Systems and methods for detecting abnormalities in non-contrast computed tomography (CT) scans

Abstract

Example embodiments described herein provide for a system for detecting features in non-contrast computed tomography scans. The system can include one or more processors configured to obtain data associated with at least one computed tomography (CT) scan for at least one patient, provide the data associated with the at least one CT scan to a model ensemble to output an indication that a condition is present, and generate a graphical user interface (GUI) including an indication of the condition. In examples, the model ensemble includes a first model to receive the data associated with the at least one CT scan as an input and generate a feature map, and a second model to receive the feature map as an input and generate the indication that the condition is present based on the feature map.

Description

06721 / 076333-1036 PATENTSYSTEMS AND METHODS FOR DETECTING ABNORMALITIES IN NON-CONTRAST COMPUTED TOMOGRAPHY (CT) SCANSCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 681,565 filed on August 9, 2024, titled “SYSTEMS AND METHODS FOR DETECTING ABNORMALITIES IN NON-CONTRAST COMPUTED TOMOGRAPHY (CT) SCANS,” the contents of which are hereby incorporated by reference in their entirety.TECHNICAL FIELD

[0002] This application generally relates to techniques for detecting features in noncontrast computed tomography (CT) scans and, in some embodiments, to techniques for training systems to detect features in non-contrast CT scans using datasets generated from contrast-based CT scans.BACKGROUND

[0003] Medical imaging techniques are regularly used to diagnose patients with critical health conditions. For example, for pulmonary embolisms (PEs) which pose a life-threatening cardiovascular emergency, computed tomography pulmonary angiography (CTPA) scans are generated as standard clinical practice when diagnosing PEs. But CTPA scans raise certain notable concerns. For example, the use of iodinated contrast to improve the visibility of PEs can cause patients to experience allergic reactions and have adverse effects on the kidneys (particularly in patients with impaired kidney function).

[0004] Alternatives to CTPA scans do not provide the same level of sensitivity. As a result, the likelihood that a clinician arrives at a false-negative diagnosis can increase, particularly in cases where the markers that would be present in a CTPA image are hidden or blended with other expected anatomical features. In the case of PEs, this can include PEs (particularly small or peripheral PEs) blending in with surrounding tissue and being imperceptible to clinicians interpreting these alternative scans.Page 1 of 434896-3752-9172.106721 / 076333-1036 PATENTSUMMARY

[0005] Embodiments described herein include systems and methods for improving on shortcomings and the art and can provide any number of additional or alternative benefits as well. For example, systems and methods are described herein for detecting features in non-contrast CT scans. When implemented, the systems and methods described herein can output data that more accurately indicates whether a PE is present in a patient in non-contrast CT scans. These systems and method can, in turn, reduce the chances of a false-negative diagnosis and allow for faster treatment of patients experiencing PEs. Further, the systems and methods described allow for the imaging of a patient without the need for contrast dyes that can lead to adverse health outcomes for the patient. While the present disclosure is discussed with respect to the identification of PEs in non-contrast CT scans, it will be understood that the techniques described herein are not limited to PEs and can be applied to identify one or more different conditions such as the presence or absence of internal bleeding, tumors, cancers, infections, blood vessel abnormalities, internal injuries, certain bone conditions such as fractures in smaller bones or complex structures (e.g., the spine), and / or the like.

[0006] In an embodiment, a system for detecting features in non-contrast computed tomography scans can include one or more processors configured to: obtain data associated with at least one computed tomography (CT) scan for at least one patient; provide the data associated with the at least one CT scan to a model ensemble to cause the model ensemble to output an indication that a condition is present, the model ensemble comprising: a first model that is configured to receive the data associated with the at least one CT scan as an input and generate a feature map based on the at least one CT scan, and a second model that is configured to receive data associated with the feature map as an input and generate the indication that the condition is present based on the feature map; and generate data associated with a graphical user interface (GUI), the GUI comprising an indication of the condition represented by the CT scan. At least a portion of the model ensemble can be trained using non-contrast CT scans that are generated based on corresponding CTPA scans.

[0007] The one or more processors configured to obtain the data associated with the at least one CT scan can be configured to: obtain the data associated with the at least one CT scan, where the CT scan is a non-contrast CT scan. The one or more processors configured to providePage 2 of 434896-3752-9172.106721 / 076333-1036 PATENT the data associated with the at least one CT scan to the model ensemble can be configured to: provide the data associated with the at least one CT scan to the first model, the first model comprising one or more layers configured to extract the feature map from the data associated with the at least one CT scan. The one or more processors configured to provide the data associated with the at least one CT scan to the model ensemble can be configured to: provide the data associated with the at least one CT scan to the first model, the first model comprising at least one first layer configured to extract an initial feature map having a first dimension from the at least one CT scan, and at least one second layer configured to extract the feature map having a second dimension that is less than the first dimension from the initial feature map.

[0008] The one or more processors can be further configured to: split the feature map into a set of patches; generate the data associated with the feature map based on the set of patches, and provide the data associated with the feature map to the second model based on generating the data associated with the feature map. The one or more processors configured to generate the data associated with the feature map based on the set of patches can be configured to: flatten each patch of the set of patches; and append a positional embedding to each patch of the set of patches, the positional embedding indicating a position of the patch relative to other patches of the set of patches. The second model can include an attention-based model; and the one or more processors can be further configured to: provide the data associated with the feature map to the second model based on appending the positional embedding to each patch of the set of patches.

[0009] In another embodiment, a computer-implemented method for detecting features in non-contrast computed tomography scans can include obtaining, by at least one processor, data associated with at least one computed tomography (CT) scan for at least one patient; providing, by the at least one processor, the data associated with the at least one CT scan to a model ensemble to cause the model ensemble to output an indication that a condition is present; and generating, by the at least one processor, data associated with a graphical user interface (GUI), the GUI comprising an indication of the condition represented by the CT scan. The model ensemble can include a first model that is configured to receive the data associated with the at least one CT scan as an input and generate a feature map based on the at least one CT scan, and a second model that is configured to receive data associated with the feature map as an input and generate the indicationPage 3 of 434896-3752-9172.106721 / 076333-1036 PATENT that the condition is present based on the feature map. At least a portion of the model ensemble can be trained using non-contrast CT scans that are generated based on corresponding CTPA scans.

[0010] Obtaining the data associated with the at least one CT scan can include: obtaining, by the at least one processor, the data associated with the at least one CT scan, where the CT scan is a non-contrast CT scan. Providing the data associated with the at least one CT scan to the model ensemble can include: providing, by the at least one processor, the data associated with the at least one CT scan to the first model, the first model comprising one or more layers configured to extract the feature map from the data associated with the at least one CT scan.

[0011] Providing the data associated with the at least one CT scan to the model ensemble can include: providing, by the at least one processor, the data associated with the at least one CT scan to the first model, the first model comprising at least one first layer configured to extract an initial feature map having a first dimension from the at least one CT scan, and at least one second layer configured to extract the feature map having a second dimension that is less than the first dimension from the initial feature map. In examples, the method can further include splitting, by the at least one processor, the feature map into a set of patches; generating, by the at least one processor, the data associated with the feature map based on the set of patches; and providing, by the at least one processor, the data associated with the feature map to the second model based on generating the data associated with the feature map. Generating the data associated with the feature map based on the set of patches can include: flattening, by the at least one processor, each patch of the set of patches; and appending, by the at least one processor, a positional embedding to each patch of the set of patches, the positional embedding indicating a position of the patch relative to other patches of the set of patches. The second model can include an attention-based model and the method can further include: providing, by the at least one processor, the data associated with the feature map to the second model based on appending the positional embedding to each patch of the set of patches.

[0012] In yet another embodiment, a non-transitory computer-readable medium for detecting features in non-contrast computed tomography scans can store instructions that, when executed by at least one processor, causes the at least one processor to obtain data associated with at least one computed tomography (CT) scan for at least one patient; provide the data associated with the at least one CT scan to a model ensemble to cause the model ensemble to output anPage 4 of 434896-3752-9172.106721 / 076333-1036 PATENT indication that a condition is present, the model ensemble and generate data associated with a graphical user interface (GUI), the GUI comprising an indication of the condition represented by the CT scan. The model ensemble can include a first model configured to receive the data associated with the at least one CT scan as an input and generate a feature map based on the at least one CT scan, and a second model that is configured to receive data associated with the feature map as an input and generate the indication that the condition is present based on the feature map. At least a portion of the model ensemble is trained using non-contrast CT scans that are generated based on corresponding CTPA scans.

[0013] The instructions that cause the at least one processor to obtain the data associated with the at least one CT scan can cause the at least one processor to: obtain the data associated with the at least one CT scan, where the CT scan is a non-contrast CT scan. The instructions that cause the at least one processor to provide the data associated with the at least one CT scan to the model ensemble can cause the at least one processor to: provide the data associated with the at least one CT scan to the first model, the first model comprising one or more layers configured to extract the feature map from the data associated with the at least one CT scan.

[0014] The instructions that cause the at least one processor to provide the data associated with the at least one CT scan to the model ensemble can cause the at least one processor to: provide the data associated with the at least one CT scan to the first model, the first model comprising at least one first layer configured to extract an initial feature map having a first dimension from the at least one CT scan, and at least one second layer configured to extract the feature map having a second dimension that is less than the first dimension from the initial feature map. The instructions can further cause the at least one processor to split the feature map into a set of patches; generate the data associated with the feature map based on the set of patches, and provide the data associated with the feature map to the second model based on generating the data associated with the feature map.

[0015] The instructions that cause the at least one processor to generate the data associated with the feature map based on the set of patches can cause the at least one processor to: flatten each patch of the set of patches; and append a positional embedding to each patch of the set of patches, the positional embedding indicating a position of the patch relative to other patches of the set of patches. The second model is an attention-based model; and the instructions can furtherPage 5 of 434896-3752-9172.106721 / 076333-1036 PATENT cause the at least one processor to provide the data associated with the feature map to the second model based on appending the positional embedding to each patch of the set of patches.

[0016] In an embodiment, a system for generating training data includes one or more processors configured to: obtain data associated with scans of a plurality of patients, each scan of the plurality of patients generated based on a first domain, wherein the scans comprise a first set of scans representing at least one condition that can be present in a patient and a second set of scans that do not represent the at least one condition that can be present in the patient; provide the data associated with the scans to a model to cause the model to generate an output, the output representing each scan of the plurality of scans in a second domain; and provide data associated with the output of the model to a system to cause the system to identify the at least one condition in at least one subsequent patient.

[0017] The first domain can be associated with computed tomography pulmonary angiography (CTPA) scans and the second domain can be associated with non-contrast computed tomography (CT) scans. The one or more processors can be configured to obtain the data associated with the scans of the plurality of patients are configured to: obtain a plurality of CTPA scans, and wherein the model is configured to receive data associated with CTPA scans and generate outputs representing the CTPA scans as non-contrast CT scans.

[0018] The one or more processors configured to provide the data associated with the scans to the model can be configured to: provide the data associated with the scans to a generative model comprising a generator system that is configured to receive the data associated with the scans and a noise vector as input and generate the output representing each scan of the plurality of scans in the second domain. The one or more processors configured to provide the data associated with the scans to a generative model can be configured to: provide the data associated with the scans to a generative model that is trained based on the generator system and a discriminator system, wherein the generator system and the discriminator system are trained together based on pairs of scans comprising at least one scan from the first set of scans and at least one second scan from a second set of scans in the second domain. The one or more processors configured to provide the data associated with the scans to a generative model can be configured to: provide the data associated with the scans to a generative model that is trained based on the generator system and a discriminator system, wherein the generator system and the discriminator system are trainedPage 6 of 434896-3752-9172.106721 / 076333-1036 PATENT together based on pairs of scans comprising at least one scan from the first set of scans and at least one second scan from a second set of scans in the second domain that corresponds to the at least one scan from the first domain. The generator system can include a first generator system, and the one or more processors configured to provide the data associated with the scans to the model can be configured to: provide the data associated with the scans to a generative model comprising the first generator system and a second generator system. The first generator system can be configured to receive the data associated with the scans and a noise vector as input and generate the output representing each scan of the plurality of scans in the second domain. And the second generator system can be configured to receive data associated with each scan of the plurality of scans in the second domain and a noise vector as input and generate an output representing each scan of the plurality of scans in the second domain as a scan in the first domain.

[0019] The one or more processors configured to provide the data associated with the scans to a generative model can be configured to: provide the data associated with the scans to a generative model that is trained based on the first generator system and the second generator system, wherein the first generator system and the second generator system are trained together during a plurality of cycles.

[0020] In an embodiment, a method can include obtaining, by at least one processor, data associated with scans of a plurality of patients, each scan of the plurality of patients generated based on a first domain, providing, by the at least one processor, the data associated with the scans to a model to cause the model to generate an output, the output representing each scan of the plurality of scans in a second domain; and providing, by the at least one processor, data associated with the output of the model to a system to cause the system to identify the at least one condition in at least one subsequent patient. The scans can include a first set of scans representing at least one condition that can be present in a patient and a second set of scans that do not represent the at least one condition that can be present in the patient.

[0021] In yet another embodiment, a non-transitory computer-readable medium stores instructions thereon that, when executed by at least one processor, cause the at least one processor to: obtain data associated with scans of a plurality of patients, each scan of the plurality of patients generated based on a first domain, provide the data associated with the scans to a model to cause the model to generate an output, the output representing each scan of the plurality of scans in aPage 7 of 434896-3752-9172.106721 / 076333-1036 PATENT second domain; and provide data associated with the output of the model to a system to cause the system to identify the at least one condition in at least one subsequent patient. The scans can include a first set of scans representing at least one condition that can be present in a patient and a second set of scans that do not represent the at least one condition that can be present in the patient.

[0022] By virtue of the implementation of the techniques described herein, more accurate determinations of whether PEs are present or not present can be generated using non-contrast CT scans. This, in turn, can reduce the probability that a clinician will arrive at a false-negative diagnosis, particularly in cases where the markers that would be present in a contrast-based CTPA scan are hidden or blended with other expected anatomical features in the non-contrast CT scan. Further, the need to use contrast dyes when treating patients can be reduced or eliminated, likewise reducing the chances that the patient will experience adverse health outcomes associated with the use of such dyes.

[0023] Because non-contrast chest CT scans are more readily available, involve lower radiation exposure, and do not need laboratory tests of renal function pre- and post-scan, the systems and methods described herein can improve the ability of facilities to diagnose PEs in a variety of situations where currently such diagnoses is not possible or prohibitively difficult. First, patients that visit hospitals and receive non-contrast chest CT scans can be evaluated in real-time or post-visit to determine whether PEs or other similarly-identifiable diseases are present. Second, patients can be evaluated in emergency scenarios where non-contrast chest CT scans were already performed or when medical facilities do not have access to contrast agents or advanced imaging scanners. In each of these scenarios, allowing for evaluation of non-contrast CT scans can result in more timely intervention and improved patient outcomes, ultimately reducing morbidity, mortality, and costs associated with managing later-treated PEs. The decreased reliance on contrast agents can also reduce the number of contraindications, adverse reactions related to contrast administration, and overall treatment cost. There are many clinical applications that these methods can be applied to. For example, it is difficult to quantify the four chamber sizes of the heart on non-contrast CT scans, and contrast CT scans are often used to visualize the four chambers. By implementing the techniques described herein, systems can be configured to segment the four chambers on contrast CT scans and then translate the contrast CT scans into non-contrast CT scans. The generated non-contrast CT scans and the four chamber maskers obtained from the contrast CTPage 8 of 434896-3752-9172.106721 / 076333-1036 PATENT scans can be used to train a model to infer the four chambers on non-contrast CT scans. Other similar applications include the quantification of the extent of fat of the liver for subjects diagnosed with fatty liver.

[0024] And by implementing the techniques as described herein, the scanning and diagnosis process can be performed more quickly when using the described model configuration as opposed to using conventional image processing techniques that are both limited ability to identify PEs and can require significant computing resources to arrive at a similar determination (whether a PE is or is not present). As such, the presently-disclosed systems can reduce the overall consumption of computing resources (e.g., processor usage and memory consumption) that would otherwise be required. Further, the presently-disclosed systems and methods obviate the need to perform multiple scans (e.g., non-contrast CT scans and corresponding CTPA scans) on patients at the point of treatment when generating training datasets, thereby conserving computing resources (e.g., processing and memory resources needed to perform the non-contrast CT and CTPA scans) that would otherwise be necessary to train the models as described herein. Additionally, the presently disclosed techniques can reduce or eliminate reliance on individuals for manual annotation of datasets, thereby allowing for the use of larger, publicly available datasets to augment training and testing of models as described herein.

[0025] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are intended to provide further explanation of the embodiments described herein.BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings constitute a part of this specification, illustrate one or more embodiments and, together with the specification, explain the subject matter of the disclosure.

[0027] FIG. 1 is a block diagram of an environment, in accordance with one or more embodiments described herein.

[0028] FIG. 2 is a flow diagram illustrating operations of a method for generating training datasets, in accordance with one or more embodiments described herein.Page 9 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0029] FIGS. 3A and 3B are a diagram of an example implementation of a system configured to generate training datasets, in accordance with one or more embodiments described herein.

[0030] FIG. 4 is an example of a set of CTPA scans and corresponding non-contrast CT scans, in accordance with one or more embodiments described herein.

[0031] FIG. 5 is an example of a set of annotated CTPA scans and corresponding segmented CTPA scans, in accordance with one or more embodiments described herein.

[0032] FIG. 6 is a flow diagram illustrating operations of a method for detecting features in non-contrast computed tomography scans, in accordance with one or more embodiments described herein.

[0033] FIGS. 7A and 7B are a diagram of an example implementation of a system for detecting features in non-contrast computed tomography scans, in accordance with one or more embodiments described herein.

[0034] FIG. 8 is a diagram of an example structure of a model ensemble in accordance with one or more embodiments described herein.DETAILED DESCRIPTION

[0035] Reference will now be made to the embodiments illustrated in the drawings, and specific language will be used here to describe the same. It will nevertheless be understood that no limitation of the scope of the disclosure is thereby intended. Alterations and further modifications of the features illustrated here, and additional applications of the principles as illustrated here, which would occur to a person skilled in the relevant art and having possession of this disclosure, are to be considered within the scope of the disclosure. While the terms “scans” and “images” are used throughout, it will be understood that such terms can be used interchangeably where contextually appropriate. Further, while the present disclosure describes techniques for training and implementing models and / or model ensembles to identify PEs in non-contrast CT scans, it will be understood that the techniques described herein are not limited to PEs and can be applied to identify one or more different features or conditions that are present and able to be detected in one domain but difficult to detect in another. Other example conditions can include, withoutPage 10 of 434896-3752-9172.106721 / 076333-1036 PATENT limitation, the presence or absence of internal bleeding, tumors, cancers, infections, blood vessel abnormalities, internal injuries, certain bone conditions such as fractures in smaller bones or complex structures (e.g., the spine), and / or the like.

[0036] FIG. l is a block diagram of an environment 100, described in accordance with one or more embodiments herein. The environment 100 can include a imaging device 110, a server 120, a model 130, client devices 140a-140n (referred to collectively as client devices 140 and individually as client device 140, unless otherwise specified), an image database 150, and a network 160. In some embodiments, the imaging device 110, server 120, client devices 140, and image database 150 can all interconnect (e.g., establish a connection to communicate with one another) via the network 160. The network 160 can be a wide area network (such as the Internet), a local area network (LAN), or any other kind of network.

[0037] The imaging device 110 can include one or more devices (including one or more computing devices) comprising hardware and software components capable of performing the various processes described herein. For example, the imaging device 110 can be any device including a memory and a processor capable of communicating, via the network 160 with one or more other devices of FIG. 1. In some implementations, the imaging device 110 includes one or more CT scanners, fluoroscopes, X-ray angiograph systems, and / or the like. In some embodiments, the imaging device 110 can be configured to be in communication with the server 120 or the client devices 140 via network 160. The imaging device 110 can provide (e.g., transmit) data to the server 120 and / or to the client devices 140 as described herein. In some embodiments, the imaging device 110 is associated with a medical facility such as a hospital or outpatient care facility.

[0038] The server 120 includes any computing device comprising hardware and software components capable of performing the various processes described herein. For example, the server 120 can be any device including a memory and processor capable of communicating, via the network 160 with one or more other devices of FIG. 1. Non-limiting examples of the server 120 includes data centers, server computers, workstation computers and / or the like. The server 120 is configured to be in communication with the imaging device 110, the client devices 140, and / or the image database 150 via network 160. In some embodiments, the server 120 is associated with one or more clinicians. In some embodiments, one or more of the operations described as beingPage 11 of 434896-3752-9172.106721 / 076333-1036 PATENT performed by the server 120 can be performed (e.g., implemented) by one or more client device 140 as described herein.

[0039] The server 120 may be associated with (e.g., generates, trains, updates, deploys, and / or the like) a model 130. The model can be associated with (e.g., include) one or more neural networks and / or other suitable network structures as described herein. In some embodiments, the model 130 is configured to receive data associated with operation of the imaging device 110 from the server 120 during training and / or updating of the model 130. For example, the model 130 can be configured to receive data associated with operation of the imaging device 110 from the server 120 as well as one or more indications (e.g., tags) that portions of the data represent one or more features (e.g., pulmonary embolisms, and / or the like). In this example, the server 120 can provide the data associated with the operation of the imaging device 110 to the model 130 to cause the model 130 to be trained and / or updated to classify portions of subsequently received data associated with the operation of the imaging device 110 (e.g., non-contrast CT scans) as representing one or more features (e.g., pulmonary embolism and / or the like). In some embodiments, the server 120 can provide data associated with the model 130 (e.g., data that represents the model 130, one or more weights of the model 130, and / or the like) to the client devices 140. For example, the server 120 can provide the data associated with the model 130 to the client devices 140 during deployment of the model 130.

[0040] The client devices 140 include any computing device comprising hardware and software components capable of performing the various processes described herein. For example, the client devices 140 can be any device including a memory and a processor capable of communicating, via the network 160 with one or more other devices of FIG. 1. Non-limiting examples of the client devices 140 include desktop computers, mobile devices (e g., cellular phones and tablets), and / or the like. The client devices 140 can be configured to be in communication with the imaging device 110 or server 120 via the network 160. In some embodiments, the client devices 140 are associated with users such as clinicians operating the imaging device 110 and / or the client devices 140 as described herein. In some embodiments, the client devices 140 can be the same as, or similar to (e.g., implemented by), the imaging device 110Page 12 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0041] The network 160 includes any computing device comprising hardware and software components capable of performing the various processes described herein. For example, the network 160 can be any device including a memory and processor capable of allowing interconnections between one or more of the devices of FIG. 1. Non-limiting examples of the network 160 includes devices associated with wired and / or wireless networks such as cellular networks (e.g., long-term evolution (LTE) networks, fourth generation (4G) networks, fifth generation (5G) networks, and / or the like), local area networks (LANs), wide area networks (WANs), the Internet, and / or the like.

[0042] FIG. 2 is a flow chart illustrating operations of a method 200 for generating training datasets, in accordance with one or more embodiments. In some implementations, one or more of the functions described with respect to method 200 can be performed (e.g., completely, partially, and / or the like) by a server that is the same as, or similar to, server 120. In some implementations, one or more of the functions described with respect to method 200 can be performed (e.g., completely, partially, and / or the like) by another device or group of devices separate from and / or including the server, such as by one or more imaging devices (e.g., imaging devices that are the same as, or similar to the imaging device 110 of FIG. 1) and / or one or more client devices (e.g., client devices that are the same as, or similar to, the client devices 140 of FIG. 1). While the present disclosure is discussed with respect to the generation of training datasets for identification of PEs in non-contrast CT scans, it will be understood that the techniques described herein are not limited to PEs and can be applied to identify one or more different conditions such as the presence or absence of internal bleeding, tumors, cancers, infections, blood vessel abnormalities, internal injuries, certain bone conditions such as fractures in smaller bones or complex structures (e.g., the spine), and / or the like. Further, while the present disclosure discusses the generation of training datasets based on scans in a first domain associated with CTPA scans and a second domain associated with non-contrast CT scans, it will be understood that the first domain and second domain are not intended to be limited to only CTPA and non-contrast CT scans. Non-limiting examples of pairs of domains can include: a first domain associated with greyscale images and a second domain associated with color images, a first domain associated with visible light and a second domain associated with infrared light, a first domain associated with light detection and ranging (LiDAR) images and a second domain associated with optical images, and / or the like.Page 13 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0043] At operation 202, the method 200 includes obtaining, by a server, a set of scans including CTPA scans and non-contrast CT scans. For example, the server can obtain pairs of images including CTPA scans and non-contrast CT scans where individual CTPA scans and noncontrast CT scans for a plurality of patients correspond to (e.g., are correlated with) each other. In other examples, the server can obtain sets of CTPA scans and non-contrast CT scans that are not correlated with each other (e.g., are generated by various CT scanners when scanning any number of patients to identify anomalies that are identified with the use of contrast scanning techniques). While the present disclosure describes the processing of CTPA scans and non-contrast CT scans, it will be understood that the description thereof is not intended to be limiting and that the presently-disclosed systems and techniques can be applied to a variety of scans and / or images in distinct domains. For example, the present disclosure can be extended to scans and / or images across a variety of domains where it is useful to transform scans and / or images from one domain to the other domain (and vice versa). For purposes of simplicity, with respect to the operations of method 200, a first domain can include one or more CTPA scans of patients that may or may not have at least one condition (e.g., a PE) and a second domain can include one or more non-contrast CT scans of patients that may or may not have the at least one condition.

[0044] At operation 204, the method 200 includes storing, by the server, the pairs of scans in a database. For example, the server can store the scans in a database. In some examples, the database can be accessible by multiple servers and can store scans for multiple patients. In some embodiments, the server can be configured to provide the scans to the server (or to different servers) that are training models that are the same as, or similar to, the model 330 described herein.

[0045] In some embodiments, the server can segment the CTPA scans. For example, the server can segment one or more regions within the CTPA scans and annotate portions of the CTPA scans (e.g., pixels and / or the like) to identify one or more anomalies. The CTPA scans can be segmented using a segmentation model (similar to those described herein) and / or manually based on input by one or more individuals (e.g., clinicians) that are trained to identify anomalies in contrast-based images such as CTPA scans. As described herein, the anomalies can include PEs or other features of CTPA scans that are identifiable with the use of a contrast dye or agent such as bone fractures, cancers, and / or the like. In this way, the server can automatically detect andPage 14 of 434896-3752-9172.106721 / 076333-1036 PATENT outline PEs within the CTPA scans to later train a model such as a segmentation model to generate corresponding annotations of the anomalies in the non-contrast CT scans.

[0046] At operation 206, the method 200 includes providing, by the server, the scans to a model to train the model to receive CTPA scans as an input and generate non-contrast CT scans as an output. For example, the server can provide individual pairs of scans to the model to train the model. During training, the model can be configured to generate non-contrast CT scans images based on CTPA scans. In this example, the model can include a generative model that is configured to receive data associated with scans in a first domain (e.g., CTPA scans and / or the like) as input and output data associated with the scans in a second domain (e.g., non-contrast CT scans).

[0047] In some embodiments, the model can include a generative adversarial network (GAN). For example, the model can include a GAN that includes two neural networks that are trained based on an adversarial training process. One network, the generator system, takes as input a scan in the first domain and a noise vector and outputs a representation of the scan in the second domain. The other network, the discriminator system, acts as a critic and analyzes both the scan that is output by the generator system and one or more scans that were obtained by the server (e.g., at operation 202). The discriminator system then determines that the output of the generator system is either generated by the generator system (e.g., a generated scan) or was obtained by the server (e g., a real scan). Based on the performance of the discriminator system, the generator system is trained to produce increasingly realistic scans in the second domain that are not identified as generated by the discriminator system.

[0048] In some embodiments, the model can include a GAN that includes a generator system that is configured to receive CTPA scans and output non-contrast CT scans, and a discriminator system that is configured to obtain the non-contrast CT scans from the generator system and output an indication of whether the non-contrast CT scans are or are not generated by the generator system as feedback during training of the generator system. For example, the generator system of the GAN can be configured to receive CTPA scans that may or may not represent at least one condition that can be present in a patient as well as a noise vector. The server can provide the CTPA scans and the noise vector(s) to the generator system to cause the generator system to output a non-contrast CT scan. The server can provide the output of the generator system to the discriminator system to cause the discriminator system to output an indication of whetherPage 15 of 434896-3752-9172.106721 / 076333-1036 PATENT the non-contrast CT scans are or are not generated by the generator system. In one example, the server can provide the output of the generator system (e.g., non-contrast CT scans that are generated based on CTPA scans) and a non-contrast CT scan that corresponds to the CTPA scan (e.g., a non-contrast CT scan that was generated by the imaging device with (e.g., prior to) the CTPA scan) to the discriminator system to cause the discriminator system to output the indication. In another example, the server can provide the output of the generator system (e.g., non-contrast CT scans that are generated based on CTPA scans) and a non-contrast CT scan that may not correspond to the CTPA scan to the discriminator system to cause the discriminator system to output the indication. In this way, the server can train the model based on individual image pairs that correspond to one another and / or the server can train the model based on a set of images in a first domain and a set of images in a second domain that do not necessarily correspond to one another. In these examples, the server can then update one or more weights of the model (e.g., one or more weights of the generator system and / or the discriminator system) based on whether the discriminator system accurately classified the output of the generator system as generated by the generator system. This set of operations can be iteratively performed until the proportion of accurate classifications to non-accurate classifications satisfies a threshold proportion (e.g., the model converges).

[0049] In some embodiments, the model can include a GAN that includes a first generator system and a second generator system. For example, the first generator system can be configured to receive CTPA scans as input and output non-contrast CT scans. In this example, the second generator system can be configured to receive the output of the first generator system (e.g., the non-contrast CT scans generated by the generator system) and output CTPA scans. In some embodiments, the server can then update one or more weights of the model (e.g., one or more weights of the first generator system and / or the second generator system) based on a difference between the CTPA scan input to the first generator system and the CTPA scan output by the second generator system. This set of operations can be performed during a plurality of cycles (e.g., iterations) until the difference between the CTPA scan input to the first generator system and the CTPA scan output by the second generator system satisfies a threshold difference (e.g., the model converges).Page 16 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0050] At operation 208, the method 200 includes obtaining, by the server, a dataset comprising a plurality of CTPA scans. For example, the server can obtain a dataset comprising a plurality of CTPA scans that were generated to diagnose whether the conditions described herein (e.g., PEs) were present in a patient. In some examples, the dataset can include scans that were generated based a plurality of patients. In other examples, the dataset can include one or more scans that were generated for a specific patient.

[0051] In some embodiments, the server can provide the CTPA scans from the dataset to the model to cause the model to generate non-contrast CT scans based on the CTPA scans. For example, the server can provide the CTPA scans to a GAN as described herein to cause the GAN to generate non-contrast CT scans in a second domain based on the CTPA scan in the first domain. In an example, the server can provide the CTPA scans to a generator system of a GAN that was trained with a discriminator system to cause the generator system to output the non-contrast CT scans. In another example, the server can provide the CTPA scans to a first generator system of a GAN that was trained with a second generator system to cause the first generator system to output the non-contrast CT scans. In these examples, the server can associate the non-contrast CT scan output by the generator system with the CTPA scan that the non-contrast CT scan is based on to enable other devices to determine a correlation between the CTPA scan and the non-contrast CT scan.

[0052] In some embodiments, the server can train a segmentation model to annotate the non-contrast CT scans to indicate the presence of one or more PEs. For example, the server can provide the segmented CTPA scans and corresponding non-contrast CT scans (generated during execution of operation 208) to a segmentation model such as a U-net and / or the like. In this example, the segmentation model can be trained based on the server providing the non-contrast CT scan as input to the segmentation model to generate an annotated version of the non-contrast CT scan. The annotations of the CTPA scan can then be used as a ground truth and compared to the annotations generated by the segmentation model. The server can then update the weights of the segmentation model based on a difference between the annotations in the non-contrast CT scan and the corresponding, ground truth annotations from the segmented CTPA scan and iteratively repeat the training process until the segmentation model converges.Page 17 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0053] At operation 210, the method 200 includes updating, by the server, the dataset comprising the plurality of CTPA scans. For example, the server can update the dataset by including the non-contrast CT scans in the dataset. In some examples, the server can also update the dataset such that each CTPA scan corresponds to the non-contrast CT scan generated by the model based on the CTPA scan.

[0054] FIGS. 3A and 3B are a diagram of an example implementation 300 of a system configured to generate training datasets, in accordance with one or more embodiments described herein. In some embodiments, one or more aspects of the implementation can be performed by an imaging device 310 that is the same as, or similar to, the imaging device 110 of FIG. 1; a server 320 that is the same as, or similar to, the server 120 of FIG. 1; and / or an image database 350 that is the same as, or similar to, the image database 150 of FIG. 1.

[0055] At operation 302, the imaging device 310 generates a set of CTPA scans and non- contrast CT scans. For example, the imaging device 310 can generate the CTPA scan and the non- contrast CT scans from one or more imaging devices 310 that include CT scanners. In some embodiments, the CT scanners can generate the CTPA scans and the non-contrast CT scans in association with one another. For example, the CT scanners can first generate one or more CT scans of a portion of a patient (e.g., a torso) and, upon administration of a contrast dye to the patient, repeat the scanning process to generate the CTPA scans of the same portion (e.g., the torso) of the patient. As will be understood, the set of CTPA scans and non-contrast CT scans can include individual scans and / or subsets of scans that correspond to a plurality of patients. Additionally, or alternatively, the set of CTPA scans and non-contrast CT scans can include individual scans and / or subsets of scans that represent a plurality of conditions for which contrast dye can be used to identify anomalies within respective CTPA scans as described herein.

[0056] At operation 304, the scans are stored in an image database 350. For example, the imaging device 310 can include an individual imaging device or a plurality of imaging devices that each provide data associated with the CTPA and non-contrast CT scans to the image database 350. In some embodiments, the scans that are stored in the image database 350 can be used to train a model such as the model 330, described below. For example, the scans can be stored in the image database 350 and later provided to a server 320 to allow the server 320 to train the model 330.Page 18 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0057] At operation 306, the server 320 receives the CTPA and non-contrast CT scans from the image database 350. For example, the server 320 can receive the CTPA and non-contrast CT scans from image database 350 during and / or after operation of the imaging device 310. In examples, the server 320 can receive CTPA and non-contrast CT scans where the scans are included as a dataset of CTPA scans and corresponding CT scans representing a variety of patients and / or a variety of conditions. In these examples, the dataset can include sets of CTPA scans and corresponding CT scans, where approximately 30% of the CTPA scans and corresponding CT scans are associated with a feature (e.g., a PE) and approximately 70% are not associated with the feature. In some embodiments, the server 320 can receive the CTPA scans and corresponding CT scans where the scans are annotated. For example, the scans can be annotated with identifiers for each scan, identifiers of one or more anatomical features represented by the scan (e.g., identifiers that a lung that is imaged is a left lung or a right lung), one or more identifiers of PEs (e.g., identifiers of one or more PEs as central or peripheral PEs), identifiers of acuteness, identifiers associated with quality assurance, and / or the like.

[0058] At operation 308, the server 320 trains a model 330. For example, the server can train the model 330 where the model includes a generative adversarial network. In one example, the model 330 can include a GAN that includes a generator network (referred to as a “generator” or a “generator system”) and a discriminator network (referred to as a “discriminator” or a “discriminator system”). In this example, the model 330 can be trained to receive CTPA scans and generate non-contrast CT scans. In other examples, the GAN can include multiple generator systems and / or multiple discriminator systems as described herein and can be trained to receive CTPA scans as an input and output non-contrast CT scans.

[0059] At operation 310, the server 320 updates the weights of the model 330 until the model 330 converges. For example, the server 320 can provide the CTPA and non-contrast CT scans to the model 330 to train the model 330. For example, at operation 310a, the server 320 can provide CTPA scans as an input to the model 330. Providing the CTPA scan to the model 330 can cause the model 330, at operation 310b, to generate an output. In this example, the output can include a non-contrast CT scan that is based on the corresponding CTPA scan provided as input. In this way, the server 320 can iteratively train the model 330 to receive CTPA scans and generate corresponding non-contrast CT scans.Page 19 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0060] In some examples where the model 330 includes a GAN, during training, the generator of the GAN can receive a random noise vector as input (e.g., a noisy image) and implement one or more convolutional layers and / or multi-layer perceptrons (MLPs) to generate an output (e.g., an image such as a non-contrast CT scan). The discriminator of the GAN can simultaneously be trained to receive real samples and the output of the generator and generate probability scores indicating a likelihood that a given sample is real or generated by the generator. The weights of the generator and discriminator can be iteratively updated until the GAN converges, at which point the probabilities output by the discriminator in response to receiving both real and generated samples indicate the outputs of the generator are indistinguishable by the discriminator. As will be understood, other features of GANs can be implemented to generate the output images. For example, a StyleGAN can be implemented. In this example, the StyleGAN can include one or more layers such as a mapping network layer that receives as input a noise vector and transforms a the noise vector into a latent code representing the image's style, a noise injection layer that introduces randomness for variation, adaptive instance normalization (AidAN) layers combined with convolutional layers to modulate activations of the convolutional layers based on a style code set by the mapping network to adjust features (e.g., color, texture, pose, and / or the like) and one or more convolutional layers to progressively upscale the resolution of the image to be output by the StyleGAN. The StyleGAN can also include a discriminator that is the same as, or similar to, the discriminators described herein.

[0061] In another example, a CycleGAN can be implemented. In this example, the GAN can include two generators are configured to take images from one domain and transform corresponding images in another domain. A first generator can be configured to receive CTPA scans and transform them into non-contrast CT scans, and a second generator can be configured to receive non-contrast CT scans and transform them into CTPA scans. During training, at a forward pass, a CTPA scan can be provided to the first generator to cause the first generator to provide an output including a non-contrast CT scan. The output can then be provided to the second generator to cause the second generator to provide an output including a CTPA scan. The CTPA scan output by the second generator can then be compared to the CTPA scan provided as input to the first generator to determine a loss (e.g., a difference) between the two scans. Once two generators converge and the loss satisfies a loss threshold, the first generator can be used asPage 20 of 434896-3752-9172.106721 / 076333-1036 PATENT described herein to convert images from a first domain (e.g., the CTPA domain) to a second domain (e.g., the non-contrast CT scan domain).

[0062] In yet another example, a BigGAN can be implemented. The BigGAN can be trained based on the CTPA images and corresponding non-contrast CT scans. For example, the BigGAN can include a generator and a discriminator as described herein. The generator can receive receives a noise vector along with a class label as input. In examples, the class label can indicate one or more anatomical features to illustrate, one or more features (e.g., PEs) to illustrate, and / or the like. A mapping network of the BigGAN can then transform the noise vector into an image. Random noise can be injected at each layer of the progressive growing architecture, and an AdalN layer can modulate activations of one or more convolutional layers based on the class information to influence the generation process (e.g., to include features represented by the label). The output image (e.g., a non-contrast CT scan) can then be provided to the discriminator of the BigGAN as well as the non-contrast CT scan that corresponds to the originally-input CTPA scan from the training data to cause the generator to output a probability score indicating how likely the generated is a real image as opposed to a generated image. A loss can then be calculated for both the generator and discriminator as described herein and the weights of both the generator and the discriminator can be updated until the BigGAN converges. The generator of the BigGAN can then be used as described herein to generate datasets for training purposes.

[0063] In some examples, the model 330 can be trained based on the CTPA and non- contrast CT scans. For example, where the model is a GAN, the server 320 can cause the model 330 to generate one or more images based on the CTPA and non-contrast CT scans. In this example, the server 320 can provide the CTPA scan as conditional input along with the noise vector to the generator system to cause the generator system to generate an output. The output of the generator system can then be compared to the non-contrast CT scan corresponding to the CTPA scan to determine a loss (e.g., a difference) between the output (the non-contrast CT scan generated by the generator system based on the CTPA scan) and the non-contrast CT scan. The weights of the generator system can be iteratively updated until the GAN converges. As described herein, the discriminator system can be trained along with the generator system until the model 330 converges (e.g., until the discriminator system is unable to determine a difference between the outputs of thePage 21 of 434896-3752-9172.106721 / 076333-1036 PATENT generator system and the non-contrast CT scans that were included in the dataset received by the server 320.

[0064] For a detailed description on various structures that can be implemented by the model 330 and techniques for training the model 330, reference can be made to Pu, et al., Automated detection and segmentation of pulmonary embolisms on computed tomography pulmonary angiography (CTPA) using deep learning but without manual outlining, Med Image Anal. (July 24, 2023) (PubMed ID: 37482032), available at https: / / pubmed.ncbi.nlm.nih.gov / 37482032 / , the contents of which are hereby incorporated by reference in their entirety. Reference can also be made to Pu, et al, Automated identification of pulmonary arteries and veins depicted in non-contrast chest CT scans, Med Image Anal. (Jan 12, 2022) (PubMed ID: 35066393), available at https: / / www.ncbi.nlm.nih.gov / pmc / articles / PMC8901546 / , the contents of which are hereby incorporated by reference in their entirety.

[0065] At operation 311, the server 320 receives a dataset of CTPA scans. For example, the server 320 can receive a dataset of one or more CTPA scans that are not correlated with non- contrast CT scans. As illustrated, the dataset of the one or more CTPA scans can be obtained from the image database 350. However, in some embodiments, the dataset of the one or more CTPA scans can be obtained from any number of datasets associated with any number of hospitals, research institutions, and / or the like. In this example, the server 320 can receive the dataset of CTPA scans to use when generating a dataset of non-contrast CT scans, where a portion of the non-contrast CT scans represent one or more PEs and are usable for training as described herein.

[0066] At operation 312, the server 320 provides the CTPA scans to the model 330 to cause the model 330 to generate corresponding non-contrast CT scans. For example, the server 320 can provide the CTPA scans to the model 330 where one or more CTPA scans represent portions of patients having a condition such as a PE. In this example, one or more CTPA scans can also represent portions of patients that do not have the condition. By providing the CTPA scans to the model 330, the server 320 can cause the model 330 to generate, as an output, non-contrast CT scans that represent portions of patients that have a condition (e g., PEs) and do not have the condition. The server 320 can then update the dataset of CTPA scans such that the dataset includes the originally-included CTPA scans as well as corresponding non-contrast CT scans for use during training as described herein.Page 22 of 434896-3752-9172.106721 / 076333-1036 PATENT

[0067] At operation 314, the server 320 transmits an updated dataset to the image database 350. For example, the server 320 can transmit the updated dataset to the image database 350 to store the dataset. In this example, the image database 350 can be configured to provide access to the updated dataset to the server 320 or any other suitable computing device for use during training as described herein.

[0068] FIG. 4 is an example of a set 400 of CTPA scans and corresponding non-contrast CT scans, in accordance with one or more embodiments described herein. As shown, the set 400 of CTPA scans 402a-402c and corresponding non-contrast CT scans 404a-404c. In some embodiments, a server (e.g., that is the same as, or similar to, the server 120 of FIG. 1), can curate a dataset including the CTPA scans 402a-402c and corresponding non-contrast CT scans 404a- 404c. For example, the server can curate a dataset including 100 CTPA scans and 100 non-contrast CT scans. The server can then train a GAN such as, for example, a cyclical generative adversarial network (CycleGAN) to generate non-contrast CT scans based on CTPA scans. As illustrated by FIG. 4, the anatomical structure in each of the transformed non-contrast CT images 404a-404c are visually similar to the original CTPA scans 402a-402c, and the arteries in the transformed non- contrast images were visually similar to non-contrasted arities.

[0069] FIG. 5 is an example of a set of annotated CTPA scans and corresponding segmented CTPA scans, in accordance with one or more embodiments described herein. In some embodiments, a server (e.g., that is the same as, or similar to, the server 120 of FIG. 1) can automatically detect and segment PEs depicted on CTPA images (e.g., in a single pass or in multiple passes). For example, the server 120 can implement a self-supervised learning strategy that does not require manually outlined PEs to train a model configured to perform one or more convolution functions (e.g., a convolutional neural network a U-net, and / or the like).

[0070] In some implementations, the server can identify PEs in the CTPAs based on implementing one or more computer vision techniques. For example, the server can identify PEs in the CTPAs by segmenting regions within the CTPAs using computer vision techniques such as thresholding, region-based segmentation, edge-based segmentation, and / or the like. The server can then assign a confidence level to each identification. The confidence level can represent a degree to which the segmented regions include a PE. While the present disclosure is discussed with respectPage 23 of 434896-3752-9172.106721 / 076333-1036 PATENT to PEs, the presently-disclosed techniques can be applied to identify any anomalous feature within an image such as a CT scan, CTPA, and / or the like.

[0071] In an implementation, the server can train a CNN-based segmentation model based on the segmented CTPAs. For example, the server can extract 3D image patches associated with the PEs that are assigned a high confidence that satisfies a threshold confidence value from the segmented regions within the set of CTPAs. The server can then train the CNN-based segmentation model using the CTPAs and the indications (e.g., segmentations) that portions of the scans correspond to PEs. In this example, the server can iteratively provide the CTPAs that are not segmented to the CNN to cause the CNN to generate outputs and, on each iteration, adjust the weights of the layers of the CNN based on a comparison of the indications of PEs represented by the output of the CNN to the indications of PEs in the segmented CTPAs. An example output is illustrated in FIG. 5, where images 502a-502c represent segmented CTPA scans (segmented using computer vision techniques) and images 504a-504c represent corresponding images that are segmented using a machine learning model that is trained as described above.

[0072] In some implementations, the CNN-based segmentation model can include a U-net model. For example, the CNN-based segmentation model can include a U-Net (e.g., a convolutional neural network specifically designed for image segmentation tasks). In these implementations, the U-net can include a set of layers corresponding to a contracting path that captures context and reduces image resolution and an expanding path that uses those features to locate and segment objects of interest in the high-resolution image. As an example, the U-net can include an encoder with multiple, stacked convolution layers that apply convolutional filters to an input of the U-net, non-linear activation functions (e g., rectified linear units (ReLUs)), and pooling layers that downsample the feature maps to reduce the spatial resolution of the input to the U-net. The U-net can also include a set of layers corresponding to an expanding path that reconstructs the image and localizes objects for segmentation. As an example, the U-net can include a decoder with multiple up-sampling blocks that are configured to perform one or more deconvolutions to increase feature map size that both from lower-dimension feature maps moving upward along the expanding path as well as feature maps forwarded at the same spatial resolution via skip connections (e.g., from corresponding encoders associated with the contracting path). In some embodiments, the U-net can further include a dynamic receptive field module that adjustsPage 24 of 434896-3752-9172.106721 / 076333-1036 PATENT the size of the receptive field of the U-net and a fusion upsampling module. For example, the dynamic receptive field module can allow the network to adaptively adjust the size of the area being analyzed within the image, enabling the U-net to capture both fine details and larger structures effectively. Additionally, the U-net can include a fusion upsampling module in the decoder path that refines the upsampled features by combining information from different scales (resolutions). This can, in turn, allow the U-net to generate more accurate segmentation boundaries compared when compared to skip connections described above, resulting in improved segmentation accuracy and a wider range of applicability for various medical image segmentation tasks and differentiation of semantically vague object boundaries. When implemented, the presently-described techniques can allow for a sensitivity of 81.5 (791 out of 970 PEs) and a false positive rate of 1.42 per scan. Additional details can be found in Beeche et al., “Super U-Net: a modularized generalizable architecture”, Pattern Recognit. 2022 Aug; 128: 108669. doi: 10.1016 / j.patcog.2022.108669. Epub 2022 Apr 1. PMID: 35528144; PMCID: PMC9070860, the contents of which are hereby incorporated by reference in their entirety.

[0073] FIG. 6 is a flow chart illustrating operations of a method 600 for detecting features in non-contrast computed tomography scans, in accordance with one or more embodiments. In some implementations, one or more of the functions described with respect to method 600 can be performed (e.g., completely, partially, and / or the like) by a server that is the same as, or similar to, server 120. In some implementations, one or more of the functions described with respect to method 600 can be performed (e.g., completely, partially, and / or the like) by another device or group of devices separate from and / or including the server, such as by one or more imaging devices (e.g., imaging devices that are the same as, or similar to the imaging device 110 of FIG. 1) and / or one or more client devices (e.g., client devices that are the same as, or similar to, the client devices 140 of FIG. 1). While the present disclosure is discussed with respect to the identification of PEs in non-contrast CT scans, it will be understood that the techniques described herein are not limited to PEs and can be applied to identify one or more different conditions such as the presence or absence of internal bleeding, tumors, cancers, infections, blood vessel abnormalities, internal injuries, certain bone conditions such as fractures in smaller bones or complex structures (e.g., the spine), and / or the like. Further, while the present disclosure discusses the detection of features based on scans in a first domain associated with CTPA scans and a second domain associated with non-contrast CT scans, it will be understood that the first domain and second domain are notPage 25 of 434896-3752-9172.106721 / 076333-1036 PATENT intended to be limited to only CTPA and non-contrast CT scans. Non-limiting examples of pairs of domains can include: a first domain associated with greyscale images and a second domain associated with color images, a first domain associated with visible light and a second domain associated with infrared light, a first domain associated with light detection and ranging (LiDAR) images and a second domain associated with optical images, and / or the like.

[0074] At operation 602, the server obtains data associated with at least one CT scan. For example, the server can obtain data associated with the at least one CT scan. The data associated with the at least one CT scan can be of at least a portion of a patient (e.g., one or more portions such as an abdomen, a neck, a head, or combinations thereof). In some embodiments, the data associated with the at least one CT scan can include data associated with a non-contrast CT scan. For example, the CT scan can be generated by an imaging device such as a CT scanner without administration of a contrast dye or other similar material that results in an increased contrast between anatomical features (e.g., blood vessels and surrounding tissue and / or the like). While the present disclosure is described with respect to CT scans, it will be understood that scans generated by fluoroscopes, X-ray angiograph systems, and / or the like can be implemented in addition to CT scanners.

[0075] At operation 604, the server provides the data associated with the at least one CT scan to a model ensemble. For example, the server can provide the data associated with the at least one CT scan to a model ensemble that is configured to receive CT scans and output predictions of the likelihood that one or more conditions are present in the patient. The model ensemble can be the same as, or similar to, the model ensemble 800 of FIG. 8. In some embodiments, the model ensemble can include a first model and a second model. For example, the model ensemble can be configured to receive the data associated with the at least one CT scan and provide the data to the first model to cause the first model to generate an output. In this example, the first model can be configured to receive the data associated with the CT scan, extract a feature map from the data associated with the at least one CT scan, and output data associated with the feature map. The server can cause the model ensemble to provide the data associated with the feature map to the second model to cause the second model to generate an output. For example, the server can cause the model ensemble to provide the data associated with the feature map to the second model to cause the second model to generate data associated with a prediction indicating a likelihood that aPage 26 of 434896-3752-9172.106721 / 076333-1036 PATENT feature is represented in the CT scan. As described herein, the prediction can indicate a likelihood that a PE is represented in the CT scan. However, the present disclosure is not limited to this feature.

[0076] In some embodiments, the first model can include multiple layers that are configured to receive the CT scan and extract one or more feature maps at one or more resolutions. For example, the first model can include a CNN and / or a U-net (e.g., a U-net having a dynamic receptive field module and / or a fusion upsampling module as described herein). In these examples, the first model can include multiple sets of layers that are configured to extract feature maps at varying resolutions (e.g., 128x128x128x2; 64x64x64x4; 32x32x32x8, and so on). In an example, the first model includes a first layer configured to receive the data associated with the CT scan and extract a first set of features that are associated with a first resolution (e.g., a first dimension) and a second layer configured to receive data output by the first layer and extract a second set of features that are associated with a second resolution (e.g., a second dimension). In this example, the second resolution can be less than the first resolution. In examples, the first model can include a third layer (or more) that is configured to receive data output by the second layer and extract a third set of features that are associated with a third resolution (e.g., a third dimension that is less than the first and second dimension).

[0077] In some embodiments, the server can be configured to update the data output by the first model. For example, where the data output by the first model represents a feature map of one or more images, the server can split the feature map into a set of patches (e.g., 4x4x4x8 3D image patches). The server can then provide data associated with the set of patches to the second model. In some embodiments, the server can flatten the set of patches. For example, the server can flatten the set of patches such that the set of patches are represented as a single patch prior to providing the single patch to the second model. In examples, the server can append one or more patch embeddings to each patch. For example, where the set of patches include 3D patches, the server can flatten each patch into a two-dimensional (2D) patch and append a patch embedding to each patch. The patch embedding can include a value representing a position of each patch relative to each other patch with regard to the set of 3D patches. In some examples, where the server receives sets of 3D patches (e.g., adjacent to one another), the server can further append aPage 27 of 434896-3752-9172.106721 / 076333-1036 PATENT positional embedding to each set of 3D patches (e.g., the group of flattened patches) prior to providing the data associated with the set of patches to the second model.

[0078] In some embodiments, the second model can be configured to receive the data associated with the set of patches and output a prediction indicating a likelihood that a feature is represented in the CT scan. For example, the second model can include an attention-based model (e g., a transformer network and / or the like) having multiple blocks as illustrated by FIG. 8. In this example, the second model can determine a query vector (e.g., representing a current element such as a flattened patch), a key (e.g., representing all of the elements such as all of the flattened patches), and a value (e.g., representing a value for each element in the sequence). The second model can then provide data associated with a plurality of corresponding queries and values to a first multi-head attention layer (a first “MatMul” layer) that is configured to perform one or more matrix multiplication operations based on the queries and values. The output of the first MatMul layer can then be provided to a scaled dot product layer (e.g., a “Scale” layer) that is configured to uses a dot product operation to calculate attention scores between queries and value vectors. The output of the Scale layer can be provided to a softmax layer that is configured to receive the attention scores and determine corresponding probabilities for each output of the Scale layer. The output of the softmax layer can then be provided to another MatMul layer that is the same as, or similar to, the earlier-described MatMul layer and the output merged. This process can be repeated through one or more iterations (represented as “Transformer Block[s]” in FIG. 8).

[0079] In some embodiments, the second model can include a feed forward neural network that is configured to receive the output of the attention-based model and generate the indication of the condition represented by the CT scan. For example, the feed forward neural network can obtain the output of the attention-based model and correlate the output with a prediction. In some examples, the prediction can be a binary prediction (e.g., that a feature is present or that a feature is not present). The server can then use the output of the feed forward neural network to generate a graphical user interface.

[0080] At operation 606, the server generates data associated with a graphical user interface (GUI) based on an output of the model ensemble. For example, the server can generate data associated with a GUI, where the GUI includes an indication of the condition represented by the CT scan. In this example, the server can determine the indication based on the output of thePage 28 of 434896-3752-9172.106721 / 076333-1036 PATENT second model. For example, the server can compare the output of the second model to one or more threshold values (or ranges of threshold values) to determine whether the output of the second model satisfies the threshold values (or ranges of threshold values). In this example, where the server determines that the output of the second model satisfies a threshold value (or range of threshold values) associated with a given state for the patient (e.g., that one or more PEs are present, that one or more PEs are not present, and / or the like) the server can correlate the state of the patient with the indication for the GUI. In these examples, the data associated with the GUI can be configured to cause a display device (e g., a display device of a client device that is the same as, or similar to, the client devices 140 of FIG. 1) to display the GUI such that the indication is output by the display device.

[0081] FIGS. 7A and 7B are a diagram of an example implementation 700 of a system configured to detect features in non-contrast computed tomography scans, in accordance with one or more embodiments described herein. In some embodiments, one or more aspects of the implementation can be performed by an imaging device 710 that is the same as, or similar to, the imaging device 110 of FIG. 1; a server 720 that is the same as, or similar to, the server 120 of FIG. 1; and / or a client device 740 that is the same as, or similar to, the client devices 140 of FIG.1

[0082] At operation 702, the imaging device 710 generates a non-contrast CT scan. For example, the imaging device can generate a non-contrast CT scan of a portion of a patient (e.g., an abdomen). In this example, the non-contrast CT scan can represent the presence or absence of one or more features. For example, the non-contrast CT scan can represent the presence or absence of one or more pulmonary embolisms affecting the patient.

[0083] At operation 704, the imaging device 710 transmits data associated with the non- contrast CT scan to the server 720. For example, the imaging device 710 can transmit the data associated with the non-contrast CT scan to the server 720 to cause the server to analyze the non- contrast CT scan and output an indication of whether the condition is present or not present in the patient.

[0084] At operation 706, the server 720 determines whether a PE is present. For example, the server 720 can determine whether a PE is present based on the server providing data associatedPage 29 of 434896-3752-9172.106721 / 076333-1036 PATENT with the non-contrast CT scan to a model 730. In this example, the model 730 can be the same as, or similar to, the model ensemble 800 of FIG. 8, described below.

[0085] In some embodiments, the server 720 can generate a graphical user interface (GUI). For example, the server 720 can generate the GUI based on the determination of whether the PE is present or not present. In an example, where the PE is present, the GUI can be generated such that the GUI indicates a visual indication (e.g., text and / or the like) that the PE is present. In examples, where the PE is present and the model 730 outputs an indication of where in the noncontrast CT scan the PE is located, the GUI can include a representation of the non-contrast CT scan representing the PE. In this example, the server 720 can generate the GUI such that one or more pixels corresponding to the PE are updated (e.g., highlighted, represented in accordance with a color value as opposed to a greyscale value, and / or the like) to indicate where the PE is located in the non-contrast CT scan.

[0086] At operation 708, the server 720 transmits data associated with a GUI. For example, the server 720 can transmit the data associated with the GUI to a client device 740. In this example, the server 720 can transmit the data associated with the GUI to a client device 740 to cause a display (e.g., a screen and / or the like) of the client device 740 to generate the GUI as an output.

[0087] At operation 709, the client device 740 displays the GUI indicating whether the PE is present. For example, the client device 740 can display the GUI based on the client device 740 receiving the data associated with the GUI. In this example, the data associated with the GUI can be configured to cause the client device 740 to control a display device such that the display device provides the GUI as output.

[0088] FIG. 8 is a diagram of an example structure of a model ensemble 800 in accordance with one or more embodiments described herein. In some embodiments, the model ensemble 800 can be implemented by a server that is the same as, or similar to, the server 120 of FIG. 1.

[0089] In some embodiments, the model ensemble 800 includes a first model 802 and a second model 804. The first model can include a neural network that is configured to receive data associated with scans (e.g., scans of at least a portion of a patient) and output a feature map. The second model can include an attention-based model (e.g., a transformer) that is configured to receive data associated with the feature map and output a prediction. In the examples described,Page 30 of 434896-3752-9172.106721 / 076333-1036 PATENT the prediction can include a binary representation of whether a feature (e.g., a condition of the patient) is represented by the non-contrast CT scan input to the model ensemble 800.

[0090] In some embodiments, a server implementing the model ensemble 800 can be configured to receive one or more non-contrast CT scans. The one or more non-contrast CT scans can be obtained based on operations of an imaging device when determining a diagnosis for a patient (e g., whether or not the patient has a PE). The server can then preprocess the non-contrast CT scans prior to providing the non-contrast CT scans to the model ensemble 800. For example, the server can crop, rotate, and / or otherwise adjust the non-contrast CT scans such that the non- contrast CT scans are configured to be received as input by the model ensemble 800. In some embodiments, to ensure consistent inputs, the non-contrast CT scans can be isotropicized with a uniform resolution (e.g., 1 x 1 x 1 mm3) and padded with a uniform dimension (e.g., 256x256x256 voxels).

[0091] In some embodiments, the server can then provide the non-contrast CT scans to the model ensemble 800. For example, the server can provide the non-contrast CT scans individually or as a set to the model ensemble 800. In this way, the server can cause the model ensemble 800 to analyze individual non-contrast CT scans or a set of non-contrast CT scans (e.g., contiguous scans) representing a portion of a patient). The server can then provide the non-contrast CT scan (or set of non-contrast CT scans) to the first model 802 of the model ensemble 800.

[0092] In some embodiments, the first model 802 can include a model that is configured to receive data associated with scans and output a feature map. For example, the first model 802 can include a model that is configured to receive the scans and output the feature map, where the feature map represents the scans at a resolution that is different from the resolution of the input scan. In an example, the first model 802 can include a convolutional neural network (CNN) that includes a first set of layers 802a, a second set of layers 802b, and a third set of layers 802c. The first set of layers 802a can be configured to receive the data associated with the scans input to the first model 802 and output a first feature map having a first resolution (e.g., 128x128x128x2). The second set of layers 802b can be configured to receive the first feature map output by the first set of layers 802a as input and output a second feature map having a second resolution (e.g., 64x64x64x4) that is reduced relative to the first feature map. The third set of layers 802c can be configured to receive the second feature map output by the second set of layers 802b as input andPage 31 of 434896-3752-9172.106721 / 076333-1036 PATENT output a third feature map having a third resolution (e.g., 32x32x32x8) that is reduced relative to the first feature map and the second feature map.

[0093] In some embodiments, the model ensemble 800 can be configured to process the output of the first model 802 prior to providing the output of the first model 802 to the second model 804. For example, where the output of the first model 802 includes a feature map that is represented as an image, the model ensemble 800 can be configured to split the feature map output by the first model 802 into image patches. The images patches can be at a resolution that is less than the resolution of the feature map output by the first model 802 (e.g., a resolution of 4x4x4x8). In some embodiments, the server can flatten the image patches. For example, the model ensemble 800 can flatten the image patches such that the set of patches are represented as a single patch prior to providing the single patch to the second model 804. The model ensemble 800 can append one or more patch embeddings to each patch. The patch embedding can include a value representing a position of each patch relative to each other patch with regard to the set of 3D patches. In some examples, where the model ensemble 800 can further be configured to append a positional embedding to each set of 3D patches prior to providing the data associated with the set of patches to the second model 804.

[0094] In some embodiments, the second model 804 of the model ensemble 800 can be configured to receive the data associated with the set of patches and output a prediction indicating a likelihood that a feature is represented in the CT scan. For example, the second model can include an attention-based model (e.g., a transformer network and / or the like) having multiple blocks. In this example, the second model 804 can determine a query vector, a key, and a value. The second model 804 can then provide data associated with a plurality of corresponding queries and values to a first multi-head attention layer that is configured to perform one or more matrix multiplication operations based on the queries and values. The output of the first MatMul layer can then be provided to a scaled dot product layer that is configured to uses a dot product operation to calculate attention scores between queries and value vectors. The output of the Scale layer can be provided to a softmax layer that is configured to receive the attention scores and determine corresponding probabilities for each output of the Scale layer. The output of the softmax layer can then be provided to another MatMul layer that is the same as, or similar to, the earlier-described MatMul layer and the output merged. This process can be repeated through one or more iterations. By virtuePage 32 of 434896-3752-9172.106721 / 076333-1036 PATENT of implementing an attention-based model as the second model 804, global information can be used to determine whether an anomaly carries valuable information (such as an unusual image pattern indicating a rare occurrence of the disease process) or is merely noise. Combining the local information extracted from the first model 802 and the global information derived from the second model 804 allows the feed-forward neural network to achieve a comprehensive understanding of data, leading to improved accuracy and robustness.

[0095] In some embodiments, the second model 804 of the model ensemble 800 can include a feed forward neural network that is configured to receive the output of the attentionbased model and generate the indication (e.g., a prediction) of the condition represented by the CT scan. For example, the feed forward neural network can obtain the output of the attention-based model and correlate the output with a prediction. In some examples, the prediction can be a binary prediction (e.g., that a feature is present or that a feature is not present). The server can then use the output of the feed forward neural network to generate a graphical user interface.

[0096] In some embodiments, the first model 802 of the model ensemble 800 can be trained to generate the feature map at the lower resolution. For example, the first model 802 can be trained independently on a large dataset of labeled images (e.g., a dataset that is the same as, or similar to, the updated dataset generated by the server 320 of FIG. 3B). This training can involve the server providing the images (e.g., the non-contrast CT scans) to the first model 802 to cause the first model 802 to output the feature map. The server can then determine a difference between the feature map and an expected feature map and update the weights of the first model 802 based on the difference to optimize the ability of the first model 802 ability to extract meaningful features from the input scans. In some embodiments, the first model 802 can be trained using an iterative selective learning (ISL) technique to mitigate the effects of data heterogeneity in a training set. The ISL technique can be efficiently implemented for any deep learning architecture. For example, the ISL begins by checking the performance of the first model 802 during training every k epochs (e.g., k=3) to identify cases that may be unsatisfactorily processed and insufficiently “learned” by the first model 802. In subsequent training epochs, these “unsatisfactory” cases with low Dice coefficients are “learned” multiple times via data augmentations. We expect that this biased or heterogeneous learning strategy will mitigate the effects of data heterogeneity in the training set. Once the CNN is trained, the model ensemble 800 is presented with non-contrast CT scans. ThePage 33 of 434896-3752-9172.106721 / 076333-1036 PATENT first model 802 extracts features, which are then processed and provided as input to the second model 804. During the ensemble training, the second model 802 is trained using backpropagation again, but this time the target is the presence or absence of a PE in the non-contrast CT scan. The second model 804 can be trained to interpret the features extracted by the first model 802 and identify patterns that signify the presence of the PE. This collaborative training refines both the feature extraction by the first model 802 and the predictions output by the second model 804, ultimately leading to a model ensemble 800 that can reliably detect PEs in non-contrast CT scans.

[0097] The foregoing method descriptions and the process flow diagrams are provided merely as illustrative examples and are not intended to require or imply that the steps of the various embodiments must be performed in the order presented. The steps in the foregoing embodiments can be performed in any order. Words such as “then,” “next,” etc., are not intended to limit the order of the steps; these words are simply used to guide the reader through the description of the methods. Although process flow diagrams can describe the operations as a sequential process, many of the operations can be performed in parallel or concurrently. In addition, the order of the operations can be re-arranged. A process can correspond to a method, a function, a procedure, a subroutine, a subprogram, and the like. When a process corresponds to a function, the process termination can correspond to a return of the function to a calling function or a main function.

[0098] Some non-limiting embodiments of the present disclosure are described herein in connection with a threshold. As described herein, satisfying a threshold can refer to a value being greater than the threshold, more than the threshold, higher than the threshold, greater than or equal to the threshold, less than the threshold, fewer than the threshold, lower than the threshold, less than or equal to the threshold, equal to the threshold, and / or the like.

[0099] No aspect, component, element, structure, act, step, function, instruction, and / or the like used herein should be construed as critical or essential unless explicitly described as such. In addition, as used herein, the articles "a" and "an" are intended to include one or more items and can be used interchangeably with "one or more" and "at least one." Furthermore, as used herein, the term "set" is intended to include one or more items (e.g., related items, unrelated items, a combination of related and unrelated items, etc.) and can be used interchangeably with "one or more" or "at least one." Where only one item is intended, the term "one" or similar language is used. Also, as used herein, the terms "has," "have," "having," or the like are intended to be openPage 34 of 434896-3752-9172.106721 / 076333-1036 PATENT ended terms. Further, the phrase "based on" is intended to mean "based at least partially on" unless explicitly stated otherwise.

[0100] The various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

[0101] Embodiments implemented in computer software can be implemented in software, firmware, middleware, microcode, hardware description languages, or any combination thereof. A code segment or machine-executable instructions can represent a procedure, a function, a subprogram, a program, a routine, a subroutine, a module, a software package, a class, or any combination of instructions, data structures, or program statements. A code segment can be coupled to another code segment or a hardware circuit by passing and / or receiving information, data, arguments, parameters, or memory contents. Information, arguments, parameters, data, etc. can be passed, forwarded, or transmitted via any suitable means including memory sharing, message passing, token passing, network transmission, etc.

[0102] The actual software code or specialized control hardware used to implement these systems and methods is not limiting. Thus, the operation and behavior of the systems and methods were described without reference to the specific software code being understood that software and control hardware can be designed to implement the systems and methods based on the description herein.

[0103] When implemented in software, the functions can be stored as one or more instructions or code on a non-transitory computer-readable or processor-readable storage medium. The steps of a method or algorithm disclosed herein can be embodied in a processor-executable software module, which can reside on a computer-readable or processor-readable storage medium.Page 35 of 434896-3752-9172.106721 / 076333-1036 PATENTA non-transitory computer-readable or processor-readable media includes both computer storage media and tangible storage media that facilitate transfer of a computer program from one place to another. A non-transitory processor-readable storage media can be any available media that can be accessed by a computer. By way of example, and not limitation, such non-transitory processor- readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other tangible storage medium that can be used to store desired program code in the form of instructions or data structures and that can be accessed by a computer or processor. Disk and disc, as used herein, include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk, and Blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media. Additionally, the operations of a method or algorithm can reside as one or any combination or set of codes and / or instructions on a non-transitory processor-readable medium and / or computer- readable medium, which can be incorporated into a computer program product.

[0104] The preceding description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the disclosure. Thus, the present disclosure is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the following claims and the principles and novel features disclosed herein.

[0105] While various aspects and embodiments have been disclosed, other aspects and embodiments are contemplated. The various aspects and embodiments disclosed are for purposes of illustration and are not intended to be limiting, with the true scope and spirit being indicated by the following claims.Page 36 of 434896-3752-9172.1

Claims

06721 / 076333-1036 PATENTCLAIMSWhat is claimed is:

1. A system comprising: one or more processors configured to: obtain data associated with at least one computed tomography (CT) scan for at least one patient; provide the data associated with the at least one CT scan to a model ensemble to cause the model ensemble to output an indication that a condition is present, the model ensemble comprising: a first model that is configured to receive the data associated with the at least one CT scan as an input and generate a feature map based on the at least one CT scan, and a second model that is configured to receive data associated with the feature map as an input and generate the indication that the condition is present based on the feature map; and generate data associated with a graphical user interface (GUI), the GUI comprising an indication of the condition represented by the CT scan, wherein at least a portion of the model ensemble is trained using non-contrast CT scans that are generated based on corresponding CTPA scans.

2. The system of claim 1, wherein the one or more processors configured to obtain the data associated with the at least one CT scan are configured to: obtain the data associated with the at least one CT scan, where the CT scan is a noncontrast CT scan.

3. The system of claim 1, wherein the one or more processors configured to provide the data associated with the at least one CT scan to the model ensemble are configured to: provide the data associated with the at least one CT scan to the first model, the first model comprising one or more layers configured to extract the feature map from the data associated with the at least one CT scan.Page 37 of 434896-3752-9172.106721 / 076333-1036 PATENT4. The system of claim 1 , wherein the one or more processors configured to provide the data associated with the at least one CT scan to the model ensemble are configured to: provide the data associated with the at least one CT scan to the first model, the first model comprising at least one first layer configured to extract an initial feature map having a first dimension from the at least one CT scan, and at least one second layer configured to extract the feature map having a second dimension that is less than the first dimension from the initial feature map.

5. The system of claim 4, wherein the one or more processors are further configured to: split the feature map into a set of patches; and generate the data associated with the feature map based on the set of patches, and provide the data associated with the feature map to the second model based on generating the data associated with the feature map.

6. The system of claim 5, wherein the one or more processors configured to generate the data associated with the feature map based on the set of patches are configured to: flatten each patch of the set of patches; and append a positional embedding to each patch of the set of patches, the positional embedding indicating a position of the patch relative to other patches of the set of patches.

7. The system of claim 6, wherein the second model is an attention-based model; and wherein the one or more processors are further configured to: provide the data associated with the feature map to the second model based on appending the positional embedding to each patch of the set of patches.

8. A method comprising: obtaining, by at least one processor, data associated with at least one computed tomography (CT) scan for at least one patient; providing, by the at least one processor, the data associated with the at least one CT scan to a model ensemble to cause the model ensemble to output an indication that a condition is present, the model ensemble comprising:Page 38 of 434896-3752-9172.106721 / 076333-1036 PATENT a first model that is configured to receive the data associated with the at least one CT scan as an input and generate a feature map based on the at least one CT scan, and a second model that is configured to receive data associated with the feature map as an input and generate the indication that the condition is present based on the feature map; and generating, by the at least one processor, data associated with a graphical user interface (GUI), the GUI comprising an indication of the condition represented by the CT scan, wherein at least a portion of the model ensemble is trained using non-contrast CT scans that are generated based on corresponding CTPA scans.

9. The method of claim 8, wherein obtaining the data associated with the at least one CT scan comprises: obtaining, by the at least one processor, the data associated with the at least one CT scan, where the CT scan is a non-contrast CT scan.

10. The method of claim 8, wherein providing the data associated with the at least one CT scan to the model ensemble comprises: providing, by the at least one processor, the data associated with the at least one CT scan to the first model, the first model comprising one or more layers configured to extract the feature map from the data associated with the at least one CT scan.

11. The method of claim 8, wherein providing the data associated with the at least one CT scan to the model ensemble comprises: providing, by the at least one processor, the data associated with the at least one CT scan to the first model, the first model comprising at least one first layer configured to extract an initial feature map having a first dimension from the at least one CT scan, and at least one second layer configured to extract the feature map having a second dimension that is less than the first dimension from the initial feature map.

12. The method of claim 11, further comprising: splitting, by the at least one processor, the feature map into a set of patches; andPage 39 of 434896-3752-9172.106721 / 076333-1036 PATENT generating, by the at least one processor, the data associated with the feature map based on the set of patches, and providing, by the at least one processor, the data associated with the feature map to the second model based on generating the data associated with the feature map.

13. The method of claim 12, wherein generating the data associated with the feature map based on the set of patches comprises: flattening, by the at least one processor, each patch of the set of patches; and appending, by the at least one processor, a positional embedding to each patch of the set of patches, the positional embedding indicating a position of the patch relative to other patches of the set of patches.

14. The method of claim 13, wherein the second model is an attention-based model; the method further comprising: providing, by the at least one processor, the data associated with the feature map to the second model based on appending the positional embedding to each patch of the set of patches.

15. A system comprising: one or more processors configured to: obtain data associated with scans of a plurality of patients, each scan of the plurality of patients generated based on a first domain, wherein the scans comprise a first set of scans representing at least one condition that can be present in a patient and a second set of scans that do not represent the at least one condition that can be present in the patient; provide the data associated with the scans to a model to cause the model to generate an output, the output representing each scan of the plurality of scans in a second domain; and provide data associated with the output of the model to a system to cause the system to identify the at least one condition in at least one subsequent patient.Page 40 of 434896-3752-9172.106721 / 076333-1036 PATENT16. The system of claim 15, wherein the first domain is associated with computed tomography pulmonary angiography (CTPA) scans and the second domain is associated with non-contrast computed tomography (CT) scans, wherein the one or more processors configured to obtain the data associated with the scans of the plurality of patients are configured to: obtain a plurality of CTPA scans, and wherein the model is configured to receive data associated with CTPA scans and generate outputs representing the CTPA scans as non-contrast CT scans.

17. The system of claim 15, wherein the one or more processors configured to provide the data associated with the scans to the model are configured to: provide the data associated with the scans to a generative model comprising a generator system that is configured to receive the data associated with the scans and a noise vector as input and generate the output representing each scan of the plurality of scans in the second domain.

18. The system of claim 17, wherein the one or more processors configured to provide the data associated with the scans to a generative model are configured to: provide the data associated with the scans to a generative model that is trained based on the generator system and a discriminator system, wherein the generator system and the discriminator system are trained together based on pairs of scans comprising at least one scan from the first set of scans and at least one second scan from a second set of scans in the second domain.

19. The system of claim 17, wherein the one or more processors configured to provide the data associated with the scans to a generative model are configured to: provide the data associated with the scans to a generative model that is trained based on the generator system and a discriminator system, wherein the generator system and the discriminator system are trained together based on pairs of scans comprising at least one scan from the first set of scans and at least one second scan from a second set of scans in the second domain that corresponds to the at least one scan from the first domain.Page 41 of 434896-3752-9172.106721 / 076333-1036 PATENT20. The system of claim 17, wherein the generator system is a first generator system, and wherein the one or more processors configured to provide the data associated with the scans to the model are configured to: provide the data associated with the scans to a generative model comprising the first generator system and a second generator system, wherein the first generator system is configured to receive the data associated with the scans and a noise vector as input and generate the output representing each scan of the plurality of scans in the second domain, and wherein the second generator system is configured to receive data associated with each scan of the plurality of scans in the second domain and a noise vector as input and generate an output representing each scan of the plurality of scans in the second domain as a scan in the first domain.

21. The system of claim 20, wherein the one or more processors configured to provide the data associated with the scans to a generative model are configured to: provide the data associated with the scans to a generative model that is trained based on the first generator system and the second generator system, wherein the first generator system and the second generator system are trained together during a plurality of cycles.Page 42 of 434896-3752-9172.1

Images