms, the p-value between two algorithms among the five was calculated using a two-tailed t-test program. For example, treating RF- HFs as a reference, the p values for the other three DL algorithms and the proposed pkMI system were computed, respectively. By the same way, treating 3D-DL-Clinic as a reference, the p values for the other two DL algorithms and the proposed pkMI system were computed, respectively. The last computation of p values was performed by treating 3D-DL-IRIS as reference and the p value was calculated for the proposed pkMI system. Tables E(1)-(4) show the p values among the five algorithms on the four datasets, respectively. Table E(1): Cross-model p-values for dataset Nodule #1 (two-tailed t-test) 3D-DL- 3D-DL- RF-HFs 3D-DL-Clinic pkMI Public IRIS a e : ross-mo e p-va ues or a ase o u e wo- a e - es 3D-DL- 3D-DL- 3D-DL- RF-HFs pkMI < < < < - Table E(3): Cross-model p-values for large polyp dataset (two-tailed t-test) 3D-DL- 3D-DL- 3D-DL- RF-HFs pkMI Clinic Public IRIS 3D-DL- 3D-DL- RF-HFs 3D-DL-IRIS pkMI Clini P bli For the two nodule datasets, since each lesion was considered indeterminant by experts, the AUCs of TL in the range of 0.60s are understandable. The DL algorithms increased AUC to 0.70s range. The pkMI increased AUCs from 0.70s to 0.90s, demonstrating an immense potential to overcome the challenge of diagnosing the IDLs. Recent reports highlight the use of artificial intelligence (AI) software to enhance diagnostic accuracy in over 1,000 nodules detected during LdCT screenings, achieving an AUC range of 0.86–0.88 [7,14,15]. These results underscore the inherent challenges in diagnosing nodules. Notably, when nodules characterized as either benign or malignant by expert interpretation, especially those in the late stages of cancer transformation, are excluded, the task of diagnosing the indeterminate lesions or IDLs becomes significantly more difficult. This increased difficulty is reflected in the outcomes presented in Table D, demonstrating a stark contrast in diagnostic precision when less definitive cases (e.g., IDLs) are considered. For the small and large polyp datasets, the pkMI showed similar increases of AUC compared to the results from the two nodule datasets. Diagnosing the small polyps, which are at an early stage of lesion tissue pathological development, presents a particularly challenging task. Despite these challenges, the presented predictive model, pkMI, has performed commendably, achieving an AUC of 0.89. This indicates a high level of accuracy in identifying these early-stage lesions, highlighting the effectiveness of the model in a critical diagnostic context. The following discussion will detail potential alternate embodiments for the pkMI diagnostic tool of the present disclosure. The effectiveness of the predictive model, pkMI, has been successfully tested through a series of experiments involving pathologically confirmed datasets of indeterminate pulmonary nodules and colorectal polyps, which are notably challenging to diagnose. The model's success is quantitatively assessed by the FOM, specifically the AUC, under various testing conditions: (1) different imaging protocols (modulated mAs vs. fixed mAs), (2) a range of sample sizes (from 68 nodules to 357 small polyps), and (3) a diversity of lesion sizes (from 0.5 cm to over 13 cm). When compared with several existing state-of-the-art medical imaging-based ML diagnostic algorithms [7,14,15,32-34], pkMI has shown significant potential. It stands out for its ability to address the critical challenge of accurately diagnosing IDL types, thereby facilitating more effective LdCT screening for early detection of lung and colorectal cancers. This demonstrates pkMI’s potential as a transformative tool in the realm of medical diagnostics, especially in the early intervention and management of these cancers. The pkMI employs quantitative models, e.g., Eq. (9), that transform prior knowledge into mathematical formulas based on established basic science principles, which are inherently consistent. Consequently, its performance is expected to be both robust and accurate, provided that the mathematical formulas are precisely modeled. The consistency and precision demonstrated in a series of experiments confirm the pkMI's robustness and accuracy for diagnosing nodules and polyps. In reference again to Fig.7C, Fig.7C depicts an embodiment of a predictive model 7002 provided by a neural network based predictive model. In reference to predictive model 7002 as shown in Fig.7C, machine learning model MLM17003 can be provided by artificial neural network (ANN) model 7501. Referring to predictive model 7002 as shown in Fig.7C, respective ones of MLM1-MLMn can be provided by respective ANNs 7501. Manager system 110 can train respective ones of MLM1 7003 to MLMn 7005 using training data defined by elastic image data. For example, MLM17003 can be trained with elastic image training data defined by iterations of elastic image ^^ఌ^derived by processing of historical instances of images associated to pathologically examined tissue, MLM2 can be trained with elastic image training data defined by iterations of elastic image ^^ఌଶ, , and machine learning model MLMn can be trained with elastic image training data defined by elastic image ^^ఌ^. Fig.9 depicts training of an ANN 7501 of respective ones of machine learning models MLM1 7003 to MLMn 7005 in the embodiment of Fig.7C where machine learning models are provided by artificial neural networks. ANN 7501 can be trained with iterations of elastic image data defined by an elastic image of a specified energy level. Iterations of training data of training ANN 7501 can include, as input training data, energy level for an applied elastic image in combination with elastic image voxel values defining the elastic image being applied as training data. The described training data can be applied for multiple iterations of elastic images derived from historical images representing pathologically examined tissue. Training data can further include, as outcome training data, a label associated to the elastic image data applied as input training data, namely the label input can include, e.g., the pathologically determined diagnosis of malignant or benign. Inferencing data for inferencing ANN 7501 can include an energy level of an applied elastic image derived from a current patient in combination with elastic image voxel values defining the applied elastic image derived from a current patient. On application of the described inferencing data, ANN 7501 can output of predicted label, e.g., the predicted label malignant or benign. Various available tools, libraries, and / or services can be utilized for the implementation of trained predictive models herein trained by machine learning. For example, a machine learning service can provide access to libraries and executable code for the support of machine learning functions. A machine learning service can provide access to a set of Representative State Transfer (REST) APIs that can be called from any programming language and that permit the integration of predictive analytics into any application. Enabled REST APIs can provide, for example, retrieval of metadata for a given predictive model, deployment of models and management of deployed models, online deployment, scoring, batch deployment, stream deployment, monitoring, and retraining deployed models. According to one possible implementation, a machine learning service can provide access to a set of REST APIs that can be called from any programming language and that permit the integration of predictive analytics into any application. Enabled REST APIs can provide, for example, retrieval of metadata for a given predictive model, deployment of models and management of deployed models, online deployment, scoring, batch deployment, stream deployment, monitoring, and retraining deployed models. Trained predictive models herein can employ the use, for example, of artificial neural networks (ANNs), support vector machines (SVM), Bayesian networks, regression-based models, and / or other machine learning technologies. Fig.10 is an illustration of an example ANN architecture for trained predictive models herein. One element of ANNs can be the structure of the information processing system, which can include a large number of highly interconnected processing elements (called “neurons”) working in parallel to solve specific problems. ANNs can be trained using a set of training data, with learning that involves adjustments to weights that exist between the neurons. Referring now to Fig.10, a generalized diagram of a neural network can be shown. Although a specific structure of an ANN can be shown, having three layers and a set number of fully connected neurons, this is intended solely for the purpose of illustration. In practice, the present embodiments can take any appropriate form, including any number of layers and any pattern or patterns of connections therebetween. ANNs can demonstrate an ability to derive meaning from complicated or imprecise data and can be used to extract patterns and detect trends that can be too complex to be detected by humans or other computer-based systems. The structure of a neural network can generally have input neurons 1002 that can provide information to one or more “hidden” neurons 1004. Weighted connections 1008 between the input neurons 1002 and hidden neurons 1004 can be weighted, and these weighted inputs can be processed by the hidden neurons 1004 according to some function in the hidden neurons 1004. There can be any number of layers of hidden neurons 1004, and neurons that perform different functions. There can also exist different neural network structures, such as a convolutional neural network, a maxout network, etc., which can vary according to the structure and function of the hidden layers, as well as the pattern of weights between the layers. The individual layers can perform particular functions and can include convolutional layers, pooling layers, fully connected layers, softmax layers, or any other appropriate type of neural network layer. Finally, a set of output neurons 1006 can accept and process weighted input from the last set of hidden neurons 1004. This can represent a “feed-forward” computation, where information propagates from input neurons 1002 to the output neurons 1006. Upon completion of a feed-forward computation, the output can be compared to a desired output available from training data. The error relative to the training data can then be processed in “backpropagation” computation, where the hidden neurons 1004 and input neurons 1002 can receive information regarding the error propagating backward from the output neurons 1006. Once the backward error propagation has been completed, weight updates can be performed, with the weighted connections 1008 being updated to account for the received error. It can be noted that the three modes of operation—feed forward, backpropagation, and weight update—do not overlap with one another. This can represent just one variety of ANN computation, and any appropriate form of computation can be used instead. To train an ANN, training data can be divided into a training set and a testing set. The training data can include pairs of an input and a known output, which can be referred to as outcome training data as referenced in connection with predictive models herein. During training, the inputs of the training set can be fed into the ANN using feed-forward propagation. After each input, the output of the ANN can be compared to the respective known output. Discrepancies between the output of the ANN and the known output that can be associated with that particular input can be used to generate an error value, which can be backpropagated through the ANN, after which the weight values of the ANN can be updated. This process can continue until the pairs in the training set can be exhausted. After the training has been completed, the ANN can be tested against the testing set to ensure that the training has not resulted in overfitting. If the ANN can generalize to new inputs beyond those on which it was trained, it can be ready for use. If the ANN does not accurately reproduce the known outputs of the testing set, additional training data may be appropriate, or hyperparameters of the ANN may need to be adjusted. ANNs can be implemented in software, hardware, or a combination of the two. For example, the weights of the weighted connections 1008 can be characterized as a weight value that can be stored in a computer memory, and the activation function of each neuron can be implemented by a computer processor. The weight value can store any appropriate data value, such as a real number, a binary value, or a value selected from a fixed number of possibilities, that can be multiplied against the relevant neuron outputs. Alternatively, the weights of the weighted connections 1008 can be implemented as resistive processing units (RPUs), generating a predictable current output when an input voltage can be applied in accordance with a settable resistance. For small polyps, which are at an early stage of lesion tissue pathological development, more sophisticated modeling strategies may be appropriate. As photon-counting CT (PCCT) technology progresses, it can reconstruct images with enhanced energy information, allowing for more precise tissue mixture segmentation within each voxel to generate EMIs. Improved extraction of tissue biological features from these detailed EMIs is expected to enhance the diagnostic capabilities of low- dose PCCT in detecting early cancers. Additionally, incorporating more specific measures of lesion malignancy, such as heterogeneity and the dynamics of lesion invasion and aggression, through the ALC can further refine the potential of the current pkMI model. In this disclosure, exemplary embodiments focused only on the 1st and 2nd order derivatives to capture characteristic features of the in vivo tissuedynamical function or elastic images ^^ఌ^^^,^^, ^^^. It will be appreciated that 3rd and higher orderderivatives could reveal additional characteristic information. It is anticipated that more accurate modeling will lead to better performance of the pkMI. The pkMI outlined offers a framework for investigating quantitative imaging biomarkers and radiomic features. Utilizing the ALC strategy enables the integration of all imaging biomarkers and radiomic features while eliminating data redundancy. This approach of integrating various datasets aims to optimize performance in diagnosing in vivo tissues. Current developments in machine learning primarily focus on emulating the anatomical structure of human neural networks, often overlooking the functionality of these architectures. For instance, encoding X-ray energy information into EMIs and translating dynamic information of in vivo tissues into biological (elastic) images as outlined in Eq. (9) exemplify how the functionality of neural network architecture can be integrated into our models. Adapting the CT-based pkMI to other imaging modalities, such as MRI, for lesion diagnosis represents another area of alternate embodiments of the present disclosure. Low-dose computed tomography (LdCT) screening has significantly advanced early lung cancer detection, enabling both medical experts and machine learning (ML) algorithms to effectively identify and delineate pulmonary nodules (PNs). Despite these advancements, distinguishing malignant PNs from those detected during screening remains a complex challenge, primarily due to the difficulty in diagnosing PNs classified as indeterminate lesions (IDLs) by medical experts. We hypothesize that this difficulty arises from the current data-driven ML algorithms' lack of integration of prior knowledge beyond the data. To address this, we present a machine intelligence system that incorporates prior knowledge into data analysis to resolve IDLs. One key piece of prior knowledge relates to X-ray energy spectrum, where different energies interact with in vivo tissues within a lesion and generate variable but reproducible image contrasts across the lesion, encapsulating tissue biological information in the image contrast variations. Typically, CT imaging devices utilize only the high-energy portion of this spectrum for data acquisition. This disclosure considers the full spectrum for lesion diagnostics. Another critical piece of prior knowledge involves dynamic or functional properties of in vivo tissues, such as elasticity and growth rate, which are indicative of pathological conditions. Embodiments disclosed in this disclosure extract these tissue pathological characteristics from the image contrast variations to diagnose IDLs, rather than relying solely on abstract image features as current ML algorithms do. The current machine intelligence system was evaluated on LdCT images of four sets of IDLs, including PNs and colorectal polyps, with their pathological reports serving as ground truth for malignancy. The outcomes achieved an AUC (area under the receiver operating characteristic curve) of 0.98, demonstrating a significant improvement over existing state-of-the-art ML algorithms, which have AUCs in the 0.70 range. Embodiments of the present disclosure provide a machine intelligence system that incorporates prior knowledge into data analysis to resolve IDLs. Two examples of prior knowledge to be incorporated to perform the task of resolving IDLs are mentioned here and more will be described later. One key piece of prior knowledge relates to X-ray energy spectrum. Within the spectrum, varying energies interact differently with in vivo tissues inside a lesion and generate specific and reproducible image contrasts across the lesion volume. Thus, tissue biological information is encapsulated in the image contrast variations. Typically, CT imaging devices utilize only the high-energy portion of this spectrum for data acquisition and image reconstruction. However, the used high-energy portion generates less image contrasts compared to the low energy portion in the spectrum. The present embodiments consider the full spectrum to more comprehensively characterize in vivo tissues and aid in lesion diagnostics. Another critical piece of prior knowledge involves dynamic or functional properties of in vivo tissues, such as elasticity and growth rate, which are indicative of pathological conditions. The present embodiments employ a mathematical model to relate these tissue dynamic properties to the image contrast variations and extracts the corresponding tissue pathological characteristics (TPCs) from the image contrast variations to resolve or diagnose IDLs, moving beyond the reliance on abstract image features as is common in current ML algorithms. Processes described herein may be performed by one or more computer systems or other processing devices, as provided, in a server as set forth herein. An examplary computer system to incorporate and use aspects described herein is depicted and described with reference to Fig.10. Computer system 500 can include one or more processor(s) 502, memory 504, and one or more I / O device 506, which may be coupled to each other by busses and other electrical hardware elements (not depicted). Processor(s) 502 may include any appropriate hardware component(s) capable of implementing functions, for instance executing instruction(s) (sometimes alternatively referred to as code, firmware and / or software) retrieved from memory 504. Execution of the instructions causes the computer system 500 to perform processes, functions, or the like, such as those described with reference to the methods set forth herein. One or more computer systems configured according to computer system 500 can host one or more virtual machine (VM) such as a hypervisor based virtual machine, and / or a container based virtual machine as set forth herein. In some examples, aspects described herein are performed by a plurality of homogenous or heterogeneous computer systems coordinated to collectively perform processes, functions, or the like, such as those described herein. Memory 504 can include hardware components or other storage devices to store data such as programs of instructions for execution, and other data. The storage devices may be magnetic, optical, and / or electrical based, as examples. Hard drives, field-programmable gate arrays (FPGAs), magnetic media, compact disks (CDs), digital versatile disks (DVDs), and flash memories are example storage devices. Accordingly, memory 504 may be volatile, non-volatile, or a combination of the two. As a specific example, memory 504 includes one or more hard drives and one or more random-access memory (RAM) devices for, respectively, non-volatile, and volatile storage of data. Example programs stored by memory include an operating system and applications that run on the operating system, such as specialized applications to perform functions described herein. Memory 504 can define a computer readable storage medium. A computer readable storage medium, as used herein, is not to be interpreted as being transitory signals per se. There is set forth herein a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a one or more processor to carry out methods and / or functions set forth herein. I / O device(s) 506 can include hardware and / or software components that support input and output of data to / from computer system 500. I / O device(s) 506 include physical components that attach physically or wirelessly to the computer system. I / O device(s) 506 can also include, but are not limited to, I / O controllers and hardware and software supporting data communication with the aforementioned components, such as network, graphics, and / or audio controller(s). An example I / O device 506 is a network adapter for communication of data between computer system 500 and another component, such as another computer system, across communication links. Examples include Ethernet, cable, WiFi, cellular and / or fiber-based communications links passing data packets between computer system 500 and other systems across one or more networks, such as the Internet. Other example I / O devices 506 include universal serial bus (USB), peripheral component interconnect (PCI), and serial adapters / interfaces configured to couple to devices of their respective kind. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprise” (and any form of comprise, such as “comprises” and “comprising”), “have” (and any form of have, such as “has” and “having”), “include” (and any form of include, such as “includes” and “including”), “contain” (and any form contain, such as “contains” and “containing”), and any other grammatical variant thereof, are open-ended linking verbs. As a result, a method or article that “comprises”, “has”, “includes” or “contains” one or more steps or elements possesses those one or more steps or elements but is not limited to possessing only those one or more steps or elements. Likewise, a step of a method or an element of an article that “comprises”, “has”, “includes” or “contains” one or more features possesses those one or more features but is not limited to possessing only those one or more features. Terms like “obtainable” or “definable” and “obtained” or “defined” are used interchangeably. This, for example, means that, unless the context clearly dictates otherwise, the term “obtained” does not mean to indicate that, for example, an embodiment must be obtained by, for example, the sequence of steps following the term “obtained” though such a limited understanding is always included by the terms “obtained” or “defined” as a preferred embodiment. It should be appreciated that all combinations of the foregoing concepts and additional concepts discussed in greater detail below (provided such concepts are not mutually inconsistent) are contemplated as being part of the subject matter disclosed herein. In particular, all combinations of claims subject matter appearing at the end of this disclosure are contemplated as being part of the subject matter disclosed herein. It should also be appreciated that terminology explicitly employed herein that also may appear in any disclosure incorporated by reference should be accorded a meaning most consistent with the particular concepts disclosed herein. This written description uses examples to disclose the subject matter, and also to enable any person skilled in the art to practice the subject matter, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the subject matter is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insubstantial differences from the literal languages of the claims. It is to be understood that the above description is intended to be illustrative, and not restrictive. For example, the above-described examples (and / or aspects thereof) may be used in combination with each other. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the various examples without departing from their scope. While the dimensions and types of materials described herein are intended to define the parameters of the various examples, they are by no means limiting and are merely exemplary. Many other examples will be apparent to those of skill in the art upon reviewing the above description. The scope of the various examples should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled. In the appended claims, the terms “including” and “in which” are used as the plain-English equivalents of the respective terms “comprising” and “wherein.” Moreover, in the following claims, the terms “first,” “second,” and “third,” etc. are used merely as labels, and are not intended to impose numerical requirements on their objects. Forms of term “based on” herein encompass relationships where an element is partially based on as well as relationships where an element is entirely based on. Forms of the term “defined” encompass relationships where an element is partially defined as well as relationships where an element is entirely defined. Further, the limitations of the following claims are not written in means-plus-function format and are not intended to be interpreted based on 35 U.S.C. § 112(f) unless and until such claim limitations expressly use the phrase “means for” followed by a statement of function void of further structure. It is to be understood that not necessarily all such objects or advantages described above may be achieved in accordance with any particular example. Thus, for example, those skilled in the art will recognize that the systems and techniques described herein may be embodied or carried out in a manner that achieves or optimizes one advantage or group of advantages as taught herein without necessarily achieving other objects or advantages as may be taught or suggested herein. The terms “substantially”, “approximately”, “about”, “relatively”, or other such similar terms that may be used throughout this disclosure, including the claims, are used to describe and account for small fluctuations, such as due to variations in processing, from a reference or parameter. Such small fluctuations include a zero fluctuation from the reference or parameter as well. For example, they can refer to less than or equal to ± 10%, such as less than or equal to ± 5%, such as less than or equal to ± 2%, such as less than or equal to ± 1%, such as less than or equal to ± 0.5%, such as less than or equal to ± 0.2%, such as less than or equal to ± 0.1%, such as less than or equal to ± 0.05%. If used herein, the terms “substantially”, “approximately”, “about”, “relatively,” or other such similar terms may also refer to no fluctuations, that is, ± 0%. It is contemplated that numerical values, as well as other values that are recited herein can be modified by the term “about”, whether expressly stated or inherently derived by the discussion of the present disclosure. Further, any description of a range herein can encompass all subranges. The terms “connect,” “connected,” “contact” “coupled” and / or the like are broadly defined herein to encompass a variety of divergent arrangements and assembly techniques. These arrangements and techniques include, but are not limited to (1) the direct joining of one component and another component with no intervening components therebetween (i.e., the components are in direct physical contact); and (2) the joining of one component and another component with one or more components therebetween, provided that the one component being “connected to” or “contacting” or “coupled to” the other component is somehow in operative communication (e.g., electrically, physically, optically, etc.) with the other component (notwithstanding the presence of one or more additional components therebetween). It is to be understood that some components that are in direct physical contact with one another may or may not be in electrical contact with one another. Moreover, two components that are electrically connected, electrically coupled, optically connected, optically coupled, may or may not be in direct physical contact, and one or more other components may be positioned therebetween. While the subject matter has been described in detail in connection with only a limited number of examples, it should be readily understood that the subject matter is not limited to such disclosed examples. Rather, the subject matter can be modified to incorporate any number of variations, alterations, substitutions or equivalent arrangements not heretofore described, but which are commensurate with the spirit and scope of the subject matter. Additionally, while various examples of the subject matter have been described, it is to be understood that aspects of the disclosure may include only some of the described examples. Also, while some examples are described as having a certain number of elements it will be understood that the subject matter can be practiced with less than or greater than the certain number of elements. Accordingly, the subject matter is not to be seen as limited by the foregoing description but is only limited by the scope of the appended claims. All publications cited in this specification are herein incorporated by reference as if each individual publication were specifically and individually indicated to be incorporated by reference herein as though fully set forth. Where one or more ranges are referred to throughout this specification, each range is intended to be a shorthand format for presenting information, where the range is understood to encompass each discrete point within the range as if the same were fully set forth herein. While several aspects and embodiments of the present disclosure have been described and depicted herein, alternative aspects and embodiments may be affected by persons having ordinary skills in the art to accomplish the same objectives. Accordingly, this disclosure and the appended claims are intended to cover all such further and alternative aspects and embodiments as fall within the true spirit and scope of the present disclosure. The following references cited herein may be relevant to the present disclosure and are incorporated herein by reference in their entireties and a skilled person is considered to be aware of disclosure of these references: [1] D.R. Aberle, A.M. Adams, A.D. Berg, et al. “Reduced lung-cancer mortality with low- dose CT screening.” New England Journal of Medicine, 365: 395-409, 2011. [2] H.J. Freeman. “Heterogeneity of colorectal adenomas, the serrated adenoma, and implications for screening and surveillance.” World Journal on Gastroenterology, 4: 3461-3463, 2008. [3] S. Chicklore, V. Goh, M. Siddique, et al. “Quantifying tumor heterogeneity in 18F-FDG PET / CT imaging by texture analysis.” European Journal of Nuclear Medicine and Molecular Imaging, 40: 133-140, 2013. [4] J.P.B. O’Connor, C.J. Rose, J.C. Waterton, et al. “Imaging intra tumor heterogeneity: Role in therapy response, resistance, and clinical outcome.” Clinical Cancer Research, 21: 249-257, 2015. [5] U. Bashir, M.M. Siddique, E. Mclean, et al. “Imaging heterogeneity in lung cancer: Techniques, applications, and challenges.” American Journal of Roentgenology, 207: 534-543, 2016. [6] R.A. Gatenby, O. Grove, and R.J. Gillies. “Quantitative imaging in cancer evolution and ecology.” Radiology, 269: 8-15, 2013. [7] S.J. Adams, D.K. Madtes, B. Burbridge, et al. “Clinical impact and generalizability of a computer-assisted diagnostic tool to risk-stratify lung nodules with CT.” Journal of the American College of Radiology, 20: p232-242, 2022. [8] W.C. Black, I.F. Gareen, S.S. Soneji, et al. “Cost-effectiveness of CT screening in the national lung screening trial.” New England Journal of Medicine, 371: 1793-1802, 2014. [9] T. Lokhandwala, M.A. Bittoni, R.A. Dann, et al. “Costs of diagnostic assessment for lung cancer: A Medicare claims analysis.” Clinic Lung Cancer, 18: e27-34, 2017.

[0010] A. McWilliams, M.C. Tammemagi, J.R. Mayo, et al. “Probability of cancer in pulmonary nodules detected on first screening CT.” New England Journal of Medicine, 369: 910-919, 2013.

[0011] F. Han, H. Wang, G. Zhang, et al. “Texture feature analysis for CADx on pulmonary nodules.” Journal of Digital Imaging, 28: 99-115, 2015.

[0012] H. Wang, T. Zhao, L.C. Li, et al. (2017), “A hybrid CNN feature model for pulmonary nodule malignancy risk differentiation.” Journal of X-Ray Science and Technology, 25: 673-6882017.

[0013] D. Ardila, A. Kiraly, S. Bharadwaj, et. al. “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography.” Nature Medicine, 25: 954-61, 2019.

[0014] D.R. Baldwin, J. Gustafson, L. Pickup, et al. “External validation of a convolutional neural network artificial intelligence tool to predict malignancy in pulmonary nodules.” Thorax, 75: 306-312, 2020.

[0015] P.P. Massion, S. Antic, S. Ather, et al. “Assessing the accuracy of a deep learning method to risk stratify indeterminate pulmonary nodules.” American Journal of Respiratory and Critical Care Medicine, 202: 241-249, 2020.

[0016] The National Institute of Standards and Technology (NIST, https: / / physics.nist.gov / PhysRefData / XrayMassCoef / tab2.html).

[0017] L. Li, D. Chen, Lakare, et al. “An image segmentation approach to extract colon lumen through colonic material tagging and hidden Markov random field model for virtual colonoscopy.” SPIE Medical Imaging, 4683: 1605-7422, 2002.

[0018] H. Han, L. Li, F. Han, et al. “Fast and adaptive detection of pulmonary nodules in thoracic CT images using a hierarchical vector quantization scheme.” IEEE Journal of Biomedical and Health Informatics, 19(2): 648-659, 2015.

[0019] L. Yu, S. Leng and C.H. “Dual-energy CT-based monochromatic imaging.” American Journal of Roentgenology, 199: S9-S15, 2012.

[0020] J. Sosna. “Photon-counting detector CT: Initial analysis of virtual noncontrast imaging.” Radiology, https: / / doi.org / 10.1148 / radiol.221173, 2022.

[0021] S. Chang, Y. Gao, M.J. Pomeroy, et al. “Exploring dual-energy CT spectral information for machine learning-driven lesion diagnosis in pre-log domain.” IEEE Transactions on Medical Imaging, 42: 1835-1845, 2023.

[0022] H. Li, X. Huang, and L. He. “Object matching using a locally Affine invariant and linear programming techniques.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 35: 411-424, 2013.

[0023] F.P. Beer, E.R. Johnston, J.T. DeWolf, et al. Mechanics of Materials. McGraw Hill. ISBN 978-0-07-015389-9, 2009.

[0024] T. Panchal, H. Patel, and A. Panchal. “License plate detection using Harris corner and character segmentation by integrated approach from an image.” Procedia Computer Science, 79: 415- 425, 2016.

[0025] B. O’Neill. Elementary Differential Geometry (revised second edition). Waltham, Massachusetts: Academic Press Elsevier, 2006.

[0026] M. Mirzakhani and A. Wright. “The boundary of an Affine invariant submanifold.” Inventiones Mathematicae, 209: 927-984.2017.

[0027] J. Kostkova, T. Suk and J. Flusser. “Affine invariants of vector fields.” IEEE Transactions on Pattern Analysis and Machine Intelligence (Early Access), 2019. DOI: 10.1109 / TPAMI.2019.2951664.

[0028] W. Cao, M.J. Pomeroy, Z. Liang Z, et al. “Lesion classification by model-based feature extraction: A differential Affine invariant model of soft tissue elasticity.” arXiv: Medical Physics. 2205.14029, 2022.

[0029] C.R. Taylor. Dynamic Programming and the Curses of Dimensionality. Applications of Dynamic Programming to Agricultural Decision Problems. Westview Press.1993.

[0030] R.M. Haralick, K. Shanmugam, and I.H. Dinstein. “Textural features for image classification.” IEEE Transactions on Systems, Man and Cybernetics, 3: 610-621, 1973.

[0031] A. Liaw and M. Wiener. Breiman and Cutler’s Random Forests for Classification and Regression.2022 on-line: https: / / cran.r-project.org / web / packages / randomForest / .

[0032] P. Wesp, S. Grosu, A. Graser, et al. “Deep learning in CT colonography: Differentiating premalignant from benign colorectal polyps.” European Radiology, 32: 4749-4759, 2022.

[0033] H. Zunair, A. Rahman, N. Mohammed, et al. “Uniformizing techniques to process CT scans with 3D CNNs for tuberculosis prediction.” arXiv:2007.13224, 2020.

[0034] L. Kuo, M.J. Pomeroy, J. Tan, et al. “Development of 3D CNN models for colorectal polyp classification in CT colonography.” Presented at the 65th Annual Meeting & Exhibition of AAPM, July 23-27, 2023.

[0035] Z. Liang and S. Wang. “An EM approach to MAP solution of segmenting tissue mixtures: A numerical analysis.” IEEE Transactions on Medical Imaging, 28: 297-310, 2009.

[0036] W. Cao, Z. Liang, Y. Gao, et al. “A dynamic lesion model for differentiation of malignant and benign pathologies.” Scientific Reports, 11: Article number: 3485, 2021.

[0037] Z. Liang, Y. Gao, S. Chang, et al. “A new look at artificial intelligence and machine learning for computer-aided diagnosis.” Presented at the 108th Annual Meeting of RSNA, November 27-December 1, 2022.

Claims

What is claimed is:

1. A computer implemented method comprising: decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing.

2. The computer implemented method of claim 1, wherein the returning the one or more action decision includes returning an action decision for remediation of a condition of the patient tissue section, wherein the condition of the patient tissue section is detected in dependence on the inferencing.

3. The computer implemented method of claim 1, wherein the predictive model includes a first machine learning model trained with training data having the first energy level, and a second machine learning model trained with training data having the second energy level.

4. The computer implemented method of claim 3, wherein the method further comprises evaluating performance of the first machine learning model and the second machine learning model, and adjusting a weight associated to one or more of the first machine learning model or the second machine learning model responsively to the evaluating.

5. The computer implemented method of claim 3, further comprising evaluating area under curve (AUC) performance of the first machine learning model and the second machine learning model, and adjusting a weight associated to one or more of the first machine learning model or the second machine learning model responsively to the evaluating.

6. The computer implemented method of claim 5, wherein the adjusting a weight includes deactivating the first machine learning model.

7. The computer implemented method of claim 5, wherein the adjusting includes adjusting a volume of training data applied to one or more of the first machine learning model or the second machine learning model.

8. The computer implemented method of claim 5, wherein the adjusting includes scaling result data output by one or more of the first machine learning model or the second machine learning model.

9. The computer implemented method of claim 3, wherein the first machine learning model includes a random forest model and a feature extractor, and wherein the second machine learning model and a feature extractor.

10. The computer implemented method of claim 3, wherein the first machine learning model includes an artificial neural network, and wherein the second machine learning model includes an artificial neural network.

11. The computer implemented method of claim 1, wherein the transforming includes, for providing the first elastic image, deriving a first derivative of the first effective mono-energy image having the first energy level.

12. The computer implemented method of claim 1, wherein the transforming includes, for providing the first elastic image, deriving a second derivative of the first effective mono-energy image having the first energy level.

13. The computer implemented method of claim 1, wherein the transforming includes, for providing the first elastic image, deriving a first derivative of the first effective mono-energy image having the first energy level, and deriving a second derivative of the first effective mono-energy image having the first energy level.

14. The computer implemented method of claim 1, wherein the returning the one or more action decision in dependence on the inferencing includes returning an action decision to present a visualization of the patient image with a result of the inferencing highlighted.

15. The computer implemented method of claim 1, wherein the returning the one or more action decision in dependence on the inferencing includes returning an action decision for controlling a treatment process for treating a condition of the patient tissue section detected responsively to the inferencing.

16. The computer implemented method of claim 3, wherein: the returning the one or more action decision includes returning an action decision for remediation of a condition of the patient tissue section;wherein the condition of the patient tissue section is detected in dependence on the inferencing; wherein the method includes evaluating area under curve (AUC) performance of the first machine learning model and the second machine learning model, and adjusting a weight associated to one or more of the first machine learning model or the second machine learning model responsively to the evaluating; wherein the first machine learning model includes a random forest model and a feature extractor; and wherein the transforming includes, for providing the first elastic image, deriving a first derivative of the first effective mono-energy image having the first energy level, and deriving a second derivative of the first effective mono-energy image having the first energy level.

17. The computer implemented method of claim 3, wherein: the returning the one or more action decision includes returning an action decision for remediation of a condition of the patient tissue section, wherein the condition of the patient tissue section is detected in dependence on the inferencing; the method further comprises evaluating area under curve (AUC) performance of the first machine learning model and the second machine learning model, and adjusting a weight associated to one or more of the first machine learning model or the second machine learning model responsively to the evaluating; the first machine learning model includes a random forest model and a feature extractor; the transforming includes, for providing the first elastic image, deriving a first derivative of the first effective mono-energy image having the first energy level, and deriving a second derivative of the first effective mono-energy image having the first energy level; and wherein the adjusting includes at least one of adjusting a volume of training data applied to one or more of the first machine learning model or the second machine learning model and scaling result data output by one or more of the first machine learning model or the second machine learning model.

18. The computer implemented method of claim 17, wherein returning the one or more action decision in dependence on the inferencing includes returning an action decision to present a visualization of the patient image with a result of the inferencing highlighted, wherein the returningthe one or more action decision in dependence on the inferencing includes returning an action decision for controlling a treatment process for treating a condition of the patient tissue section detected responsively to the inferencing.

19. A system comprising: a memory; at least one processor in communication with the memory; and program instructions executable by one or more processor via the memory to perform a method comprising: decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level and the second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing.

20. A non-transitory computer readable storage medium storing instructions for execution a computer processor for performing a method comprising: decomposing a patient image representing a patient tissue section into a plurality of effective mono-energy images, wherein the plurality of effective mono-energy images include a first effective mono-energy image having a first energy level and a second effective mono-energy image having a second energy level; transforming the first and second effective mono-energy image into respective first and second elastic images, wherein the first and second elastic images represent dynamic properties of tissue elasticity of the patient tissue section at the first energy level andthe second energy level respectively; inferencing a trained predictive model, wherein the inferencing includes inputting, as inferencing data, the first and second elastic images into the predictive model for return of result data, wherein the trained predictive model has been trained with historical pathology sample images; and returning one or more action decision in dependence on the inferencing.

Images