Non-contrast computer tomography with ai-denoiser and dynamic range correction for net water uptake assessment

Abstract

An exemplary non-contrast computer tomography (CT) system and method are disclosed employing AI-denoiser and dynamic range correction that remaps pixels on a portion of the brain to preserve stroke or net water uptake information in the image to accurately generate a whole brain image that readily show changes in net water uptake that could used to identify earlier stage or outset of a stroke or like injuries or conditions. The exemplary system and method can additionally provide quantification of net water uptake, e.g., for stroke diagnosis, in addition to visualization. The AI-denoiser and dynamic range correction may operate with an image inhomogeneity correction operation that can correct for subtle residual artifacts (e.g., cupping artifacts) that increases with skull thickness.

Description

NON-CONTRAST COMPUTER TOMOGRAPHY WITH AI-DENOISER ANDDYNAMIC RANGE CORRECTION FOR NET WATER UPTAKE ASSESSMENTRelated Application

[0001] This is a PCT application that claims priority to, and the benefit of, U.S. Provisional Patent Application No. 63 / 589,199, filed October 10, 2023, which is incorporated by reference herein in its entirety.Background

[0002] A stroke comes on when your brain doesn't get the blood and oxygen it needs. That could be due to a clot, known as an ischemic stroke, or it can happen with a burst blood vessel, as with a hemorrhagic stroke. The first few hours following a stroke can be critical as the availability of treatment options can change drastically. For an ischemic stroke, a clot-busting drug may be given to a patient. For hemorrhagic stroke, surgery may be warranted to repair a broken blood vessel.

[0003] X-ray Computed tomography (CT) is often employed for most head-related injuries or conditions as they are more affordable and widely accessible, especially in emergency care settings. While magnetic resonance imaging and fluoroscopy imaging may more readily identify or quantify stroke-related symptoms or injuries, they are often only employed when strokes are already identified. Existing CT systems have a low resolution in the visualization of stroke and stroke -related injuries or conditions. A doctor may order a head CT for a patient who is experiencing unexplained headaches or dizziness.

[0004] There is a benefit to improving CT imaging for stroke and other head injuries and in being able to identify early presentation of stroke and various head injuries.Summary

[0005] An exemplary non-contrast computer tomography (CT) system and method are disclosed employing Al-denoiser and dynamic range correction that remaps pixels on a portion of the brain to preserve stroke or net water uptake information in the image to accurately generate a whole brain image that readily shows changes in net water uptake that could be used to identify earlier stage or outset of a stroke or like injuries or conditions. The exemplary system and method can additionally provide quantification of net water uptake, e.g., for stroke diagnosis, in addition to visualization. The Al-denoiser and dynamic range correction may operate with an image inhomogeneity correction operation that can correct for subtle residual artifacts (e.g., cupping artifacts) that increase with skull thickness.

[0006] The time of stroke onset can be challenging to define, and even when known, it may not reflect tissue viability. Some patients (“fast progressors”) may present early with large irreversible lesions, while others (“slow progressors”) may still have a sizeable amount of salvageable tissue later. The exemplary system and method can provide visualization and quantification of such early lesions by correcting for the subtle residual artifacts (e.g., cupping artifacts) that increase with skull thickness. The exemplary image inhomogeneity correction operation can also preserve contrast in the ischemic stroke region, including those in large vessel occlusion patients.

[0007] In addition, the exemplary denoiser model shows a significant increase in contrast-to-noise by more than 2-fold and noise reduction by 2.4 to 2.8-fold, while the gray / white matter contrast is simultaneously improved as compared to state-of-the-art imaging. The exemplary denoiser model can be applied to an NCCT of the brain to remove image noise.

[0008] In an aspect, a method is disclosed comprising receiving, by a processor, a whole brain CT image; generating, by the processor, a denoised whole-brain CT image from the received whole-brain CT image using a trained Al model (e.g., neural network) configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15); correcting, by the processor (e.g., via a fully automated NWU brain mapping), the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image, wherein the correction employs one or more segmented regions of the denoised CT image consisting of healthy tissue (e.g., grey matter) in the denoised whole-brain CT image (e.g., wherein the healthy tissue is determined from a tissue probability maps having defined regions for gray matter, white matter, cerebrospinal fluid); and outputting, by the processor, the remapped denoised whole brain CT image in a report or graphical user interface, wherein the remapped denoised whole brain CT image preserves information for disease / injury if present in the whole brain CT image.

[0009] In some embodiments, the method described herein further comprises adjusting, by the processor, the denoised whole-brain CT image prior to the correction via an image inhomogeneity correction operation (e.g., employing a skull-shape-based bias field correction algorithm), wherein the image inhomogeneity correction operation applies a bias correction factor k derived from a distance map to a skull surface (e.g., inner surface, outer surface, or therebetween) (e.g., to correct for residual cupping artifacts attributable to skull thickness).

[0010] In some embodiments, the method described herein further comprises determining, by the processor, a fluid-uptake region (swelling) in the remapped denoised whole brain CT image having uptake of water or fluids in the brain; and outputting, by the processor,visualization of the fluid-uptake region via the report or graphical user interface, wherein the visualization is overlaid over, or fused into, the remapped denoised whole brain CT image or a portion thereof.

[0011] In some embodiments, the method described herein further comprises determining, by the processor, a quantification value of the fluid-uptake region (swelling) in the remapped denoised whole brain CT image; and outputting, by the processor, the quantification value of the fluid-uptake region via the report or graphical user interface.

[0012] In some embodiments, the quantification value NWU is determined byNWU = 1 — Diesion / Dnormai, wherein Diesionis a mean CT density of stroke lesion region-of- interest (RO I) (e.g., fluid-uptake region) and Dnormatis the modal value (mode) of CT density for the normal tissue of the same tissue type, such as gray matter or white matter, over the entire brain. The modal value is chosen to represent the normal tissue CT density under the assumption that more than 50% of the brain tissue within each tissue type is normal. The size of the ROI can vary and may be as small as a single pixel, allowing for high resolution NWU mapping.

[0013] In some embodiments, at least one of the remapped denoised whole brain CT image, outputted visualization of the fluid-uptake region, or the outputted visualization of the quantification value of the fluid-uptake region, is employed by a clinician for a clinical assessment or detection of a stroke, stroke severity (e.g., whether the stroke is treatable), or for directing a treatment thereof.

[0014] In some embodiments, at least one of the remapped denoised whole brain CT image, outputted visualization of the fluid-uptake region, or the outputted visualization of the quantification value of the fluid-uptake region, is employed for a clinical assessment or detection of cerebral hemorrhage or cerebral edema (e.g., Traumatic Brain Injury (TBI), Ruptured Aneurysms, Brain Tumors, Hypertensive Encephalopathy, Cerebral Venous Sinus Thrombosis, Idiopathic Intracranial Hypertension, Cerebral Amyloid Angiopathy), or for directing a treatment thereof.

[0015] In some embodiments, the Al trained model includes convolutional neural network (e.g., Unet-based) configured with (i) skip and residual encoder-decoder and (ii) rotation and reflection equivariant group configurations.

[0016] In some embodiments, (i) the trained Al model or the model derived therefrom and (ii) the correcting for the denoised whole-brain CT image are executed on a CT scanner control system.

[0017] In some embodiments, (i) the trained Al model or the model derived therefrom and (ii) the correcting for the denoised whole-brain CT image, are executed on an externalcomputing device that receives the whole brain CT image or image set acquired from a CT scanner.

[0018] In an aspect, a system is disclosed comprising a processor; a memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to: receive a whole brain CT image; generate a denoised whole-brain CT image from the received whole-brain CT image using a trained Al model (e.g., neural network) configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15); correct (e.g., via a fully automated NWU brain mapping) the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image, wherein the correction employs one or more segmented regions of the denoised CT image consisting of healthy tissue (e.g., grey matter) in the denoised wholebrain CT image (e.g., wherein the healthy tissue is determined from a tissue probability maps having defined regions for gray matter, white matter, cerebrospinal fluid); and output the remapped denoised whole brain CT image in a report or graphical user interface, wherein the remapped denoised whole brain CT image preserves information for disease / injury if present in the whole brain CT image.

[0019] In some embodiments, the instructions, when executed by the processor, causes the processor to adjust the denoised whole-brain CT image prior to the correction via an image inhomogeneity correction operation (e.g., employing a skull-shape based bias field correction algorithm), wherein the image inhomogeneity correction operation applies a bias correction factor k derived from a distance map to a skull surface (e.g., inner surface, outer surface, or therebetween) (e.g., to correct for residual cupping artifacts associated with the distance from the skull).

[0020] In some embodiments, the instructions, when executed by the processor, causes the processor to determine a fluid-uptake region (swelling) in the remapped denoised whole brain CT image having uptake of water or fluids in the brain; and output visualization of the fluid-uptake region via the report or graphical user interface, wherein the visualization is overlaid over, or fused into, the remapped denoised whole brain CT image or a portion thereof.

[0021] In some embodiments, the instructions, when executed by the processor, causes the processor to determine a quantification value of the fluid-uptake region (swelling) in the remapped denoised whole brain CT image; and output the quantification value of the fluiduptake region via the report or graphical user interface.

[0022] In some embodiments, the quantification value NWU is determined by NWU = 1 — Dlesion / Dnormal, wherein Dlesionis a mean CT density of stroke lesion region-of-interest (RO I) (e.g., fluid-uptake region) and Dnormaiis the modal value (mode) of CT density for the normal tissue of the same tissue type, such as gray matter or white matter, over the entire brain. The modal value is chosen to represent the normal tissue CT density under the assumption that more than 50% of the brain tissue within each tissue type is normal. The size of the ROI can vary and may be as small as a single pixel, allowing for high-resolution NWU mapping.

[0023] In an aspect, a non-transitory compute readable medium is disclosed having instructions stored thereon, wherein execution of the instructions by a processor causes the processor to receive a whole brain CT image; generate a denoised whole-brain CT image from the received whole-brain CT image using a trained Al model (e.g., neural network) configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15); correct (e.g., via a fully automated NWU brain mapping) the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image, wherein the correction employs one or more segmented regions of the denoised CT image consisting of healthy tissue (e.g., grey matter) in the denoised wholebrain CT image (e.g., wherein the healthy tissue is determined from a tissue probability maps having defined regions for gray matter, white matter, cerebrospinal fluid); and output the remapped denoised whole brain CT image in a report or graphical user interface, wherein the remapped denoised whole brain CT image preserves information for disease / injury if present in the whole brain CT image.

[0024] In some embodiments, the instructions, when executed by the processor, cause the processor to adjust the denoised whole-brain CT image prior to the correction via an image inhomogeneity correction operation (e.g., employing a skull-shape-based bias field correction algorithm), wherein the image inhomogeneity correction operation applies a bias correction factor k derived from a distance map to a skull surface (e.g., inner surface, outer surface, or therebetween) (e.g., to correct for residual cupping artifacts attributable to skull thickness).

[0025] In some embodiments, the instructions, when executed by the processor, cause the processor to determine a fluid-uptake region (swelling) in the remapped denoised whole brain CT image having uptake of water or fluids in the brain; and output visualization of the fluiduptake region via the report or graphical user interface, wherein the visualization is overlaid over, or fused into, the remapped denoised whole brain CT image or a portion thereof.

[0026] In some embodiments, the instructions, when executed by the processor, cause the processor to determine a quantification value of the fluid-uptake region (swelling) in the remapped denoised whole brain CT image; and output the quantification value of the fluiduptake region via the report or graphical user interface.

[0027] In some embodiments, the quantification value NWU is determined byNWU = 1 — Desion / Dnormai, wherein Diesionis a mean CT density of stroke lesion region-of- interest (RO I) (e.g., fluid-uptake region) and Dnormaiis the modal value (mode) of CT density for the normal tissue of the same tissue type, such as gray matter or white matter, over the entire brain. The modal value is chosen to represent the normal tissue CT density under the assumption that more than 50% of the brain tissue within each tissue type is normal. The size of the ROI can vary and may be as small as a single pixel, allowing for high-resolution NWU mapping.Brief Description of the Drawings

[0028] Figs. 1A - 1C each shows an exemplary non-contrast computer tomography (CT) system employing a CT denoiser (i.e., Al denoiser), a dynamic range correction module for net water uptake assessment, and a visualizer. Fig. IB shows an image inhomogeneity correction operation module adjusting a denoised whole-brain CT image prior to the dynamic range correction. Fig. 1C shows an a fluid-uptake region quantifier determining a fluid-uptake region (swelling) in a remapped denoised whole-brain CT image having uptake of water or fluids in the brain.

[0029] Figs. 2A - 2B each shows an example operation flow of the exemplary system.

[0030] Fig. 3 shows a visualization of 3D convolution and a two-dimensional (2D) Carley graph for the dihedral group DA.

[0031] Figs. 4A - 4B each shows the architecture of a group convolutional neural network (e.g., GCNN-nnUNet, SRED-GCNN) and its end-to-end operation pipeline.

[0032] Fig. 5 shows a visualization of the encoder path for UNetR and SwinUNetR to show their difference.

[0033] Fig. 6A shows a visual comparison of the segmented results of the GCNN- nnUNet denoising model and baseline models (e.g., nnU-Net, SPM12, UNETR, SwinUNETR).

[0034] Fig. 6B shows a visual comparison between the SRED-GCNN denoising model and state-of-the-art denoising models.

[0035] Fig. 7 shows the statistical Wilcoxon test result of the spatial heterogeneity for the exemplary Al-denoising model coupled with the skull-shape-dependent bias-field correction.

[0036] Figs. 8A - 8B show the original NCCT image, the SPM12 bias -corrected image, the exemplary bias-corrected image, and the exemplary denoised and bias-corrected image in clinical cases.

[0037] Figs. 9A - 9B show the usage of the Z-score map and net water uptake (NWU) map for early stroke onset.

[0038] Fig. 10 shows a standalone proof-of-concept (PoC) deployment of the exemplary system.Detailed Description

[0039] Some references, which may include various patents, patent applications, and publications, are cited in a reference list and discussed in the disclosure provided herein. The citation and / or discussion of such references is provided merely to clarify the description of the disclosed technology and is not an admission that any such reference is “prior art” to any aspects of the disclosed technology described herein. In terms of notation, “[n]” corresponds to the nth reference in the list. For example, [1] refers to the first reference in the list. All references cited and discussed in this specification are incorporated herein by reference in their entirety and to the same extent as if each reference was individually incorporated by reference.

[0040] Figs. 1A - 1C each shows an exemplary non-contrast computer tomography (CT) system (shown as 100a, 100b, 100c) that employs a CT image denoiser (i.e., Al denoiser), a dynamic range correction module for net water uptake assessment, and a visualizer. Fig. IB shows the CT image denoiser (i.e., Al denoiser) and the dynamic range correction module operating with an image inhomogeneity correction operation module. Fig. 1C further shows a fluid-uptake region quantifier that can determine a fluid-uptake region (swelling) in a remapped denoised whole-brain CT image having uptake of water or fluids in the brain.

[0041] The image analysis and / or adjustments described herein may be incorproated into a CT scanner or an external system that operates with the CT scanner or incorporated into a radiologic software to be used by hospitals or healthcare practice groups to view patient CT images.

[0042] Example System #1

[0043] As shown in Fig. 1A, the example system 100a comprises a CT scanner 102, a CT image denoiser 106, a dynamic range correction 112 (e.g., for net water uptake), and a visualizer 118. The CT scanner 102 scans the user’s brain and generates a whole-brain CT image 104.

[0044] The CT image denoiser 106, coupled with the CT scanner 102, receives the whole-brain CT image 104 from the CT scanner 102. The CT image denoiser 106 then generates a denoised whole-brain CT image 110 from the whole-brain CT image 104 using a trained Al model 108 (e.g., neural network), wherein the trained Al model 108 is configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15). An example of an Al-denoiser 106 is provided later herein. Additional examples of Al-denoiser include published PCT patent application no. WO2022225580, which is incorporated by reference herein, among others.

[0045] The dynamic range correction module 112, coupled with the CT image denoiser 106, is configured to generate a remapped denoised whole-brain CT image 116 from the denoised whole-brain CT image 110 using pixel remapping module 114 that remaps pixels of the denoised whole-brain CT image 110 based on the segmented region of interest of healthy tissue (e.g., grey matter). The healthy tissue can be determined from tissue probability maps having defined regions for grey matter, white matter, and cerebrospinal fluid. The remapping preserves stroke or net water uptake information in the whole brain image that can be readily observed for changes in net water uptake that could used to identify earlier stages or outset of a stroke or like injuries or conditions. The exemplary system and method can additionally provide quantification of net water uptake, e.g., for stroke diagnosis, in addition to visualization.

[0046] For example, were the denoised whole-brain CT image 110 to have a few pixel values of resolution that have correspondence to the cerebral anomaly, the exemplary dynamic range correction module 112 can preserve that information by remapping the pixels based on the segmented portion of the brain tissue determined to be a baseline. By remapping or adjusting for the whole image, the correction has drowned out the pixel values of interest corresponding to the cerebral anomaly. In some embodiments, the healthy tissue is determined from the segmented region of interest using known toolsets or algorithms. An example that may be employed in the SPM12 toolkit. For segmentation tasks, SPM12 uses a unified model that can perform tissue segmentation, including those of healthy tissue.

[0047] The visualizer 118, coupled with the dynamic range correction module 112, outputs the remapped denoised whole-brain CT image 116 in a report or a graphical user interface, wherein the remapped denoised whole-brain CT image 116 preserves information for disease / injury if present in the whole-brain CT image 104.

[0048] An output 119 (i.e., remapped denoised whole-brain CT image 116) of the visualizer 118 shows a whole brain image. Notably, the whole brain image in this example shows a swelling region 119a (i.e., region of interest) of the patient’s brain to the dynamic range / remapping operation that would otherwise be different to view and quantify. The swelling region 119a can be recognizable from other regions of the brain in grey matter, as shown in the output 119.

[0049] The trained Al model 108 comprises a convolutional neural network (e.g., UNet- based), preferably configured with skip and residual encoder-decoder and rotation and reflection equivariant group configurations. The trained Al model 108 and the dynamic range correction module 112 can be executed on a CT scanner control system or on an external computing device that receives the whole brain CT image 104 or image set acquired from the CT scanner 102.

[0050] Example System #2

[0051] Fig. IB shows the CT image denoiser (i.e., Al denoiser) and the dynamic range correction module operating with an image inhomogeneity correction operation module. In the example shown in Fig. IB, the example system 100b includes the CT scanner 102, the CT image denoiser 106, and the dynamic range correction 112, as described in relation to Fig. 1A, and further includes an image inhomogeneity correction operation 124.

[0052] The CT scanner 102 scans the user’s brain and generates a whole-brain CT image 104. The CT image denoiser 106, coupled with the CT scanner 102, receives the whole-brain CT image 104 from the CT scanner 102. The CT image denoiser 106 then generates a denoised whole-brain CT image 110 from the whole-brain CT image 104 using a trained Al model 108 (e.g., neural network), wherein the trained Al model 108 is configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15).

[0053] The image inhomogeneity correction operation 124, coupled with the CT image denoiser 106, is configured to receive the denoised whole-brain CT image 110 from the CT image denoiser 106, where the denoised whole-brain CT image 110 contains a skull surface (e.g., inner surface, outer surface, or therebetween) of the user. The image inhomogeneity correction operation 124 then generates an adjusted denoised whole-brain CT image 128 by applying a bias factor k 122, derived from a distance map 120, to the skull surface using an image inhomogeneity correction algorithm 126.

[0054] The dynamic range correction module 112, coupled with the image inhomogeneity correction operation 126, then generates a remapped denoised whole-brain CT image 116 from the adjusted denoised whole-brain CT image 128 using module 114, wherein the module 114 remaps pixels of the adjusted denoised whole- brain CT image 128 based on segmented region of interest of healthy tissue (e.g., grey matter, e.g., determined via SPM12 or via other means described or referenced herein). The healthy tissue can be determined from tissue probability maps having defined regions for grey matter, white matter, and cerebrospinal fluid.

[0055] The visualizer 118, coupled with the dynamic range correction module 112, outputs the remapped denoised whole-brain CT image 116 in a report or a graphical user interface, wherein the remapped denoised whole-brain CT image 116 preserves information for disease / injury if present in the whole-brain CT image 104.

[0056] Example System #3

[0057] Fig. 1C further shows a fluid-uptake region quantifier that can determine a fluiduptake region (swelling) in a remapped denoised whole-brain CT image having uptake of water or fluids in the brain.

[0058] In the example shown in Fig. 1C, the example system 100c comprises a CT scanner 102, a CT image denoiser 106, an image inhomogeneity correction operation 124, a dynamic range correction 112 (e.g., for net water uptake), a fluid-uptake region quantifier 130, and a visualizer 118. As described in relation to Figs. 1A and IB, the CT scanner 102 scans the user’s brain and generates a whole-brain CT image 104. The CT image denoiser 106, coupled with the CT scanner 102, receives the whole-brain CT image 104 from the CT scanner 102. The CT image denoiser 106 then generates a denoised whole- brain CT image 110 from the wholebrain CT image 104 using a trained Al model 108 (e.g., neural network), wherein the trained Al model 108 is configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15).

[0059] The image inhomogeneity correction operation 124, coupled with the CT image denoiser 106, receives the denoised whole-brain CT image 110 from the CT image denoiser 106, wherein the denoised whole-brain CT image 110 contains a skull surface (e.g., inner surface, outer surface, or therebetween) of the user. The image inhomogeneity correction operation 124 then generates an adjusted denoised whole-brain CT image 128 by applying a bias factor k 122, derived from a distance map 120, to the skull surface using an image inhomogeneity correction algorithm 126.

[0060] The dynamic range correction module 112, coupled with the image inhomogeneity correction operation 126, generates a remapped denoised whole-brain CT image 116 from the adjusted denoised whole-brain CT image 128 using module 114, wherein the module 1 14 remaps pixels of the adjusted denoised whole-brain CT image 128 based on segmented region of interest of healthy tissue (e.g., grey matter, e.g., via SPM12 and etc.). The healthy tissue can be determined from tissue probability maps having defined regions for grey matter, white matter, and cerebrospinal fluid.

[0061] Fig. 1C further shows a fluid-uptake region quantifier 130. The quantifier 130 is coupled with the dynamic range correction module 112 and is configured to determine a fluiduptake region (swelling) and a quantification value of the fluid-uptake region of the brain in the remapped denoised whole-brain CT image, then transmits the remapped denoised whole-brain CT image and the quantification value of the fluid-uptake region to the visualizer 118.

[0062] The visualizer 118, coupled with the fluid-uptake region quantifier 130, outputs the combination 132 of the remapped denoised whole-brain CT image 116, the fluid-uptake region, and the quantification value of the fluid-uptake region in a report or a graphical user interface, wherein the remapped denoised whole-brain CT image 116 preserves information for disease / injury if present in the whole-brain CT image 104.

[0063] The quantification value of the fluid-uptake region may be determined by Equation 1.NWU 1 t) lesion / ^normal(Eq. 1)

[0064] In Equation 1, Diesionis a mean CT density of stroke lesion region-of-interest (ROI) (e.g., fluid-uptake region), Dnormaiis the modal value (mode) of CT density for the normal tissue of the same tissue type, such as gray matter or white matter, over the entire brain. The modal value may be chosen to represent the normal tissue CT density under the assumption that more than 50% of the brain tissue within each tissue type is normal. The size of the ROI may vary and may be as small as a single pixel, allowing for high-resolution NWU mapping.

[0065] The outputted visualization of the fluid-uptake region via the report or graphical user interface by the visualizer 118 can be overlaid over, or fused into, the remapped denoised whole brain CT image 116 or a portion thereof.

[0066] At least one of the remapped denoised whole brain CT image 116, outputted visualization of the fluid-uptake region, or the outputted visualization of the quantification value of the fluid-uptake region, is employed by a clinician for a clinical assessment or detection of a stroke, stroke severity (e.g., whether the stroke is treatable), or for directing a treatment thereof.

[0067] At least one of the remapped denoised whole brain CT image 116, outputted visualization of the fluid-uptake region, or the outputted visualization of the quantification value of the fluid-uptake region, is employed for a clinical assessment or detection of cerebral hemorrhage or cerebral edema (e.g., Traumatic Brain Injury (TBI), Ruptured Aneurysms, Brain Tumors, Hypertensive Encephalopathy, Cerebral Venous Sinus Thrombosis, Idiopathic Intracranial Hypertension, Cerebral Amyloid Angiopathy), or for directing a treatment thereof.

[0068] Example Method

[0069] Figs. 2A - 2B each shows an example operation flow of the exemplary system. Specifically, Fig. 2A shows an operation flow 200a of the exemplary system corresponding to the operation at system 100a described in relation to Fig. 1 A.

[0070] At step 202, the exemplary system receives, by a processor, a whole-brain CT image.

[0071] At step 204, the exemplary system generates, by the processor, a denoised wholebrain CT image from the received whole-brain CT image using a trained Al model (e.g., neural network) configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT) (e.g., having an image signal-to-noise of at least 12, e.g., 12-15).

[0072] At step 206, the exemplary system corrects, by the processor (e.g., via a fully automated NWU brain mapping), the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image.

[0073] At step 208, the exemplary system outputs, by the processor, the remapped denoised whole brain CT image in a report or graphical user interface.

[0074] Fig. 2B shows an operation flow 200b of an exemplary system, as described in relation to Fig. IB.

[0075] At step 210, the exemplary system receives an original non-contrast computer tomography (NCCT) image of a whole brain (i.e., a whole-brain CT image). At step 212, the exemplary system performs image harmonization on the whole-brain image received from step 210.

[0076] At step 212, the exemplary system performs image denoising on the whole-brain image using a skip and residual encoder-decoder design group convolutional neural network (SRED-GCNN) (e.g., an Al denoising model). The exemplary system generates a denoised whole-brain CT image 216 (i.e., denoised NCCT) as a result of step 212.

[0077] At step 218, the exemplary system performs a CT skull-shape dependent biasfield correction operation (i.e., image inhomogeneity correction operation), which comprises Al brain tissue segmentation operation 218a and skull-shape dependent bias-field algorithm 218b, using statistical parametric mapping version 12 (SMP12).

[0078] At operation 218a, the exemplary system segments the denoised whole-brain CT image 216 into one or more regions consisting of different types of tissues (e.g., gray / white / deep gray matter and / or cerebrospinal fluid). In other words, the exemplary system initializes modal density values of normal gray matter and white matter based on the denoised whole-brain CT image (e.g., denoised NCCT) masked by tissue probability maps.

[0079] At algorithm 218b, the exemplary system applies skull-shape bias correction to the denoised whole-brain CT image 216 and corrects the modal values of the denoised wholebrain CT image 216. In other words, the algorithm 218b optimizes the modal values by maximizing the normal tissue volume that is within ±5% NWU in each tissue type.

[0080] The exemplary system generates a corrected denoised whole-brain CT image as a result of step 218. At step 220, the exemplary shades the corrected denoised whole-brain CTimage and proceeds to step 222. At step 222, the exemplary system differentiates normal tissues from pathologic tissues and determines Hounsfield units (HU) of normal tissue types from the shading of the corrected denoised whole-brain CT image at step 220.

[0081] At step 224, the exemplary system employs a pixel-to-pixel net water uptake (NWU) brain mapping to determine a fluid-uptake region (swelling) and a quantification value of the fluid-uptake region in the corrected denoised whole-brain CT image. The NWU mapping uses the mean CT density of stroke lesion (denoted as D(esion) region of interest and normal tissue (denoted as Dnormal) region of interest in HU to determine the quantification value of the fluid-uptake region, as shown in Equation 1.

[0082] The generalized formulation for the the pixel-to-pixel NWU mapping, based on the voxel weighted contributions between the gray matter and white matter, can be expanded to include the weighted contribution from different tissue types as shown in Equation 2.(Eq. 2)

[0083] In Equation 2, element-wise / Hadamard product, HUGM, HUWMand HUCSFdenote the modal values of normal gray matter, white matter, and cerebrospinal fluid (CSF), respectively.

[0084] To help visualize the hypodense ischemic stroke in the denoised whole-brain CT image, at step 226, the exemplary system generates a standard Z-score map by subtracting the modal density values of normal gray matter and white matter from the respective segmented tissues and normalized by the standard error. The standard error in normal white matter is used for normalization in both gray and white matter, as the intensity distribution in white matter is typically more stable and symmetrical. Ischemic stroke is represented by a negative Z-score in this map, whereas hemorrhage / calcification has a positive Z-score. To remove the contribution from tissue with severe CSF contamination, both the Z-score and net water uptake maps are weighted by the (1 -probability of CSF).

[0085] Al-denoising with Group Equivariant Convolutional Neural Networks

[0086] Three-dimensional (3D) Group Equivariant Convolutional Neural Networks. Compared to data augmentation through various transformations, group convolution exploits the inherent symmetry in the data, enabling the learning of more robust invariant features and enhancing generalizability.

[0087] A standard 3D convolution is computed by performing a dot product of a 3D learnable kernel with a 3D feature map at each position to generate the output. In a 3D group convolution, the learnable kernel can undergo transformation through rotation and reflection togenerate multiple copies of the filter to encourage weight sharing. This process can be expressed as shown in Equation 1.(Eq. 1)

[0088] In Equation 1 , x and y represent the 3D coordinates of the output volume and the coordinates of the input volume '-channel / ' relative to the filter , respectively. The corresponding group operation can be shown in Equation 2, wherein g denotes generators for the group.(Eq. 2)

[0089] Example 3D Convolution Operation. Fig. 3 provides a visualization of 3D convolution and a two-dimensional (2D) Carley graph 302 for the dihedral group D4 to explain the concept of the group and extend it to the 3D scenario. The group D4 shown in Carley graph 302 corresponds to the symmetry of a square and represents a set of transformations that preserves the structure of the square, including rotation (r, r2, r3) and reflection ( / ). Combining or repeating those actions (e.g., rotations, reflections) (i.e., generators) can generate the entire group D4.

[0090] Extending to 3D rectangular cuboid (e.g., cuboid with square base) and cube cases, group 4 gets more complicated with the extra dimension. First, the rectangular cuboid cases (not shown), height = width length (h = w c), can be further divided into two subgroups, orientation preserving and non-orientation preserving

[0019] . In the orientationpreserving case, only rotation is being considered. The group D4 consists of 90° rotations around z axes and 180° rotations around y or -axes, which yield 8 different copies of the cuboid. With the addition of the reflection (non-orientation preserving), the total variations increase to 16.

[0091] In the case of cube 304, height = width = length (h = w = c), the rotations can be further extended to the diagonal axis (d1, d2, d d4) at 120° intervals. This group is referred to as group O, with 24 variations. The corresponding non-orientation preserving group Oh includes the reflection, resulting in 48 copies of the cube derived from rotation and reflection. All copies share the same weight, and with rotation and reflection operations, the convolution operation is equivariant to input rotation and reflection.

[0092] Group Convolutional Neural Network - No New UN el ( GCNN-nnUNet). Previous studies

[0016] ,

[0019] ,

[0016] ,

[0018] have demonstrated that using GCNN can achieve betterperformance in segmentation and classification tasks. However, the performance of group equivariant networks in state-of-the-art nnU-Net [5] segmentation pipelines remains less explored. To bridge this gap, GCNN-nnUNet - a rotation / reflection equivariant modem UNet architecture is developed. Using the principles of group equivariance and rotation symmetry can enhance the segmentation accuracy and robustness of the UNet neural network.

[0093] 1 ) Preprocess pipeline. To simplify usage, the GCNN-nnUNet is integrated into the No New UNet (nnU-Net) design, leveraging its preprocessing pipeline. Upon the target dataset being analyzed, nnU-Net 404 generates a set of configuration parameters 406 to configure the network’s depth, convolution type, etc. For GCNN-nnUNet, the convolution type can be the group convolution, and related settings like normalization and transpose convolution can be adjusted accordingly.

[0094] 2) Architecture design. Fig. 4A shows the architecture of GCNN-nnUNet and its end-to-end operation pipeline. As shown in Fig. 4 A, subpanel (a), dataset 402 is first analyzed by the preprocessing pipeline of nnU-Net 404. Based on the dataset property and the available resources on the training machine, nnU-Net 404 generates configurations 406, including batch size, patch size, network architecture, type of convolution operation, etc. The group equivariant convolution operation is integrated to be part of the configuration. Thus, if group equivariant convolution is selected by the user, the network architecture adjusts correspondingly, and the standard convolution at each layer is replaced with group equivariant convolution as shown in subpanel (a).

[0095] 3) Operation. The operation of the GCNN-nnUNet shown in Fig. 4A, subpanel(a) includes two operations: encoding 409 and decoding 411. Once the dataset 402 is preprocessed and the configurations 406 are generated, the dataset goes through the encoding operation 409. During the encoding operation 409, dataset 402 undergoes group convolution operations (e.g, 408a - 408f), including rotations and reflections, and downsampling operations (e.g., 410a - 4 lOd).

[0096] After the encoding operation 409, dataset 402 may go through the decoding operation 411. During decoding operation 411 , the dataset 402 may undergo group convolution operations (e.g., 408g - 408i), including, e.g., rotations and reflections, and upsampling operations (e.g., 412a - 412d). At the end of the decoding operation, dataset 402 undergoes a convolution operation 414 and becomes a segmented dataset 416 as an equivariant convolution result of the GCNN-nnUNet.

[0097] 4) Equivariant convolution result. Fig. 4A, Subpanel (b) shows the equivariant property of the 3D group convolution under rotation in an orientation-preserving group (C4H).The input is a 3x3x3 cube (for illustration purposes, it can be any arbitrary size of rectangle, cuboid, or cube). As shown in Fig. 4A, subpanel (b), the top row shows the 180-degree rotation of the cube over opposing faces, where the top layer becomes the bottom layer, and the bottom layers become the top layer, the middle layer remains but with the value flipped. Subsequently, standard 3D convolution is performed on those two same (but rotated) volumes with the same kernel, resulting in different outputs, as shown in the second row. Despite the input rotation, group equivariant convolution ensures the output from the convolution operation to be invariant as shown in the rest rows.

[0098] 5) Loss function. For the loss function design, both dice coefficient loss and crossentropy loss may be combined together to create a compound loss. This design leverages the advantages of both losses, focusing on detailed pixel-level classification while also addressing the class imbalance inherent in medical imaging segmentation tasks. The loss function may be expressed as shown in Equation 3.L(y, y) —a’ CE (y, y) + P • DC(y, yj,(Eq. 3) where c CE y, y = - ^ l(y = c)log (yc) c— 1

[0099] In Equation 3, y is the ground truth, y is the predicted segmentation mask, and a and f are tunable parameters to balance the importance of each criterion set to 1 to give equal importance.

[0100] Group Convolutional Neural Network - Skip and Residual Encoder-Decoder (SRED-GCNN). To increase the ischemia detection sensitivity of NCCT in a tissue-specific manner, the SRED-GCNN is developed for denoising even in the presence of intensity change due to stroke ischemia. The local intensity change from different NCCTs does not create a bias in denoising if the loss between individual CT and the averaged CT is minimized in the same patient even with beam hardening effects typically seen in CT.

[0101] 1 ) Model training. Acute ischemic stroke patients who have at least three 2.5mmNCCT from May 2016 to Dec 2000 are included in the model training with variations in scanner manufacturers or model, imaging protocols (120 kV, variable exposure time, and tube current), head position, and imaging plane. A total of 463 NCCT cases from 325 patients who hadreceived multiple head CT scans within a 5-day window were included. Two patients had 7 CT scans, one patient had 6 CT scans, four patients had 5 CT scans, 14 patients had 4 CT scans, 32 patients had 3 CT scans, and the rest of them had 1 CT scan after excluding scans due to head movement. Most of these repeated scans are due to hemorrhagic transform in the ischemic stroke. CT cases were grouped by patient, with 83 cases for model training and 380 cases for holdout testing.

[0102] 2) Loss function. An averaged CT may be created by spatially co-registering all the CTs (n) for each patient. Image noise is roughly reduced by a factor of Vn in the normal brain tissue. A special loss function was designed to reduce the difference between individual CT and the averaged CT using the rotation-reflection equivariant group UNet (GCNN), featuring both skip and residual connections from encoder to decoder (SRED). The network was implemented on Keras with Tensorflow core vl.13 backend [34’]. A custom loss function, defined per Equation 4, was used to train the network to reduce the difference between individual CT and the averaged CT.

[0103] 3) Operation. Fig. 4B shows the architecture of SRED-GCNN and its end-to-end operation pipeline. As shown in subpanel (a), the operation of the SRED-GCNN includes two operations: encoding 409 and decoding 411. Once a dataset is preprocessed and the configuration parameters are generated, the dataset goes through the encoding operation 409. During the encoding operation 409, the dataset undergoes 2D D4 group convolution operations (e.g., 420a - 420e) (denoted as D4 GConv2D), including rotations and reflections, and 2D downsampling operations (e.g., 422a - 422d) (denoted as Maxpool2D).

[0104] After the encoding operation 409, the dataset goes through the decoding operation 411. During the decoding operation 411, the dataset undergoes group convolution operations (e.g., 420f - 420i) (denoted as D4 GConv2D), including rotations and reflections, and 2D upsampling operations (e.g., 424a - 424c) (denoted as Upsample2D). At the end of the decoding operation, the dataset undergoes a 2D convolution operation 426 (denoted as D4 to Z2 GConv2D) and becomes a segmented dataset as an equivariant convolution result of the SRED- GCNN.

[0105] Between encoding operation 409 and decoding operation 41 1, the data set also undergoes concatenation operations (e.g., 428a - 428e) and add operations (e.g., 430a - 430d).

[0106] Fig. 4B, Subpanel (b) shows the formulation of the loss function defined per Equation 4.(Eq. 4)

[0107] In Equation 4 and subpanel (b), y is the input noisy CT slice from patient z, 0 represents the model configuration parameters (e.g., batch size, patch size, network architecture, type of convolution operation), Ld z(y) is the model loss from the input y slice from patient z, fe (y) isthe denoised CT slice, E(ylz) is the corresponding mean CT slice from patient z, M(y, fe(y)) is the modal value of y -fe, S(y, / 0(y)) is the skew of y -fe, and k and fo are positive definite weights.

[0108] Two-dimensional (2D) Group Equivariant Convolutional Neural Networks.Group Equivariant Convolutional Networks (GCNN)

[0013] employs group equivariant filters to preserve the spatial relationships between features. A standard convolution of 2D input image / with filter yz has K channels is defined as shown in Equation 5, wherein the convolution is equivariant to an operation like translation, but not rotation or reflection.(Eq. 5)

[0109] On the other hand, filters are designed to be equivariant to transformations like rotation and reflection in GCNN. In the first layer, the feature maps are thought of as functions on group G that contain symmetry elements like rotation. The group lifting convolution process maps Z2— > G and can be defined per Equation 6, wherein g denotes generators for the group.(Eq. 6)

[0110] For the rest of the layers, both filter and feature maps are functions on G through G —> G convolution. The output from the group convolution is a stack of feature maps that are equivariant to the input transformation within the group.

[0111] Building on GCNN, more advanced equivariant networks like steerable CNN

[0014] and spherical CNN

[0015] are proposed. GCNN offers parameter efficiency by sharing weights across different spatial locations or transformations, which is particularly beneficial in biomedical imaging, where data acquisition and labeling demand expertise and, consequently, training data is sparse. In

[0016] , GCNN DenseNet is developed for classification tasks, in which they achieve equal or even better performance compared to the baselines with only half of the training data.

[0017] integrated GCNN into 2D UNet and performed nuclear segmentation while

[0018] demonstrates that GCNN with 2D UNet works well in segmenting stroke lesions at different vascular territories within a large patient pool. In

[0019] , 3D-GCNN is used for pulmonary noduleclassification. Moreover,

[0020] utilized the special Euclidean group SE(3) group to perform diffusion-weighted imaging (DWI) segmentation and showed promising results.

[0112] No New UNet. No New Unet is a self-configurable framework for medical image segmentation. Despite the promising results delivered by the deep learning-based segmentation methods, tuning or reconfiguring DNN for specific tasks still requires significant expertise. [5] developed nnU-Net, which integrates preprocessing, neural network architecture configuration, and post-processing into a unified framework, streamlining the segmentation process.

[0113] By formulating explicit dependencies between dataset properties and design choices into heuristic rules, nnU-Net systematically configures the neural network architecture without requiring task-specific expertise. It also integrates more recent activation and normalization options, such as leaky ReLU and instance normalization

[0021] , into the configuration pipeline. This automated system constructs customized network architecture for individual segmentation tasks and preprocesses data to appropriate sizes that comply with the physical constraints of the training machine. This dual optimization, fine-tuning both the network and leveraging machine capabilities together, contributes to nnUNet’s superior performance in a wide range of segmentation tasks without the need for manual tuning, setting the state-of-the-art performance on several datasets for diverse segmentation tasks.

[0114] Transformer-based segmentation architecture. Transformer-based segmentation architecture includes two frameworks, UNetR

[0010] and SwinUNetR

[0011] ,

[0023] . Both frameworks adopt the contract and expanding path of UNet, with different designs at the encoder path. Fig. 5 provides a visualization of the encoder path for UNetR and SwinUNetR to show their difference.

[0115] As shown in Fig. 5, subpanel (a), UNetR

[0010] replaces the CNN feature extraction layers in the UNet encoder with transformers that consist of multi-head self-attention and multilayer perceptron. At each embedding level, a sequence of representations is extracted and reshaped to ensure features from each resolution level are available and can be concatenated with a decoder path for improved location and boundary prediction.

[0116] In SwinUNetR

[0011] shown in Fig. 5, subpanel (b), the transformer in the encoder path is substituted with a swin transformer

[0024] , employing a hierarchical design by merging patches at each layer as the network deepens. Additionally, it uses a novel shift window partition design to enhance the inter-layer connection at each level and shows improvement in MRI datasets.

[0117] Statistical parametric mapping (SPM). SPM12 is a standard software package commonly used for functional and structural neuroimage analysis. For segmentation tasks, SPM 12 uses a unified model

[0025] to perform registration, tissue segmentation, and biascorrection simultaneously. SPM12 models image intensity using Gaussian mixtures, with each Gaussian characterized by its mean, variance, and mixing proportion.

[0118] A tissue probability map (TPM) serves as a prior to the model, and a posterior probability is generated through the Bayes rule for each tissue class. The default segmentation routine in SPM12 offers six tissue classes: gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), bone, soft tissue, and background. A customized TPM with additional tissue classes deep gray matter and cerebellum gray matter, was produced for the exemplary system and method. For the ground truth label generation, the labels for each CT image were derived from the corresponding MR images using the methodology outlined in [9].

[0119] Examplelmage Inhomogeneity Correction (Skull-shape dependent bias-field correction)

[0120] The image inhomogeneity correction operation employs a skull-shape dependent NCCT bias-field correction algorithm that corrects the intensity inhomogeneity by using a distance map to inner skull surface (distinner). The bias correction factor (k) is determined by minimizing the standard deviation of whole brain normal gray matter after bias correction as shown in Equation 7. log(distinner))(Eq. 7)

[0121] Shading correction. Image inhomogeneity from residual cupping artifacts is common in non-contrast head CT. The artifact is obvious when cortical gray matter that is close to the skull has a much higher Hounsfield unit than cortical gray matter that are deep inside the skull. Current state-of-the-art image inhomogeneity correction (N3 / N4) algorithms in SPM12, FSL, ANT etc., are not able to correct for the inhomogeneity caused by the residual cupping artifact when large vessel occlusion ischemic stroke is present.

[0122] Experimental Results and Additional Examples

[0123] A study was conducted to develop a non-contrast computer tomography (CT) system and method with Al-denoiser and dynamic range correction for net water uptake assessment.

[0124] Datasets. The study employed two datasets, including a publicly available dataset derived from the SynthRad2023 challenge

[0026] and a private dataset obtained from a hospital (referred to as Center D) for extensive testing. The datasets comprised MRI-CT pairs of brain scans acquired from four distinct facilities, each employing different equipment, technologies, and acquisition protocols.

[0125] Table 1 shows information on the acquisition protocols (e.g., MRI, CT) on the open and private datasets.Table 1

[0126] The original SynthRad2023 dataset

[0026] was developed for the purpose of synthesizing CT images for radiotherapy from MRI scans. As shown in Table 1, the dataset consisted of two separate sub-datasets: one containing MRI-CT pairs and the other containing CBCT-CT pairs. The study exclusively utilized the MRI-CT pairs of brain scans. The dataset comprised a total of 180 MRI-CT volume pairs from three Centers (A, B, and C). All volumes were stored in nifti (.nii.gz) format and resampled with a uniform grid of Ixlxl mm3to give the same voxel spacing. The patients’ age ranges from 3 to 93 years old, with 43.9% being male. The study split the 180 cases into 150 training and 30 testing cases. During the training phase, the study further split the 150 cases into training and validation subsets in an 8:2 ratio.

[0127] The study also used a separate dataset obtained from Center D during evaluation to enrich the comprehensiveness of the dataset and to further evaluate the robustness of the model. The dataset included initial non-contrast head CT images and follow-up MRI imagesfrom patients with acute ischemic stroke who were within 4.5 hours of stroke onset and subsequently treated with intravenous thrombolysis. The study preprocessed these CT images using Skipped Residual Encoder-Decoder Group Convolutional Neural Network (SRED-GCNN)

[0027] to reduce image noise and generate better CT tissue segmentation maps using SPM12. The study registered the follow-up axial T1 FLAIR MRI to the CT using SPM12 with rigid affine transform. The study then resampled the T1 FLAIR at the same voxel coordinates as the CT and resulted in spatially matching 388 MRI-CT volume pairs. All volumes were stored in nifti (.nii.gz) format and resampled with a uniform grid of Ixlxl mm3to give the same voxel spacing. The study randomly selected 100 volumes for training, and the study used the rest 288 volumes for testing.

[0128] SPM12 created the ground truth tissue labels by the standard atlas-based segmentation of the follow-up T1 FLAIR MRI, except for deep gray matter, which was segmented from non-contrast CT as MRI did not have good tissue contrast in deep gray matter. The study created a customized brain atlas that included white matter, gray matter, deep gray matter, and cerebellum gray matter tissue types, as deep gray and cerebellum gray matter were presented with higher radiodensity than cortical gray matter in the non-contrast CT. The deep gray and cerebellum gray matter labels were created using brain regions defined by the Neuromorphometrics brain atlas. The deep gray matter tissue class included Caudate, Putamen, and Pallidum regions, whereas the cerebellum gray matter class included the cerebellum gray matter as well as the cerebellar vermal lobules LX.

[0129] Performance metric. The study employed the dice coefficient as the performance metric to compare the segmentation quality. The computation of dice coefficients can be expressed as shown in Equation 8.2* |y n y||y| + lyl(Eq. 8)

[0130] In Equation 8, y is the predicted segmentation mask, and y is the ground truth. The study only computed the foreground dice coefficient score during the evaluation.

[0131] The study highlighted the advantage of GCNN-nnUNet through a visual analysis of the segmented results, a comparison of dice coefficient between GCNN-nnUNet and multiple baselines on open and private datasets, the performance evaluation of GCNN-nnUNet against multiple base-lines with reduced parameters, and an assessment of GCNN-nnUNet’ s performance with limited training data. All volumes were registered before the training and inference processes, eliminating the need for any additional spatial transformation.

[0132] Visualization of GCNN-nnUNet 's segmentation results. Fig. 6A shows a visual comparison of the segmented results of the GCNN-nnUNet denoising model and baseline models (e.g., nnU-Net, SPM12, UNETR, SwinUNETR). Multiple slices from the same volume were shown in Fig. 6A, with the top row including the original CT scans (shown in subpanel a), which were the input to the neural network, as well as its corresponding MRI scans (shown in subpanel b), which were shown as a reference to better visualize the performance. The performance improvement of GCNN-nnUNet over SPM12, UNETR, and SwinUNetR was significant, especially in the whiter matter region. GCNN-nnUNet’ s results were more continuous and aligned with the anatomical structure. Compared to the closest baseline nnUNet, GCNN-nnUNet achieved better performance at details as shown in the second row (subpanels c and d). Such details may be crucial when used for tracking disease progression, surgical planning, etc.

[0133] Visual comparisons between SRED-GCNN and state-of-the-art CT models. The study evaluated the SRED-GCNN-based CT denoising model against block matching and 3D filtering (BM3D) [35 ’-37’] denoising algorithm as well as residual encoder-decoder convolutional neural network (RED-CNN) [22’] deep learning CT denoising algorithm in the 380 hold-out testing cases using signal-to-noise and contrast-to-noise ratios in the cortical gray matter (GM), corpus collosum white matter and thalamus deep gray matter respectively. The study evaluated denoising performance using root mean square error (RMSE) and structural similarity (SSIM) index between the denoised image and the mean CT while excluding the stroke regions to confirm statistically significant improvements.

[0134] Fig. 6B shows a visual comparison between the SRED-GCNN denoising model and state-of-the-art denoising models. As shown in Fig. 6B, the CT image denoised by the SRED-GCNN was smoother and less noisy compared to the CT images denoised by state-of-the- art denoising models. This is because the signal-to-noise ratio (SNR) of the denoised CT using SRED-GCNN ranged around 12 to 15 and reached the same level commonly seen in MRI, which improved visualization of ischemic stroke lesions in NCCT that were non-diagnostic and missed. The SNR significantly increased by 2.5x to 2.8x and the contrast-to-noise ratio (CNR) by at least 2x in cerebral cortex / gray matter (GM), corpus callosum / white matter (WM), and thalamus / deep gray matter (DG) (one-tailed paired / -test, with probability p < 0.0001). RED-CNN denoise to similar SNR range but it contained excessive blurring and was not useful in stroke diagnosis.

[0135] Dice coefficient comparison. Table 2 shows the dice coefficient of GCNN- nnUNet and the baseline models on both the open dataset as well as the private dataset.Table 2

[0136] As shown in Table 2, compared to the deep learning-based framework, the traditional statistical method performs the worst for CT image segmentation in most tissue classes except in deep gray matter, where SPM12 performed better than SwinUNetR. All deep learning-based methods had comparable performance, especially in deep gray matter and cerebellum gray matter; all frameworks achieved a dice coefficient close to 0.9. GCNN-nnUNet showed steady improvement in all tissue classes compared to baseline models, achieving the highest dice coefficient on the SynthRad dataset on all tissue classes. As for the private dataset, GCNN-nnUNet outperformed the best baseline model nnU-Net in WM, CSF, tied in GM, and CERB, and performed competitively in DG.

[0137] The smaller improvement in the private dataset was due to the CT images being acquired in 2D and then interpolated to 3D, which reduced the advantages of utilizing the 3D group equivariant kernel. Although the transformer-based UNETR and SwinUNETR showed remarkable performance in organ segmentation, their performance in brain tissue segmentation on non-contrast CT still lagged behind the CNN-based segmentation network.

[0138] Performance comparison with reduced parameters. To demonstrate the efficient weight sharing of GCNN-nnUNet, the study reduced the input filter (feature size) number for GCNN-nnUNet and baseline models to reduce the model parameters and evaluate itsperformance. For every model except SwinUNetR, the study reduced its input filter number by half from the default 32 input filter (32, 16, 8). Since SwinUNetR required the input feature size to be a multiple of 12, the study reduced its feature size from default 48 to 24 and 12, respectively.

[0139] Table 3 shows the performances of GGCN-nnUNet and baseline deep-learning methods (e.g., nnUNet, UNetR, SwinUNetR, SPM12) on a full dataset. As shown in Table 3, the segmentation performance degraded as the number of model parameters reduced for all models. The dice coefficient of nnUNet, UNetR, and SwinUNetR were reduced between 2-3%. However, with GCNN-nnUNet, the performance degradation was less than 0.4%.Table 3

[0140] Performance comparison with limited data. Obtaining high-quality training data in the biomedical imaging field was a challenging task due to the expertise and training required for data labeling. Therefore, a robust model that can learn the invariant feature from a few training examples can be beneficial in this field. To this end, the study used half and a quarter of the original dataset to train the model and evaluate the performance, respectively. Table 4 shows the performances of GCNN-nnUNet and baseline deep-learning methods (e.g., nnUNet, UNetR, SwinUNetR, SPM12) on a limited (i.e., reduced) dataset. As shown in Table 4, as the number of training cases (i.e., data) reduced, the performance of all models degraded. However, GCNN-nnUNet achieved the best segmentation performance among all baselines with limited training samples.Table 4

[0141] Performance comparison without spatial augmentation. To demonstrate that the observed equivariance stems from the network design rather than extensive data augmentation, the study disabled spatial data augmentation for both GCNN-nnUNet and nnU-Net during training. The test results, shown in Table 5, revealed that GCNN-nnUNet continued to outperform nnU-Net by a significant margin, highlighting the effectiveness of the group equivariant property.Table 5

[0142] Evaluation of Al-denoising and skull-shape dependent bias -field correction. The study tested the skull-shape dependent bias-field correction method in a patient cohort who had thrombectomy (number of patients: n=391) either with or without tissue plasminogen activator (tPA) treatment between 2017 and 2021. Fig. 7 shows the statistical Wilcoxon test result of the spatial heterogeneity for the exemplary Al-denoising model coupled with the skull-shape-dependent bias-field correction. As shown in Fig. 7, the Al image denoising employed in the exemplary system inherently improved spatial heterogeneity in gray matter and white matter. Combining it with the skull-shape dependent bias-field correction, the spatial heterogeneity further improved significantly (Wilcoxon test, with probability p < 0.0001).

[0143] Figs. 8A - 8B show the original NCCT image, the SPM12 bias -corrected image, the exemplary bias-corrected image, and the exemplary denoised and bias-corrected image in clinical cases. Specifically, as shown in Fig. 8A, compared to the original NCCT, the skull-shape dependent bias-field correction operation generates images with higher precision in difficult clinical cases, e.g., patients with incomplete skull (top), and hemorrhagic strokes with metal artifacts (bottom).

[0144] Fig. 8B shows a clinical case where a large vessel occlusion ischemic stroke NCCT case (stroke region is slightly darker, white arrow) with cupping artifact was bias- corrected using the exemplary method with / without SRED-GCNN denoising and compared with SPM12 bias-correction. The image spatial inhomogeneity was measured by the corresponding density standard deviation in the normal gray matter (gray matter probability pGM> 0.95), in which the threshold can effectively exclude intracranial hemorrhage / old stroke. The density standard deviations in the normal gray matter were 4.6 HU, 4.6 HU, 3.7 HU, 2.5 HU, whereas the standard deviations in the normal white matter (pWM> 0.95) were 4.5 HU, 4.5 HU, 3.5 HU, 2.1 HU in the original, SPM12 bias-corrected, exemplary bias-corrected, and exemplary bias- corrected with denoising, respectively.

[0145] After SPM12 bias-correction, the stroke region contrast was reduced compared to the contralateral normal side, along with improved homogeneity near the frontal region. The exemplary bias correction preserves the image contrast in the stroke region while improving the homogeneity in the gray matter throughout the brain. After bias correction and Al denoising, the ischemic stroke region was clearly visible.

[0146] Net water uptake mapping and Z- score mapping for early stroke onset. Figs. 9A - 9B show the usage of the Z-score map and net water uptake (NWU) map for early stroke onset.

[0147] In the clinical case shown in Fig. 9A, a patient presented in the emergency department (ED) 1.25 hours after stroke onset. The Z-score map showed a clear hypodense region in the posterior left insula, which had a mean NWU of 13.9%, indicating an acute ischemic stroke, which was not obvious in the admission NCCT. The location and extent of the stroke were confirmed by restricted diffusion in the posterior left insula in the follow-up diffusion- weighted imaging (DW1). The ischemic stroke reversed its course after IV alteplase was given to the patient.

[0148] In the clinical case shown in Fig. 9B, a patient was presented in ED with neurological symptoms 4 hours after stroke onset. The net water uptake map showed the increased NWU at the ischemic stroke location, with mean NWU at 16%. The location was confirmed by the hyperintense region with restricted diffusion in the follow-up DWI at 21 hours after stroke onset. Although the patient was treated with IV alteplase, the ischemic infarct did not reverse, as indicated by the hyperintense lesion in the T2 FLAIR MRI.

[0149] CT image denoising on an edge computing device. Fig. 10 shows a standalone proof-of-concept (PoC) deployment of the exemplary system using a GPU-embedded edge computing device (NVIDIA Jetson AGX Xavier, 8-core ARM CPU, 512 CUDA cores Volta GPU, 2TB SSD upgrade) using Django backend (v3.2.6) and Python 3. The device received head NCCT via the Digital Imaging and Communications in Medicine (DICOM) connectivity standard, and the denoised NCCT was sent back via the same protocol. A local DICOM listener for ARM64 AArch64 architecture was compiled and installed on the Jetson using the DICOM toolkit (DCMTK) library to receive the original NCCT and send denoised NCCT back to the workstation (e.g., CT scanner). The images were sorted by image series upon arrival. The system can poll the receiving file system within a preset timeout to confirm the edge computing device received the entire imaging study. The study series that required denoising was automatically identified based on Study Description DICOM tag (0008,1030) matching, and the images from that series are converted to an image volume in NIFTI format.

[0150] The NCCT was automatically deskulled, followed by ALdenoising using the SRED-GCNN model. The Al denoising model (8 M parameters) was converted to a lightweight TensorRT model engine with 16-bit precision. After denoising, the skull and face were restored, and the image result was sent back to the original DICOM sender as a new image series. The image data may be moved to a backup location which may be deleted automatically in a predefined hold-out time. Image preprocessing and deskull took less than 10 seconds, and denoising takes ~200ms / slice at full 512x512 resolution. A typical 3mm head scan took less than 30 seconds to complete when tested on a local network. Currently, the additional software pipeline necessary for NWU map and standard Z-score map generations was brain image segmentation from SPM12, which ran on Matlab software (Mathworks Inc) but not on edge computing devices.

[0151] Discussion

[0152] Stroke eliminates about 140,000 Americans every year. As the population ages, the incidence of this acute, severe disease will increase, making the delivery of effective interventions paramount to reducing the burden of disease and the cost to the healthcare system.A missed early diagnosis means a missed treatment opportunity. About two-thirds of acute ischemic stroke patients have mild to moderate symptoms without a large vessel occlusion but with a clear opportunity for benefit from thrombolysis. While diffusion-weighted (DWI) MRI is the most sensitive method to diagnose acute ischemic stroke, MRI is not routinely part of the initial emergency department (ED) evaluation because of time constraints, costs, the need for patient cooperation, and the presence of implanted devices that are not MRI-compatible. Furthermore, many rural and free-standing facilities do not have MRI capability and even large centers do not all have around-the-clock staffing.

[0153] Stroke diagnosis is challenging because clinical symptoms can be subtle, subjective, and fluctuating. The current state of diagnosis begins with the clinical impression of triage nurses and physicians and a non-contrast CT (NCCT) to rule out brain bleed or alternative pathology. Current NCCT technology is not sensitive enough to positively identify early ischemia [1 ’ -6’]. In patients with suspected large vessel stroke, a CT angiography is performed to identify the presence and location of large vessel occlusion. Importantly, though, an LVO should be suspected, and in a registry of LVO patients, 21% had low NIHSS [7’]. Stroke centers struggle to identify subtle stroke presentations and consequently err on the side of treating patients when the clinical diagnosis of stroke can possibly be made and the time of stroke onset is within 4.5 hours, even if the diagnosis is not clear. While this approach is reasonable in the current state, improved diagnostics may reduce the risk of adverse effects of unnecessary treatments and reduce unnecessary costs and prolonged hospitalizations.

[0154] In addition to the challenges of assessing symptoms, time of onset can be a challenge to define and, even when known, may not reflect tissue viability as “fast progressors” may present early with large irreversible lesions, and “slow progressors” may still have a sizeable salvageable tissue later. Advanced imaging, such as DWI MRI and / or perfusion imaging, can assist in selecting candidates for treatment (DAWN, DEFUSE 3) [8’], [9’]. Furthermore, 1 in 5 patients present with stroke symptoms but an unknown onset time [10’]. The current recommendation for this group of patients is to give IV thrombolytic if the diffusion MRI shows an ischemic stroke lesion and the lesion is negative in fluid-attenuated inversion recovery (FLAIR) MRI as this “DWI / FAIR mismatch” suggests stroke onset within 4.5 hours (WAKE-UP trial) [11 ’]. Patients would usually miss the treatment opportunity as most EDs cannot perform an MRI efficiently.

[0155] Additional discussion. Brain tissue segmentation, the process of dividing brain images into distinct anatomical structures such as gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), is crucial in medical imaging and neuroscience research. It aids inearly diagnosis of neurological disorders, planning and guiding surgical procedures and mapping brain function for research and clinical purposes.

[0156] Computed tomography (CT) and magnetic resonance imaging (MRI) are two common structural neuroimaging modalities. CT uses X-rays to create cross-sectional images, providing excellent contrast for dense structures like bones. It is ideal for detecting fractures and other bone-related abnormalities. MRI, on the other hand, detects signals emitted by hydrogen nuclei as they return to their equilibrium state after being disturbed by a radio frequency pulse, excelling in imaging soft tissues with superior details. It is particularly useful for identifying abnormalities in the brain and spinal cord.

[0157] Several practical considerations must be taken into account when choosing imaging modality, with availability being a major factor. In the United States, there are 27.01 CT units [1] in hospitals per million inhabitants but only 17.93 MRI units [2]. Additionally, the speed of imaging is also a critical consideration. CT scans are much faster than MRI scans, often taking just a few minutes compared to the longer duration required for MRI. This speed advantage makes CT particularly valuable in emergency room settings, where rapid diagnosis is essential for conditions such as stroke and trauma. Therefore, the ability to perform accurate brain tissue segmentation on CT images has significant value, especially in time- sensitive emergencies. By accurately distinguishing between different brain tissues on CT images, medical professionals can make more informed decisions quickly.

[0158] However, brain tissue segmentation on CT images is challenging due to low soft tissue contrast [3] compared to MRI. Moreover, image noise due to statistical uncertainty in physical measurements can further blur tissue boundaries, making accurate labeling difficult even for experienced physicians.

[0159] Deep Neural Network (DNN) has shown promising results in classification and segmentation tasks. The UNet [4] architecture, in particular, demonstrates robustness and accuracy in medical image segmentation compared to traditional statistical methods. Many taskspecific models [5-7] have been built on top of UNet, but focus on MRI images. These models trained on MRI images cannot be generalized to CT images, making training on CT images with proper labeling crucial. Studies on CT-based segmentation using statistical methods [8] for brain tissues often yield unsatisfactory results due to low contrast between grey and white matter. [9] first developed to use deep learning-based methods for brain tissue segmentation on CT images by acquiring CT and T1 -weighted MR volume pairs from the same day and using MRI-derived labels as the ground truth to train their plain U-Net model. The results showed significant improvement.

[0160] The instant study developed the exemplary system and method employing an automatic segmentation pipeline that used group equivariant convolutional neural network (GCNN) in a modem 3D UNet architecture or other Al model to perform brain tissue segmentation on non-contrast CT scans. Compared to traditional data augmentation methods, which altered the CT noise distribution from Poisson to Gaussian, GCNN was able to preserve the original noise distribution. The study integrated GCNN into the nnU-Net [5] pipeline to leverage nnU-Net’ s self-configuration capabilities to streamline the setup for GCNN-nnUNet. The study compared GCNN-nnUNet with state-of-the-art deep learning methods, including nnU- Net [5], UNetR

[0010] , and SwinUNetR[l 1], as well as the traditional brain atlas-based statistical method SPM12

[0012] in CT brain tissue segmentation. The results demonstrate that GCNN- nnUNet improved overall segmentation performance.

[0161] The instant study also further developed an Al-enabled high resolution fully automated Net Water Uptake (NWU) brain mapping technique that relies only on NCCT and does not require a CTP imaging procedure. The NWU can replace CTP / MRP and be used as both a tissue clock and a positive indication of early ischemia to identify patients who will benefit from treatment with thrombolysis or EVT. This approach can more efficiently and accurately make a stroke diagnosis, saving cost and valuable time during the acute stroke treatment process.

[0162] Net Water Uptake, a measurement derived from NCCT that is similar to MRI DWI, can quantitatively measure the degree of ionic edema in the ischemic brain tissue. When using an NWU cutoff at 11.5%, Brooks and colleagues have found that the NWU‘s predictive power to identify patients within 4.5 hours from stroke onset is comparable and potentially superior to MRI-based measures [12’]. This could replace the unknown stroke onset time with an NCCT tissue clock, which closely mirrors tissue viability. Recent multicenter clinical observation studies showed that quantitative NWU measurements from NCCT can not only detect, but monitor the progression of ionic edema in ischemic stroke from early stroke onset up to 21 hours [13’]. NWU increases almost linearly with increasing stroke onset time in a tight error band up to 4.5 to 6 hours after stroke onset. Studies of NWU in successful recanalization of acute middle cerebral artery stroke patients [14’] showed that an early increase of NWU (“fast progressors”) during admission is an independent prognostic indicator of poor stroke outcome at 90 days when adjusted for other confounding covariates. Another study also independently confirms the predictive value of the NWU in stroke outcomes of thrombectomy patients [15’]. To perform NWU measurements, the ischemic stroke lesion needs to be positively identified either in NCCT or in CT perfusion (CTP). Due to the poor contrast-to-noise ratio of strokelesions in NCCT, small and moderate ischemic strokes are often missed during the first 6 hours of stroke onset [16’- 18’ ]. Therefore, most NWU studies [10’]. [14’], [15’], [19’ - 33’] require a CTP scan to better identify the location and extent of the stroke. A technologist can manually trace the stroke lesion in the perfusion map at multiple slice levels as a region-of-interest (RO I) and copy the mirrored version of the ROI to the contralateral side as a normal tissue comparison. This process is labor intensive as ROIs cannot include brain sulcus, cerebrospinal fluid, or intraparenchymal hemorrhage, and these components must be manually excluded.

[0163] A major advance in this technology goes to the core of the underdiagnosis challenge in stroke. Even in ambiguous cases with low NIHSS or vague non-focal symptoms, such as a patient with a fall or delirium, this technology has an advantage over MRI and CTP. For MRI or CTP to be considered, a patient would need to be triaged by a stroke algorithm. In the case of NWU NCCT, a patient who receives an NCCT for any reason could be evaluated and alerted for possible stroke and treatment in a timely manner. This overcomes a significant cause of underdiagnosis. The exemplary system maps NWU to the tissue with high spatial accuracy, not relying on average measurements of large ROIs. Denoising allows the identification of reversible ischemic lesions and differentiation of acute stroke more than 5 mm in diameter from old ischemic lesions. The distribution of ischemic tissue may suggest large vessel occlusion, allowing rapid evaluation with CTA and transfer to centers where EVT can be performed. With a low-cost edge computing solution that is compatible with different CT scanners, it can potentially be deployed to rural and standalone clinics, thus democratizing timely stroke treatment for underserved communities as well as vulnerable patients with underrecognized symptoms.

[0164] Conclusion

[0165] The construction and arrangement of the systems and methods as shown in the various implementations are illustrative only. Although only a few implementations have been described in detail in this disclosure, many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes, and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.). For example, the position of elements may be reversed or otherwise varied, and the nature or number of discrete elements or positions may be altered or varied. Accordingly, all such modifications are intended to be included within the scope of the present disclosure. The order or sequence of any process or method steps may be varied or re-sequenced according to alternative implementations. Other substitutions, modifications, changes, and omissions may be made in the design, operatingconditions, and arrangement of the implementations without departing from the scope of the present disclosure.

[0166] The present disclosure contemplates methods, systems, and program products on any machine-readable media for accomplishing various operations. The implementation of the present disclosure may be implemented using existing computer processors, or by a special purpose computer processor for an appropriate system, incorporated for this or another purpose, or by a hardwired system. Implementations within the scope of the present disclosure include program products including machine-readable media for carrying or having machine-executable instructions or data structures stored thereon. Such machine-readable media can be any available media that can be accessed by a general purpose or special purpose computer or other machine with a processor. By way of example, such machine-readable media can comprise RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code in the form of machine-executable instructions or data structures, and which can be accessed by a general purpose or special purpose computer or other machine with a processor.

[0167] When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or a combination of hardwired or wireless) to a machine, the machine properly views the connection as a machine-readable medium. Thus, any such connection is properly termed a machine-readable medium. Combinations of the above are also included within the scope of machine-readable media. Machine-executable instructions include, for example, instructions and data which cause a general-purpose computer, special purpose computer, or special purpose processing machines to perform a certain function or group of functions.

[0168] Although the figures show a specific order of method steps, the order of the steps may differ from what is depicted. Also, two or more steps may be performed concurrently or with partial concurrence. Such variation will depend on the software and hardware systems chosen and on designer choice. All such variations are within the scope of the disclosure. Likewise, software implementations could be accomplished with standard programming techniques with rule-based logic and other logic to accomplish the various connection steps, processing steps, comparison steps and decision steps.

[0169] Each and every feature described herein, and each and every combination of two or more of such features, is included within the scope of the present invention, provided that the features included in such a combination are not mutually inconsistent.

[0170] Although example embodiments of the disclosed technology are explained in detail herein, it is to be understood that other embodiments are contemplated. Accordingly, it is not intended that the disclosed technology be limited in its scope to the details of construction and arrangement of components set forth in the following description or illustrated in the drawings. The disclosed technology is capable of other embodiments and of being practiced or carried out in various ways.

[0171] It must also be noted that, as used in the specification and the appended claims, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” or “approximately” one particular value and / or to “about” or “approximately” another particular value. When such a range is expressed, other exemplary embodiments include from the one particular value and / or to the other particular value.

[0172] By “comprising” or “containing” or “including” is meant that at least the named compound, element, particle, or method step is present in the composition or article or method, but does not exclude the presence of other compounds, materials, particles, method steps, even if the other such compounds, material, particles, method steps have the same function as what is named.

[0173] Unless otherwise expressly stated, it is in no way intended that any method set forth herein be construed as requiring that its steps be performed in a specific order.Accordingly, where a method claim does not actually recite an order to be followed by its steps or it is not otherwise specifically stated in the claims or descriptions that the steps are to be limited to a specific order, it is no way intended that an order be inferred, in any respect. This holds for any possible non-express basis for interpretation, including: matters of logic with respect to arrangement of steps or operational flow; plain meaning derived from grammatical organization or punctuation; the number or type of embodiments described in the specification.

[0174] While the methods and systems have been described in connection with certain embodiments and specific examples, it is not intended that the scope be limited to the particular embodiments set forth, as the embodiments herein are intended in all respects to be illustrative rather than restrictive.

[0175] The following patents, applications, and publications, as listed below and throughout this document, are hereby incorporated by reference in their entirety herein. Reference List #1[1] OECD. “Computed tomography (CT) scanners (indicator)”. In: (2023), p. 234.[2] OECD. “Magnetic resonance imaging (MRI) units (in-dicator)”. In: (2023), p. 234.[3] Edwin J.R. van Beek and Eric A. Hoffman. “Functional Imaging: CT and MRI”. In: Clinics in Chest Medicine 29.1 (2008). Contemporary Chest Imaging, pp. 195-216. ISSN: 0272-5231.[4] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks forBiomedical Image Segmentation. 2015. arXiv: 1505.04597 [cs.CV].[5] Fabian Isensee, Paul F Jaeger, Simon A A Kohl, Jens Petersen, and Klaus H Maier-Hein.“nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation”. In: Nature methods 18.2 (Feb. 2021), pp. 203-211. ISSN: 1548-7091.[6] Sijing Cai, Yunxian Tian, Harvey Lui, Haishan Zeng, Yi Wu, and Guannan Chen. “Dense-UNet: a novel multi-photon in vivo cellular image segmentation model based on a convolutional neural network”. In: Quantitative imaging in medicine and surgery 10.6 (2020), p. 1275.[7] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, andManning Wang. “Swin-unet: Unet-like pure transformer for medical image segmentation”. In: European conference on computer vision. Springer. 2022, pp. 205- 218.[8] C. Chu, D.A. Handwerker, P.A. Bandettini, and J. Ash-bumer. “Measuring the Consistency of Global Func-tional Connectivity Using Kernel Regression Methods”. In: Pattern Recognition in NeuroImaging (PRNI), 2011 International Workshop on. 2011, pp. 41- 44.[9] Meera Srikrishna, Joana B Pereira, Rolf A Heckemann, Giovanni Volpe, Danielle vanWesten, Anna Zettergren, Silke Kem, Lars-Olof Wahlund, Eric Westman, Ingmar Skoog, and Michael Sch' oll. “Deep learning from MRLderived labels enables automatic brain tissue classifica-tion on human brain CT”, en. In: Neuroimage 244 (Sept. 2021), p. 118606.

[0010] Ali Hatamizadeh, Dong Yang, Holger R. Roth, and Daguang Xu. “UNETR: Transformers for 3D Medical Image Segmentation”. In: 2022 IEEE / CVF Winter Con-ference on Applications of Computer Vision (WACV)(2021), pp. 1748-1758.

[0011] Ah Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger Roth, and DaguangXu. “Swin UN-ETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images”. In: arXiv preprint arXiv:2201.01266 (2022).

[0012] Karl J Friston, Andrew P Holmes, Keith J Worsley, J-P Poline, Chris D Frith, and RichardSJ Frackowiak. “Statistical parametric maps in functional imaging: a general linear approach”. In: Human brain mapping 2.4 (1994), pp. 189-210.

[0013] Taco Cohen and Max Welling. “Group Equivariant Convolutional Networks”. In:Proceedings of The 33rd International Conference on Machine Learning. Vol. 48. Proceedings of Machine Learning Research. PMLR, 20-22 Jun 2016, pp. 2990-2999.

[0014] Taco S. Cohen and Max Welling. “Steerable CNNs”. In: International Conference onLearning Representations. 2017.

[0015] Taco S. Cohen, Mario Geiger, Jonas K 'ohler, and Max Welling. “Spherical CNNs”. In:International Confer-ence on Learning Representations. 2018.

[0016] Bastiaan S. Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling.“Rotation Equivariant CNNs for Digital Pathology”. In: Medical Image Com-puting and Computer Assisted Intervention - MICCAI 2018. Springer International Publishing, 2018, pp. 210-218.

[0017] Benjamin Chidester, That-Vinh Ton, Minh-Triet Tran, Jian Ma, and Minh N. Do.“Enhanced Rotation-Equivariant U-Net for Nuclear Segmentation”. In: Pro-ceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. June 2019.

[0018] Kelvin K Wong, Jonathon S Cummock, Guihua Li, Rahul Ghosh, Pingyi Xu, John J Volpi, and Stephen TC Wong. “Automatic segmentation in acute ischemic stroke: Prognostic significance of topological stroke volumes on stroke outcome”. In: Stroke 53.9 (2022), pp. 2896-2905.

[0019] Marysia Winkels and Taco S. Cohen. “Pulmonary nod-ule detection in CT scans with equivariant CNNs”. In: Medical Image Analysis 55 (2019), pp. 15-26. ISSN: 1361 - 8415.

[0020] Renfei Liu, Francois Lauze, Erik Bekkers, Kenny Er-leben, and Sune Darkner. “GroupConvolutional Neural Networks for DWI Segmentation”. In: Proceedings of the First International Workshop on Geometric Deep Learning in Medical Image Analysis. Vol. 194. Proceed-ings of Machine Learning Research. PMLR, 18 Nov 2022, pp. 96-106.

[0021] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lem-pitsky. Instance Normalization: TheMissing Ingredi-ent for Fast Stylization. 2017. arXiv: 1607 . 08022 [cs.CV].

[0022] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, XiaohuaZhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Geliy, Jakob Uszkoreit, and Neil Houlsby. “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale”. In: International Conference on Learning Representations. 2021.

[0023] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu,Vishwesh Nath, and Ali Hatamizadeh. “Self-supervised pre-training of swin transformers for 3d medical image analysis”. In: Pro-ceedings of the IEEE / CVF Conference on Computer Vi-sion and Pattern Recognition. 2022, pp. 20730-20740.

[0024] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, andBaining Guo. “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows”. In: 2021 IEEE / CVF Interna-tional Conference on Computer Vision (ICCV). 2021, pp. 9992-10002.

[0025] N. Weiskopf, A. Lutti, G. Helms, M. Novak, J. Ash-burner, and C. Hutton. “Unified segmentation based correction of R1 brain maps for RF transmit field inhomogeneities (UNICORT)”. In: NeuroImage 54.3 (2011), pp. 2116-2124.

[0026] Adrian Thummerer, Erik van der Bijl, Arthur Galapon, Joost J.C. Verhoeff, Johannes A.Langendijk, Stefan Both, Cornells (Nico) A.T. van den Berg, and Matteo Maspero. “SynthRAD2023 Grand Challenge dataset: Generating synthetic CT for radiotherapy”. In: Medical Physics 50.7 (July 2023), pp. 4664-4674. ISSN: 0094-2405.

[0027] Kelvin K Wong, Jonathon S Cummock, Yunjie He, Rahul Ghosh, John J Volpi, andStephen TC Wong. “Retrospective study of deep learning to reduce noise in noncontrast head CT images”. In: Computerized Medical Imaging and Graphics 94 (2021), p. 101996.Reference List #2[1'] Shake A, Oostema A, Jones J. Bet 2: Missed diagnosis of ischaemic stroke in the emergency department. Emerg Med J. 2018;35:768-769[2’] Tamutzer AA, Lee SH, Robinson KA, Wang Z, Edlow JA, Newman-Toker DE. Ed misdiagnosis of cerebrovascular events in the era of modern neuroimaging: A metaanalysis. Neurology. 2017;88: 1468-1477[3'] Newman-Toker DE, Moy E, Valente E, Coffey R, Hines AL. Missed diagnosis of stroke in the emergency department: A cross-sectional analysis of a large population-based sample. Diagnosis (Berl). 2014;1 :155-166[4'] Arch AE, Weisman DC, Coca S, Nystrom KV, Wira CR, Schindler JL. Missed ischemic stroke diagnosis in the emergency department by emergency medicine and neurology services. Stroke. 2016;47:668-673[5’] Madsen TE, Khoury J, Cadena R, Adeoye O, Alwell KA, Moomaw CJ, et al. Potentially missed diagnosis of ischemic stroke in the emergency department in the greater cincinnati / northem kentucky stroke study. Acad Emerg Med. 2016;23:1128-1135[6’] Liberman AL, Prabhakaran S. Stroke chameleons and stroke mimics in the emergency department. Curr Neurol Neurosci Rep. 2017; 17: 15[7'] Mazya MV, Cooray C, Lees KR, Toni D, Ford GA, Bar M, et al. Minor stroke due to large artery occlusion. When is intravenous thrombolysis not enough? Results from the sits international stroke thrombolysis register. Eur Stroke J. 2018;3:29-38[8’] Nogueira RG, Jadhav AP, Haussen DC, Bonafe A, Budzik RF, Bhuva P, et al. Thrombectomy 6 to 24 hours after stroke with a mismatch between deficit and infarct. New England Journal of Medicine. 2017;378:11-21[9'] Albers GW, Marks MP, Kemp S, Christensen S, Tsai JP, Ortega-Gutierrez S, et al. Thrombectomy for stroke at 6 to 16 hours with selection by perfusion imaging. New England J ournal of Medicine. 2018;378:708-718[10'] Rubin MN, Barrett KM. What to do with wake-up stroke. The Neurohospitalist. 2015;5:161-172[I T] Thomalla G, Simonsen CZ, Boutitie F, Andersen G, Berthezene Y, Cheng B, et al. Mri- guided thrombolysis for stroke with unknown time of onset. N Engl J Med. 2018;379:611-622[12'] Broocks G, Leischner H, Hanning U, Flottmann F, Faizy TD, Schon G, et al. Lesion age imaging in acute stroke: Water uptake in ct versus dwi-flair mismatch. Ann Neurol. 2020;88:1144-1152[13'] Minnerup J, Broocks G, Kalkoffen J, Langner S, Knauth M, Psychogios MN, et al.Computed tomography-based quantification of lesion water uptake identifies patients within 4.5 hours of stroke onset: A multicenter observational study. Ann Neurol. 2016;80:924-934[14'] Nawabi J, Flottmann F, Kemmling A, Kniep H, Leischner H, Sporns P, et al. Elevated early lesion water uptake in acute stroke predicts poor outcome despite successful recanalization - when "tissue clock" and "time clock" are desynchronized. Int J Stroke. 2021;16:863-872[15'] Broocks G, Kemmling A, TeBarek S, McDonough R, Meyer L, Faizy TD, et al. Quantitative lesion water uptake as stroke imaging biomarker: A tool for treatment selection in the extended time window? Stroke. 2022;53:201-209[16'] Patel SC, Levine SR, Tilley BC, Grotta JC, Lu M, Frankel M, et al. Lack of clinical significance of early ischemic changes on computed tomography in acute stroke. JAMA. 2001;286:2830-2838[17'] Gao J, Parsons MW, Kawano H, Levi CR, Evans TJ, Lin L, et al. Visibility of ct early ischemic change is significantly associated with time from stroke onset to baseline scan beyond the first 3 hours of stroke onset. J Stroke. 2017;19:340-346[18'] Hopyan J, Ciarallo A, Dowlatshahi D, Howard P, John V, Yeung R, et al. Certainty of stroke diagnosis: Incremental benefit with ct perfusion over noncontrast ct and ct angiography. Radiology. 2010;255: 142-153[19'] Cohen TS, Welling M. Group equivariant convolutional networks. arXiv e-prints. 2016[20'] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI). 2015;9351:234-241[21 '] Wong KK, Cummock JS, He Y, Ghosh R, Volpi JJ, Wong STC. Retrospective study of deep learning to reduce noise in non-contrast head ct images. Comput Med Imaging Graph. 2021 ;94: 101996[22'] Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao P, et al. Low-dose ct with a residual encoder-decoder convolutional neural network. IEEE Trans Med Imaging. 2017;36:2524-2535[23'] Ghosh R, Cummock JS, Wong K, Volpi JJ, Wong ST. Retrospective study of deep learning denoising to improve contrast-to-noise ratio in missed hyperacute ischemic stroke lesions. European Stroke Organization Conference 2021: ESOC 202 E 2021[24'] Rorden C, Bonilha L, Fridriksson J, Bender B, Kamath HO. Age-specific ct and mri templates for spatial normalization. Neuroimage. 2012;61 :957-965[25'] Broocks G, Flottmann F, Scheibel A, Aigner A, Faizy TD, Hanning U, et al. Quantitative lesion water uptake in acute stroke computed tomography is a predictor of malignant infarction. Stroke. 2018;49: 1906-1912[26'] Xu T, Y ang J, Han Q, Wu Y, Gao X, Xu Y, et al. Net water uptake, a neuroimaging marker of early brain edema, as a predictor of symptomatic intracranial hemorrhage after acute ischemic stroke. Front Neurol. 2022; 13:903263[27'] Nawabi J, Kniep H, Schon G, Flottmann F, Leischner H, Kabiri R, et al. Hemorrhage after endovascular recanalization in acute stroke: Lesion extent, collaterals and degree of ischemic water uptake mediate tissue vulnerability. Front Neurol. 2019;10:569[28'] Wintermark M, Flanders AE, Velthuis B, Meuli R, van Leeuwen M, Goldsher D, et al.Perfusion-ct assessment of infarct core and penumbra: Receiver operating characteristic curve analysis in 130 patients suspected of acute hemispheric stroke. Stroke. 2006;37:979-985[29'] Srikrishna M, Pereira JB, Heckemann RA, Volpe G, van Westen D, Zettergren A, et al. Deep learning from mri-derived labels enables automatic brain tissue classification on human brain ct. Neuroimage. 2021 ;244:118606[30'] Boned S, Padroni M, Rubiera M, Tomasello A, Coscojuela P, Romero N, et al. Admission ct perfusion may overestimate initial infarct core: The ghost infarct core concept.Journal of Neurolnterventional Surgery. 2017;9:66-69[31'] Rava RA, Snyder KV, Mokin M, Waqas M, Zhang X, Podgorsak AR, et al. Assessment of computed tomography perfusion software in predicting spatial location and volume of infarct in acute ischemic stroke patients: A comparison of sphere, vitrea, and rapid.Journal of Neurolnterventional Surgery. 2021 ;13:130-135[32'] Garcia-Tomei A, Campos D, Rubiera M, Boned S, Olive-Gadea M, Requena M, et al. Ischemic core overestimation on computed tomography perfusion. Stroke. 2021 ;52:1751-1760[33'] Sivan-Hoffmann R, Saban M, Buxbaum C, Srour F, Sprecher E, Eran A, et al. Infarct core reliability by ct perfusion is a time-dependent phenomenon. Journal of Neuroimaging. 2020;30:240-245[34'] Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, et al. Tensorflow: Large- scale machine learning on heterogeneous systems. 2015[35'] Dabov K, Foi A, Katkovnik V, Egiazarian K. Bm3d image denoising with shape-adaptive principal component analysis. SPARS'09 - Signal Processing with Adaptive Sparse Structured Representations. 2009[36'] Danielyan A, Katkovnik V, Egiazarian K. Bm3d frames and variational image deblurring. IEEE Transactions on Image Processing. 2011;21: 1715-1728[37'] Dabov K, Foi A, Katkovnik V, Egiazarian K. Image denoising by sparse 3-d transformdomain collaborative filtering. IEEE Transactions on image processing. 2007;16:2080- 2095[38'] Wong KK, Cummock JS, Li G, Ghosh R, Xu P, Volpi JJ, et al. Automatic segmentation in acute ischemic stroke: Prognostic significance of topological stroke volumes on stroke outcome. Stroke. 2022;53:2896-2905[39'] Purushotham A, Campbell BC, Straka M, Mlynash M, Olivot JM, Bammer R, et al. Apparent diffusion coefficient threshold for delineation of ischemic core. Int J Stroke. 2015;10:348-353[40'] Wong KK, Cummock J, Volpi J, Wong ST. Ai segmentation of acute ischemic infarct from diffusion mri robust to image reflection and rotation (oral presentation). JointEuropean Stroke Organisation and the World Stroke Organization Conference (ESO- WSO). 2020[41'] Veeling BS, Linmans J, Winkens J, Cohen T, Welling M. Rotation equivariant cnns for digital pathology. arXiv e-prints. 2018[42'] Linmans J, Winkens J, Veeling BS, Cohen TS, Welling M. Sample efficient semantic segmentation using rotation equivariant convolutional networks. arXiv e-prints. 2018[43'] Goyal M, Ospel JM, Menon B, Almekhlafi M, Jayaraman M, Fiehler J, et al. Challenging the ischemic core concept in acute ischemic stroke imaging. Stroke. 2020;51 :3147-3155[44'] Broocks G, Faizy TD, Flottmann F, Schon G, Langner S, Fiehler J, et al. Subacute infarct volume with edema correction in computed tomography is equivalent to final infarct volume after ischemic stroke: Improving the comparability of infarct imaging endpoints in clinical trials. Invest Radiol. 2018;53:472-476[45'] Inoue M, Mlynash M, Christensen S, Wheeler HM, Straka M, Tipirneni A, et al. Early diffusion- weighted imaging reversal after endovascular reperfusion is typically transient in patients imaged 3 to 6 hours after onset. Stroke. 2014;45: 1024-1028[46'] Neuberger U, Mohlenbruch MA, Herweh C, Ulfert C, Bendszus M, Pfaff J. Classification of bleeding events: Comparison of ecass iii (european cooperative acute stroke study) and the new heidelberg bleeding classification. Stroke. 2017;48:1983-1985.

Claims

What is claimed:

1. A method comprising: receiving, by a processor, a whole brain CT image; generating, by the processor, a denoised whole-brain CT image from the received wholebrain CT image using a trained Al model configured to reduce image noise and improve tissue contrast in non-contrast CT (NCCT); correcting, by the processor, the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image, wherein the correction employs one or more segmented regions of the denoised CT image consisting of healthy tissue in the denoised whole-brain CT image; and outputting, by the processor, the remapped denoised whole brain CT image in a report or graphical user interface, wherein the remapped denoised whole brain CT image preserves information for disease / injury if present in the whole brain CT image.

2. The method of claim 1 further comprising: adjusting, by the processor, the denoised whole-brain CT image prior to the correction via an image inhomogeneity correction operation, wherein the image inhomogeneity correction operation applies a bias correction factor k derived from a distance map to a skull surface.

3. The method of claim 1 or 2 further comprising: determining, by the processor, a fluid-uptake region in the remapped denoised whole brain CT image having uptake of water or fluids in the brain; and outputting, by the processor, visualization of the fluid-uptake region via the report or graphical user interface, wherein the visualization is overlaid over, or fused into, the remapped denoised whole brain CT image or a portion thereof.

4. The method of claim 3 further comprising: determining, by the processor, a quantification value of the fluid-uptake region in the remapped denoised whole brain CT image; and outputting, by the processor, the quantification value of the fluid-uptake region via the report or graphical user interface.

5. The method of claim 4, wherein the quantification value NWU is determined by:NWU — 1 — Dlesion / Dnormalwherein Diesi onis a mean CT density of stroke lesion region-of-interest (ROI) andDnormai is the modal value (mode) of CT density for the normal tissue of the same tissue type .

6. The method of any one of claims 1-5, wherein at least one of the remapped denoised whole brain CT image, outputted visualization of the fluid-uptake region, or the outputted visualization of the quantification value of the fluid-uptake region, is employed by a clinician for a clinical assessment or detection of a stroke, stroke severity, or for directing a treatment thereof.

7. The method of any one of claims 1-6, wherein at least one of the remapped denoised whole brain CT image, outputted visualization of the fluid-uptake region, or the outputted visualization of the quantification value of the fluid-uptake region, is employed for a clinical assessment or detection of cerebral hemorrhage or cerebral edema (e.g., Traumatic Brain Injury (TBI), Ruptured Aneurysms, Brain Tumors, Hypertensive Encephalopathy, Cerebral Venous Sinus Thrombosis, Idiopathic Intracranial Hypertension, Cerebral Amyloid Angiopathy), or for directing a treatment thereof.

8. The method of any one of claims 1-7, wherein the Al trained model comprises convolutional neural network configured with (i) skip and residual encoder-decoder and (ii) rotation and reflection equivariant group configurations.

9. The method of any one of claims 1-8, wherein (i) the trained Al model or the model derived therefrom and (ii) the correcting for the denoised whole-brain CT image, are executed on a CT scanner control system.

10. The method of any one of claims 1-8, wherein (i) the trained Al model or the model derived therefrom and (ii) the correcting for the denoised whole-brain CT image, are executed on an external computing device that receives the whole brain CT image or image set acquired from a CT scanner.

11. A system comprising: a processor; a memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to: receive a whole brain CT image;generate a denoised whole-brain CT image from the received whole-brain CT image using a trained Al model configured to reduce image noise and improve tissue contrast in noncontrast CT (NCCT); correct the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image, wherein the correction employs one or more segmented regions of the denoised CT image consisting of healthy tissue in the denoised whole-brain CT image; and output the remapped denoised whole brain CT image in a report or graphical user interface, wherein the remapped denoised whole brain CT image preserves information for disease / injury if present in the whole brain CT image.

12. The system of claim 11, wherein the instructions, when executed by the processor, causes the processor to: adjust the denoised whole-brain CT image prior to the correction via an image inhomogeneity correction operation, wherein the image inhomogeneity correction operation applies a bias correction factor k derived from a distance map to a skull surface.

13. The system of claim 11 or 12, wherein the instructions, when executed by the processor, causes the processor to: determine a fluid-uptake region in the remapped denoised whole brain CT image having uptake of water or fluids in the brain; and output visualization of the fluid-uptake region via the report or graphical user interface, wherein the visualization is overlaid over, or fused into, the remapped denoised whole brain CT image or a portion thereof.

14. The system of claim 13, wherein the instructions, when executed by the processor, causes the processor to: determine a quantification value of the fluid-uptake region in the remapped denoised whole brain CT image; and output the quantification value of the fluid-uptake region via the report or graphical user interface.

15. The system of claim 14, wherein the quantification value NWU is determined by:NWU — 1 — Dlesion / Dnormalwherein Diesionis a mean CT density of stroke lesion region-of-interest (ROI) (e.g., fluid-uptake region) and Dnormaiis the modal value (mode) of CT density for the normal tissue of the same tissue type.

16. A non-transitory compute readable medium having instructions stored thereon, wherein execution of the instructions by a processor causes the processor to: receive a whole brain CT image; generate a denoised whole-brain CT image from the received whole-brain CT image using a trained Al model configured to reduce image noise and improve tissue contrast in noncontrast CT (NCCT); correct the denoised whole-brain CT image that remaps intensity values of pixels of the received whole-brain CT image, wherein the correction employs one or more segmented regions of the denoised CT image consisting of healthy tissue in the denoised whole-brain CT image; and output the remapped denoised whole brain CT image in a report or graphical user interface, wherein the remapped denoised whole brain CT image preserves information for disease / injury if present in the whole brain CT image.

17. The non-transitory compute readable medium of claim 16, wherein the instructions, when executed by the processor, causes the processor to: adjust the denoised whole-brain CT image prior to the correction via an image inhomogeneity correction operation, wherein the image inhomogeneity correction operation applies a bias correction factor k derived from a distance map to a skull surface.

18. The non-transitory compute readable medium of claim 16 or 17, wherein the instructions, when executed by the processor, causes the processor to: determine a fluid-uptake region in the remapped denoised whole brain CT image having uptake of water or fluids in the brain; and output visualization of the fluid-uptake region via the report or graphical user interface, wherein the visualization is overlaid over, or fused into, the remapped denoised whole brain CT image or a portion thereof.

19. The non-transitory compute readable medium of claim 18, wherein the instructions, when executed by the processor, causes the processor to:determine a quantification value of the fluid-uptake region (swelling) in the remapped denoised whole brain CT image; and output the quantification value of the fluid-uptake region via the report or graphical user interface.

20. The non-transitory compute readable medium of claim 19, wherein the quantification value NWU is determined by:NWU 1 ^lesion / ^normal wherein Diesionis a mean CT density of stroke lesion region-of-interest (RO I) and Dnormaiis the modal value of CT density for the normal tissue of the same tissue type, such as gray matter or white matter.

Images