Method and system for generating text description of an image
By classifying and segmenting elements of image data, detecting exceptions using domain-independent models, and generating text descriptions for specific fields, it solves the problem that existing systems are difficult to combine images and text, and improves the utilization rate and description efficiency of image data.
Patent Information
- Application Number
- CN201980033783.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-05-24
- Filing Date
- 2019-05-13
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2039-05-13
AI Technical Summary
It is difficult for existing inference systems to directly combine image data in semantic networks with text data, resulting in image data being ignored when answering questions. The existing tools require a large amount of labeling data and calculation time, making it difficult to generate image text descriptions in specific fields.
By classifying image data into categories of specific domains, segmenting image elements, and detecting statistical anomalies using domain-independent models, generating field-independent text phrases, and then converting them into specific domain descriptions, and generating text descriptions of images based on context information.
It realizes the generation of text descriptions covering a wide range of images without the need for a large amount of labeled data and calculation time, which improves the utilization rate of image data in the inference system and saves time and effort in building specialized detectors.
Smart Images

Figure CN112136155B_ABST
Abstract
Description
Technical Field
[0001] The present invention generally relates to methods, systems, and computer program products for generating textual descriptions of images. More specifically, the present invention relates to methods, systems, and computer program products for generating textual descriptions of images using domain independent anomaly analysis. Background Art
[0002] There are a variety of artificial intelligence-based reasoning systems for analyzing text-based or semantic network-based representations of data to extract desired information from the data, such as determining answers to questions about a given topic associated with the data. These reasoning systems typically include applications executed on data processing hardware. Sometimes, the problems posed to these systems also have image components. For example, some questions from medical license examinations include annotated anatomical diagrams or radiological images. Unfortunately, there is no direct way to convert these semantically-based (e.g., language-based) networks to provide reasoning systems with the ability to analyze image data as well as text data to determine answers to extract desired information (such as answers to questions related to text data and image data). Therefore, existing reasoning systems typically ignore these images, although they have obvious utility for answering questions. Summary of the invention
[0003] Illustrative embodiments provide a method, system, and computer program product. An embodiment of a method for generating a text description of an image includes: classifying an image represented by image data into domain-specific categories; and segmenting one or more elements in the image data based on the domain-specific categories. Embodiments also include comparing each of the one or more elements to a domain-independent model to detect one or more statistical anomalies in the one or more elements, and characterizing the one or more detected statistical anomalies using one or more domain-independent text phrases. Embodiments also include converting one or more domain-independent text phrases into one or more domain-specific descriptions based on the domain-specific categories.
[0004] One embodiment further comprises outputting the one or more domain-specific descriptions.Another embodiment further comprises receiving the image data, and receiving context information indicating a context of the image data.
[0005] In an embodiment, the classification of the image into a domain-specific category is based on context information. In another embodiment, the context information includes metadata associated with the image data. In another embodiment, the context includes the morphology and viewpoint of the image data.
[0006] In another embodiment, segmenting the one or more elements uses a domain-specific class-specific segmentation process. In another embodiment, segmenting the one or more elements uses a segmentation process specific to the element. In yet another embodiment, segmenting the one or more elements aligns the generally segmented elements with a coarse model associated with a domain-specific class.
[0007] In another embodiment, the domain-independent model is a contour-based shape model. In another embodiment, the domain-independent model is a texture-based internal model.
[0008] In another embodiment, characterizing the detected statistical anomalies includes noting missing or under-represented elements. In another embodiment, characterizing the detected statistical anomalies includes over-annotated or over-represented elements.
[0009] Embodiments include a computer usable program product. A computer usable program product includes one or more computer readable storage devices and program instructions stored on at least one of the one or more storage devices.
[0010] Embodiments include a computer system comprising one or more processors, one or more computer-readable memories, and one or more computer-readable storage devices, and program instructions stored on at least one of the one or more storage devices for execution by at least one of the one or more processors via at least one of the one or more memories. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] Certain novel features which are believed to be characteristic of the invention are set forth in the appended claims. However, the invention itself, as well as the preferred mode of use, further objects and advantages thereof, will be best understood by reference to the following detailed description of illustrative embodiments when read in conjunction with the accompanying drawings in which:
[0012] Figure 1 depicts a block diagram of a network of data processing systems in which the illustrative embodiments may be implemented;
[0013] Figure 2 depicts a block diagram of a data processing system in which the illustrative embodiments may be implemented;
[0014] Figure 3 depicts a block diagram of an example configuration of an application for generating a text description of an image using domain-independent anomaly analysis in accordance with an illustrative embodiment;
[0015] Figure 4 depicts a block diagram of an example configuration of an application for generating textual descriptions of medical images using domain-independent anomaly analysis according to another illustrative embodiment;
[0016] Figure 5A-5B depicts an example of a medical image analysis surface primitive in accordance with an illustrative embodiment;
[0017] Figure 6A-6B depicts an example of a medical image analysis pipe primitive in accordance with an illustrative embodiment;
[0018] Figure 7A-7B depicts an example of a medical image analysis volumetric three-dimensional (3D) object primitive in accordance with an illustrative embodiment; and
[0019] Figure 8 Depicted is a flowchart of an example process for generating a textual description of an image using domain-independent anomaly analysis in accordance with an illustrative embodiment. DETAILED DESCRIPTION
[0020] The illustrative embodiments described herein generally relate to generating textual descriptions of images using domain-independent anomaly analysis.The illustrative embodiments recognize that currently available tools or solutions for generating textual descriptions of images do not address these needs or provide solutions for these needs.
[0021] One existing approach for generating textual descriptions of images is to identify the image as a member of some scene class, such as "beach," "grass," or "city scene." Weaker versions of this process exist that attempt only partial descriptions of the image, such as "indoor" or "x-ray," which describe the environment more than the content of the scene itself. While such partial descriptions are clearly better than treating the image data as a large "packet of bits," these partial descriptions are typically not specific enough to allow intelligent reasoning.
[0022] Another existing approach for generating text descriptions of images is to characterize the image by the objects it contains. This characterization can be general such as "person" or "animal", or more specific such as "John Smith" and "Labradoodle". Unfortunately, specific models must be learned for each fine-grained category, so such approaches require a large amount of labeled training data and computation time.
[0023] Even listing objects may not be particularly helpful to some reasoning systems. For example, indicating that a particular medical diagnosis is a "chest x-ray" and that it includes "left lung," "right lung," and "heart" has little or no diagnostic value. For such a system, even finer-grained concepts must be trained, such as "lung nodule," "infarction," or "astrocytoma." Thus, a potentially excessively large number of categories will typically be required to describe medical images.
[0024] Existing specialized medical image analysis systems are available for various organs, conditions, and modalities, such as cardiac blood flow, electrocardiogram (EKG), cell counts. However, these specialized systems require carefully hand-coded programs to be effective. In addition, there are machine learning systems for detecting and / or diagnosing specific organs, conditions, or modalities, such as melanoma, breast cancer, and diabetic retinopathy. However, these machine learning systems require many accurately labeled training examples. Further, there are deep learning image-to-text generation systems that are primarily used for web photos rather than medical images. However, these existing deep learning image-to-text generation systems require supervised training data.
[0025] The illustrative embodiments used to describe the present invention generally address and cope with the above-mentioned problems and other problems associated with generating textual descriptions of images.
[0026] One or more embodiments described herein provide an application that first classifies an image into certain broad categories. From the broad categories, it is expected that there are certain elements of scene composition and content. These elements are modeled in a domain-independent manner. For example, in a specific exemplary model, it is not "beach sand" that is the model, but "an extended horizontal light strip, the color of which may be light yellow near the middle of the image". In one or more embodiments, these models contain regular conditions, such as "level", which have been statistically derived from multiple labeled instances. In an embodiment, the system attempts to find candidates for each specified element and records any anomalies about its model. For example, the system can identify that there are colored areas in the image that look similar to the model, but are slightly diagonal rather than purely horizontal. Once the anomaly has been detected in the domain-independent model, the system uses appropriate domain-dependent terms to re-characterize the anomaly. For example, when the sand portion of a beach image is diagonal, it is usually evidence of the presence of a "pit". Therefore, the entire image can be given a text description of "beach view of the bay".
[0027] One or more embodiments provide the advantage that a small set of domain-independent model types can potentially cover all images that the system may encounter. For example, 3D "speckles", 2D "surfaces", and 1D "tubes" are sufficient for almost all visual objects. This is in contrast to specialized representations such as "endocardial wall" for computed tomography (CT) scans of the heart and "red blood cells" for peripheral blood smears.
[0028] In one or more embodiments, it may be advantageous to describe an image as some base class plus deviations from that base class. In the medical field, such a description may be particularly useful because, while not all abnormalities are pathological, often pathology causes the abnormalities. For example, detecting a lung as having a 3D mass with "rounded corners" rather than sharp corners on the distal edge (a domain-independent characterization) can be used to infer the presence of fluid accumulation in the costonephric angle (a domain-dependent characterization). Similarly, erroneous texture, color, or count of some elements (all domain-independent characterizations) often have diagnostic ramifications. In one or more embodiments, detecting such conditions in a domain-independent manner in a first pass saves the time and effort required to build a much larger number of more specialized detectors that would otherwise be required.
[0029] In an embodiment, an application receives an input image and applies a set of domain-specific classifiers to determine a wide-domain classification of the image. In this embodiment, the application selects a plurality of domain-independent element models based on the classification results. In an embodiment, the application attempts to segment elements in the input image corresponding to these models, and examines the found elements to determine domain-independent anomalies or deviations of elements from domain-independent models. In one or more embodiments, an element is a feature found within an image that has an expected appearance, position, size, and / or shape. For example, an element may include an expected appearance, position, size, and / or shape of a body organ. In one or more embodiments, a statistical anomaly may include a deviation from the expected occurrence position, size, and / or shape of an element by more than a determined amount.
[0030] In this embodiment, the application generates a domain-independent anomaly description that describes the characteristics of the anomaly in a domain-independent manner. In an embodiment, the application translates or recharacterizes the domain-independent anomaly description into a domain-specific text description of the anomaly that includes domain-specific terms. In an embodiment, the application generates an output text description representing the domain-specific text description. In one or more embodiments, the domain-specific text description may include the location of the anomaly, such as one or more of the following: an identification of an area near a landmark of a body organ (e.g., the "top" of a lung), a portion of the contour of the anomaly (e.g., the "concentric angle"), or the overall anatomical coordinates of the anomaly (e.g., "left", "far").
[0031] In one or more embodiments, communities for many visual properties can be learned. For example, general appearance terms such as "dark," "muddy," and "striped" can be learned. In another example, geometric terms such as "swollen," "narrowed," and "enlarged" can be learned.
[0032] In a number of specific embodiments, the initial classification of an image using a domain-specific classifier need not necessarily be performed based on the appearance of the image itself. In one or more embodiments, the initial classification may be based on alternative information, additional information, or metadata associated with the image (rather than replacing the image data or in addition to the image data). For example, the text of a question containing the image may state "Fetal ultrasound taken to confirm." In this instance, the application may infer the basic properties of the image directly from the text alone. In another example, suppose that the image is found to reside on the disk of an x-ray machine in a podiatrist's office. The application may infer that this information is strong evidence that the category of the image being examined is "X-ray of the foot."
[0033] Merely by way of example, the illustrative embodiments are described with respect to certain types of image data, classifiers, models, domains, devices, data processing systems, environments, components, and applications. Any particular representation of these and other similar products is not intended to limit the invention. Any suitable representation of these and other similar products may be selected within the scope of the illustrative embodiments.
[0034] Furthermore, the illustrative embodiments may be implemented with respect to any type of data, data source, or access to a data source over a data network. Any type of data storage device may provide data to embodiments of the present invention locally at a data processing system or over a data network within the scope of the present invention. Where embodiments are described using a mobile device, any type of data storage device suitable for use with a mobile device may provide data to such embodiments locally at the mobile device or via a data network within the scope of the illustrative embodiments.
[0035] The illustrative embodiments are described only as examples using specific codes, designs, architectures, protocols, layouts, schematics, and tools, and examples are not limited to the illustrative embodiments. In addition, for the sake of clarity of description, only specific software, tools, and data processing environments are used as examples to describe the illustrative embodiments in some instances. The illustrative embodiments can be used in conjunction with other comparable or similarly used structures, systems, applications, or architectures. For example, within the scope of the present invention, other similar mobile devices, structures, systems, applications, or their architectures can be used in conjunction with such embodiments of the present invention. The illustrative embodiments can be implemented in hardware, software, or a combination thereof.
[0036] The examples in this disclosure are only for clarity of description and are not limited to the illustrative embodiments. Additional data, operations, actions, tasks, activities, and manipulations can be imagined from this disclosure, and these additional data, operations, actions, tasks, activities, and manipulations are contemplated within the scope of the illustrative embodiments.
[0037] Any advantages listed herein are merely examples and are not intended to limit these illustrative embodiments. Additional or different advantages may be achieved by specific illustrative embodiments. In addition, specific illustrative embodiments may have some, all, or none of the advantages listed above.
[0038] With reference to the drawings, in particular to Figure 1 and 2 , these figures are exemplary diagrams of data processing environments in which illustrative embodiments may be implemented. Figure 1 and 2 These are examples only and are not intended to assert or imply any limitation with regard to the environments in which different embodiments may be implemented. Specific embodiments may make numerous modifications to the depicted environments based on the following description.
[0039] Figure 1 A block diagram of a network of data processing systems in which the illustrative embodiments may be implemented is depicted. Data processing environment 100 is a network of computers in which the illustrative embodiments may be implemented. Data processing environment 100 includes network 102. Network 102 is a medium used to provide communication links between different devices and computers connected together within data processing environment 100. Network 102 may include connections such as wired, wireless communication links, or fiber optic cables.
[0040] Client or server are merely example roles for certain data processing systems connected to network 102 and are not intended to exclude other configurations or roles for these data processing systems. Server 104 and server 106 are coupled to network 102 along with storage unit 108. In one or more embodiments, storage unit 108 may be configured to store one or more models, domain-specific classifiers, domain-independent component models, domain-independent anomaly descriptions, domain-specific anomaly descriptions, and / or component descriptions in database 109. Software applications may be executed on any computer in data processing environment 100. Clients 110, 112, and 114 are also coupled to network 102. Data processing systems such as servers 104 or 106 or clients 110, 112, or 114 may contain data and may have software applications or software tools executed thereon.
[0041] This is for example purposes only and does not imply any limitation on such architecture. Figure 1Certain components that may be used in an example implementation of an embodiment are depicted. For example, servers 104 and 106 and clients 110, 112, 114 are depicted as servers and clients only as an example, without implying limitation to a client-server architecture. As another example, an embodiment may be distributed across several data processing systems and data networks as shown, while another embodiment may be implemented on a single data processing system within the scope of the illustrative embodiments. Data processing systems 104, 106, 110, 112, and 114 also represent example nodes in a cluster, partitions, and other configurations suitable for implementing an embodiment.
[0042] Device 132 is an example of a device described herein. For example, device 132 may take the form of a smartphone, a PDA, a tablet computer, a laptop computer, a fixed or portable form of client 110, a wearable computing device, or any other suitable device. Figure 1 Any software application executed in another data processing system in may be configured to execute in device 132 in a similar manner. Figure 1 Any data or information stored or generated in another data processing system in may be configured to be stored or generated in device 132 in a similar manner.
[0043] The application 105 of the server 104 implements an embodiment of a process for generating a text description of an image using domain-independent anomaly analysis as described herein. The cognitive application 107 of the server 106 implements a cognitive reasoning system as described herein with respect to different embodiments. The imaging device 116 is configured to capture image data representing a subject or object (such as a body part) for which a text description is desired to be generated. In a number of specific embodiments, the imaging device 116 may include a camera, an x-ray machine, a scanner, or any other suitable imaging device. In other embodiments, the imaging device 116 may be omitted, and the image data may be obtained from another suitable source.
[0044] Servers 104 and 106, storage unit 108, and clients 110, 112, and 114 may be coupled to network 102 using wired connections, wireless communication protocols, or other suitable data connectivity. Clients 110, 112, and 114 may be, for example, personal computers or network computers.
[0045] In the depicted example, server 104 may provide data, such as boot files, operating system images, and applications to clients 110, 112, and 114. In this example, clients 110, 112, and 114 may be clients to server 104. Clients 110, 112, 114, or some combination thereof, may include their own data, boot files, operating system images, and applications. Data processing environment 100 may include additional servers, clients, and other devices not shown.
[0046] In the depicted example, data processing environment 100 may be the Internet. Network 102 may represent a collection of networks and gateways that use Transmission Control Protocol / Internet Protocol (TCP / IP) and other protocols to communicate with each other. The core of the Internet is a backbone of data communication links between master nodes or host computers, including thousands of commercial, government, educational and other computer systems that route data and messages. Of course, data processing environment 100 may also be implemented as many different types of networks, such as an intranet, a local area network (LAN), or a wide area network (WAN). Figure 1 This is intended as an example, and not as an architectural limitation for the different illustrative embodiments.
[0047] Among other uses, data processing environment 100 can be used to implement a client-server environment in which illustrative embodiments can be implemented. The client-server environment enables software applications and data to be distributed across a network so that the application works by using the interaction between the client data processing system and the server data processing system. Data processing environment 100 can also adopt a service-oriented architecture in which interoperable software components distributed across a network can be packaged together as consistent business applications. Data processing environment 100 can also take the form of a cloud, and adopt a cloud computing model of service delivery so as to facilitate on-demand network access to a shared pool of configurable computing resources (e.g., networks, network bandwidth, servers, processing, memory, storage, applications, virtual machines and services), which can be quickly configured and released with minimal management effort or interaction with a service provider.
[0048] See also Figure 2 , which depicts a block diagram of a data processing system in which the illustrative embodiments may be implemented. Data processing system 200 is an example of a computer, such as Figure 1 Servers 104 and 106, or clients 110, 112, and 114, or another type of device in which computer usable program code or instructions implementing the processes may be located for use in the illustrative embodiments.
[0049] Data processing system 200 also represents a data processing system or configuration thereof in which computer usable program code or instructions implementing the processes of the illustrative embodiments may be located. Data processing system 200 is described as a computer only as an example and is not limited thereto. Implementations in other device forms may modify data processing system 200, such as by adding a touch interface, and even eliminate certain depicted components from data processing system 200, without departing from the general description of the operation and functionality of data processing system 200 described herein.
[0050] In the depicted example, data processing system 200 employs a hub architecture including north bridge and memory controller hub (NB / MCH) 202 and south bridge and input / output (I / O) controller hub (SB / ICH) 204. Processing unit 206, main memory 208, and graphics processor 210 are coupled to north bridge and memory controller hub (NB / MCH) 202. Processing unit 206 may include one or more processors and may be implemented using one or more heterogeneous processor systems. Processing unit 206 may be a multi-core processor. In some implementations, graphics processor 210 may be coupled to NB / MCH 202 via an accelerated graphics port (AGP).
[0051] In the depicted example, a local area network (LAN) adapter 212 is coupled to the south bridge and I / O controller hub (SB / ICH) 204. An audio adapter 216, a keyboard and mouse adapter 220, a modem 222, a read-only memory (ROM) 224, a universal serial bus (USB) and other ports 232, and a PCI / PCIe device 234 are coupled to the south bridge and I / O controller hub 204 via bus 238. A hard disk drive (HDD) or solid-state drive (SSD) 226 and a CD-ROM 230 are coupled to the south bridge and I / O controller hub 204 via bus 240. The PCI / PCIe devices 234 may include, for example, Ethernet adapters, add-in cards, and PC cards for notebook computers. PCI uses a card bus controller, while PCIe does not. ROM 224 may be, for example, a flash binary input / output system (BIOS). Hard disk drive 226 and CD-ROM 230 may use, for example, an integrated drive electronics (IDE), a serial advanced technology attachment (SATA) interface, or variations such as external SATA (eSATA) and micro SATA (mSATA). Super I / O (SIO) device 236 may be coupled to south bridge and I / O controller hub (SB / ICH) 204 via bus 238.
[0052] Memories such as main memory 208, ROM 224, or flash memory (not shown) are some examples of computer usable storage devices. Hard disk drives or solid state drives 226, CD-ROM 230, and other similarly available devices are some examples of computer usable storage devices, including computer usable storage media.
[0053] An operating system runs on processing unit 206. The operating system coordinates and provides Figure 2 The operating system may be a commercially available operating system for any type of computing platform, including but not limited to server systems, personal computers, and mobile devices. Object-oriented or other types of programming systems may operate in conjunction with the operating system and provide calls to the operating system from programs or applications executing on data processing system 200.
[0054] for operating systems, object-oriented programming systems, and applications or programs such as Figure 1 205 and document authoring application 107), are located on a storage device, such as in the form of code 226A on hard drive 226, and can be loaded into at least one of one or more memories, such as main memory 208, for execution by processing unit 206. The processes of the illustrative embodiments may be performed by processing unit 206 using computer-implemented instructions, which may be located in a memory, such as main memory 208, ROM 224, or one or more peripheral devices.
[0055] In addition, in one case, code 226A can be downloaded from remote system 201B via network 201A, where similar code 201C is stored on storage device 201D. In another case, code 226A can be downloaded to remote system 201B via network 201A, where downloaded code 201C is stored on storage device 201D.
[0056] Figure 1-2 The hardware in may vary depending on the implementation. Figure 1-2 In addition to or in lieu of the hardware depicted Figure 1-2 In addition to the hardware depicted in the drawings, other internal hardware or peripheral devices may be used, such as flash memory, equivalent non-volatile memory, or optical disk drives, etc. In addition, the processes of the illustrative embodiments may be applied to a multi-processor data processing system.
[0057] In some illustrative examples, data processing system 200 may be a personal digital assistant (PDA), which is typically configured with flash memory to provide non-volatile memory for storing operating system files and / or user-generated data. The bus system may include one or more buses, such as a system bus, an I / O bus, and a PCI bus. Of course, the bus system may be implemented using any type of communication structure or architecture that provides data transfer between different components or devices attached to the structure or architecture.
[0058] The communications unit may include one or more devices for sending and receiving data, such as a modem or a network adapter. A memory may be, for example, main memory 208 or a cache, such as that found in north bridge and memory controller hub 202. A processing unit may include one or more processors or CPUs.
[0059] Figure 1-2 The examples depicted in and above-described are not meant to imply architectural limitations. For example, data processing system 200 may be a tablet computer, laptop computer, or telephone device in addition to taking the form of a mobile or wearable device.
[0060] Where a computer or data processing system is depicted as a virtual machine, virtual device, or virtual component, the virtual machine, virtual device, or virtual component uses virtualized representations of some or all of the components depicted in data processing system 200 to operate in the manner of data processing system 200. For example, in the virtual machine, virtual device, or virtual component, processing unit 206 is represented as a virtualized instance of all or some number of hardware processing units 206 available in a host data processing system, main memory 208 is represented as a virtualized instance of all or some portion of main memory 208 available in the host data processing system, and disk 226 is represented as a virtualized instance of all or some portion of disk 226 available in the host data processing system. In this case, the host data processing system is represented by data processing system 200.
[0061] See also Figure 3 , which depicts a block diagram of an example configuration of an application 302 for generating text descriptions of images using domain-independent anomaly analysis in accordance with an illustrative embodiment. The application 302 is Figure 1. The application 302 includes a domain-specific classifier component 308 configured to receive an input image 304 and text and / or metadata 306. The input image 304 is an image containing one or more features or elements that are expected by a domain-specific text description. The text and / or metadata 306 includes text, metadata, or other additional information indicating the context of the input image 304, such as the morphology or viewpoint of the input image 304. In particular embodiments, the modality of the image includes the type of imaging device used to acquire structural or functional images of a body part, including but not limited to radiography, ultrasound, nuclear medicine, computed tomography (CT), magnetic resonance imaging (MRI), and visible light.
[0062] The domain-specific classifier component 308 is configured to apply a set of one or more domain-specific classifiers to determine a broad classification of the input image 304 based on the input image 304 and the text and / or metadata 306. In this embodiment, the domain-independent element model component 310 is configured to select one or more domain-independent element models based on the classification results. The one or more domain-independent element models include models of elements or features that can be found in the input image 304. In an embodiment, the element segmentation component 312 attempts to segment elements in the input image 304 corresponding to the one or more domain-independent element models. In one or more embodiments, an element is a feature with an expected appearance, position, size and / or shape found within an image. For example, an element may include an expected appearance, position, size and / or shape of a body organ. The domain-independent anomaly detection component 314 is configured to examine the segmented elements to determine / detect domain-independent statistical anomalies or deviations of the segmented elements based on the domain-independent model. In one or more embodiments, the statistical anomaly may include deviations from the expected position, size and / or shape of the element exceeding an expected or acceptable amount.
[0063] In an embodiment, the domain-independent anomaly description component 316 generates a domain-independent anomaly description that describes characteristics of the statistical anomaly in a domain-independent manner. In a particular embodiment, the domain-independent anomaly description includes a textual description of the anomaly using terms that are not specific to a particular domain.
[0064] In an embodiment, the domain-specific re-characterization component 318 translates or re-characterizes the domain-independent anomaly description into a domain-specific textual description of the statistical anomaly, including terms specific to the determined specific domain to describe the statistical anomaly. In an embodiment, the textual description component 320 uses the domain-specific terms to generate an output textual description representing the domain-specific textual description to describe the statistical anomaly.
[0065] See also Figure 4, which depicts a block diagram of an example configuration of an application 402 for generating textual descriptions of medical images using domain-independent anomaly analysis according to another illustrative embodiment. In this embodiment, the application 402 is configured to perform operations related to medical image interpretation to identify abnormal regions or portions of body parts in medical images.
[0066] In an embodiment, application 402 receives image 404 and modality and viewpoint information 406. In a specific embodiment, from a source such as Figure 3 A domain-specific classifier, such as the domain-specific classifier 308 of , receives as output modality and viewpoint information 406. In various specific embodiments, for medical images, the modality and viewpoint information 406 typically consists of several parts of classification information. In specific embodiments, one part of the classification information is the type of radiology or image used, such as X-ray, CT, positron emission tomography (PET), MRI, microscopy, etc. Another part of the classification information is a portion of the anatomical structure within the image 404, such as the foot, heart, kidney, brain, etc. Another part of the classification information is the imaging direction used. For example, a chest X-ray can be acquired from the front to back direction ("anterior-posterior" or "AP") with the patient lying on a table, or from the side ("lateral") direction with the patient standing. The elements visible in these two viewpoints are sometimes different, and their expected appearance (position, shape, texture, etc.) is also different.
[0067] In an embodiment, based on the morphology and viewpoint information 406, the expected elements 410 in the image 404 are looked up in a semantic database within the database 109. For example, for an AP chest X-ray, the left lung, right lung, heart, diaphragm, and clavicle can be expected to be seen. These elements are expected to have a specific rough characterization 412 including rough appearance, size, shape, and / or location. For example, the left lung appears on the right side of the AP X-ray as a large vertically elongated area, darker in color, and approximately half the height of the image. There are similar rough characterizations for other expected elements that are looked up from the graphical database within the database 109. The input image 404 is subdivided into possible regions based on apparent appearance using structural segmentation 408 to form candidate regions. In an embodiment, the candidate regions are matched with the rough descriptions to form correspondences and labels 414 between the expected element names and the image regions.
[0068] In an alternative embodiment, instead of a coarse graphical description, a more detailed model such as an active shape model (ASM) or an active appearance model (AAM) may be retrieved for each expected element, and an associated more complex segmentation method applied to the input image to locate each element. It should be understood that the methods described herein (such as ASM and AAM) are not specific to the medical field, but can be used with many different types of objects (e.g., airplanes, cars, and horses).
[0069] In an embodiment, the partially segmented image is examined in a domain-independent manner for anomalies. In a specific embodiment, a comparison is made to ensure that all expected elements are present, and it is determined whether the expected elements are missing / under-represented in the correct amount 416. For example, if the right lung cannot be found, it may be that the right lung has collapsed and therefore the absence should be noted. Similarly, if a different object (e.g., a bullet) is found that is not a member of the expected element list, the additional / over-represented element should also be recorded 418. For some types of images, it is not enough to be present or absent, but the expected number of elements of some type is important. For example, in the case of peripheral blood smear microscopy images and red blood cells-for every M other elements (e.g., white blood cells), N of these elements should be seen. In a specific example, the domain-independent text description of such anomalies may include "missing element 2" or "excessive number of elements 7".
[0070] In this embodiment, a more detailed analysis of the image element also examines the shape and texture of the element. In an embodiment, a statistical model of the associated image element is retrieved from the database 109 to identify the statistical profile 422 of the element and to identify the abnormal segment 420 of the element. In one embodiment, the application 402 uses a contour-based shape model to compare each element. In another embodiment, the application 402 uses a texture-based internal model to compare each element. In an embodiment, a statistical appearance model is further retrieved from the database 109 to identify the statistical appearance 426 of the element, thereby determining the abnormal area 424 of the statistical appearance 426. In a number of specific embodiments, the statistical model learns by acquiring the average appearance and likelihood of various patterns of changes on appropriate elements in multiple "healthy" example images of the same type. For example, an abnormality can be marked when the detected change is more than two standard deviations away from the mean of the set. Alternatively, in a more complex analysis, an abnormality can be declared when the likelihood of the observed change is less than one percent (1%) in the principal component analysis (PCA) decomposition of the training example. Example initial domain-independent descriptions of these types of anomalies may include "excessive curvature on boundary segment 12" or "insufficient striated texture in region 8." In certain embodiments, overall characterizations such as "element 4 is too large" or "element 23 is mottled texture rather than transparent" are also possible.
[0071] Applicant 402 is also configured to convert domain-independent abnormal descriptions into domain-related text. For example, if too few round structures with indentations are found in a peripheral blood smear, this can be translated into the medically relevant text string "Low RBC" as a domain-specific text representation 428 of a missing / underrepresented element 416, indicating a low red blood cell (RBC) count. Similarly, if an unexpected dark element appears within an oval structure in a CT image (computed tomography) of the abdomen, then this observation can be converted into the phrase "deep nodule in left kidney" as a domain-specific text representation 430 of an additional / overrepresented element 418.
[0072] Abnormalities in contour shape and internal appearance may also be converted into more meaningful forms. For example, if a segment of an object's contour bulges outward at a certain location, this abnormality is first converted into the domain-independent term "bulge in segment 45 of the contour of element 6." Then, due to the known identity of the element with the abnormal contour and the location of the aberration, an unexpected finding in a leg X-ray may be rendered as "a bulge near the femoral head" as a domain-specific text representation 432 of the abnormal segment element 422. In another example, a portion of a large oval object was found to be darker than expected in a chest X-ray. Due to the known identity of the containing element and the location of the abnormality within this element, the description may be converted from just "dark area in element 9" to "shaded lower lobe" as a domain-specific text representation 434 of the abnormal area element 424. In one or more embodiments, such a domain-specific text description not only converts the category of the abnormality into a specific term, but also provides the associated anatomical location of the abnormality.
[0073] One or more embodiments rely on the use of domain-independent element models and domain-independent methods to detect anomalies in elements determined from an input image. In one or more embodiments, most items of interest can be modeled as one or more of three types: surfaces (primarily two-dimensional), tubes (primarily one-dimensional), and objects or blobs (primarily three-dimensional). In a number of specific embodiments, a specific type of measurement is associated with each of these types and, based on the nature of the measurement, used in the process of finding anomalies.
[0074] See now Figure 5A-5B , these figures depict examples of medical image analysis surface primitives in accordance with illustrative embodiments. Figure 5A An example of a surface primitive depicting the diaphragm containing an AP chest x-ray. The basic measure of a surface primitive is its curvature pattern 502, for example, starting out straight and then curving sharply downward. In an embodiment, the left pattern of curvature 502 is not represented by words, but by a Figure 5B504. In the example surface primitive for the diaphragm, the curvature pattern may be characterized by starting at one end and moving along the profile for optimal alignment. In one embodiment, a collection of such graphs from healthy individuals defines a "normal band". In this embodiment, portions of the test stop profile derived from the input image that fall outside of the band are considered abnormal.
[0075] In other embodiments, curvature is not the only property that can be evaluated for a surface. Another possibility is to observe image patches 503A and 503B on either side of the profile to determine the texture of the aperture. In a number of specific embodiments, image patches 503A and 503B can be represented by their grayscale mean, standard deviation of intensity, and distribution of edge directions. Again, statistical analysis can be performed on patches 503A and 503B (or differences in patches across a surface) to provide a quantitative definition of "normal".
[0076] In some embodiments, "normal" does not have to be defined only with respect to other individuals. Many anatomical structures appear in pairs, both visible in the same image. Thus, the left side of the diaphragm may appear flat compared to the entire image acquisition, but it is similar in curvature to the right side of the diaphragm seen in that same image. So, while somewhat unusual, it is not necessarily labeled abnormal. In another example, the relative position of the diaphragm can be examined with respect to a landmark.
[0077] See now Figure 6A-6B , these figures depict examples of medical image analysis pipeline primitives in accordance with illustrative embodiments. Fig. 6A An example of a tube primitive is depicted that includes a blood vessel 602. In one or more embodiments, the tube primitive is a linearly extended term. A tube generally has an orientation that can vary along the length of the tube, just as the curvature of a surface can vary. Perhaps more interestingly, a tube has a characteristic width that varies along an axis of symmetry. Thus, the measurement of the tube primitive is the vessel width. Figure 6B A graph of vessel width variation 604 of a counter along a vessel 602 is depicted. The graph of vessel width variation 604 can be compared to similar graphs obtained from healthy individuals to highlight abnormalities. For example, unusual narrowing in a coronary artery is a sign of atherosclerosis and is therefore an important aspect of the domain-specific text description presenting an abnormality. Another abnormality that may be identified is an uneven texture that may indicate, for example, a fracture. Other abnormalities may be identified using axial shape or relative position (such as looking for kinks, stretching, or branching of the vessel). Other abnormalities may be identified by looking for deflections relative to landmarks.
[0078] See now Figure 7A-7B , these figures depict examples of medical image analysis volumetric three-dimensional (3D) object primitives in accordance with illustrative embodiments. Fig. 7AAn example of a 3D object primitive for a lung is depicted. Similar to a surface primitive, a 3D object primitive has a profile with curvature that can be plotted as a function of position relative to some starting reference point. Similar to a tube, a 3D object primitive has a width that can be plotted, perhaps as a function of radial angle from the object's centroid rather than axial tube position. Fig. 7A A radial direction 702 of a 3D object primitive of a lung is shown. Figure 7B Shows the drawing distance Fig. 7A A graph of radial distances 704 of the centroids of 3D object primitives may be used to develop a statistical model of the population, thereby providing guidance on normal versus abnormal shapes.
[0079] In addition to boundaries, objects also have interior regions whose appearance can be characterized in a variety of ways. In a number of specific embodiments, the small image patches 703A-703C can be described by their intensity distribution and response to the convolution texture kernel. In addition, the local location of each patch 703A-703C is known, so a local characterization of the expected appearance can also be developed. That is, just because a particular patch is darker than the average of all patches extracted from the lungs, it is not necessarily abnormal. If the patch is from near the center, the patch may be darker given such a regional model.
[0080] Likewise, as described for surface primitives, comparisons can be performed between corresponding contralateral elements (rather than just an aggregate model derived from numerous other individuals). For example, in the case of the lungs, the distance between the inner border and the center of the spine may appear to be within normal limits on both sides. However, if one lung is significantly closer to that line than the other, this may be a noteworthy abnormality.
[0081] Although the examples given are mainly from the medical field, the method can be used with other types of images. Consider an image that was originally classified as a "beach scene" (whether by appearance, associated text, or other means). In view of this category, it is expected to find a light band at the bottom of the image ("sand"), a dark blue band near the middle of the image ("ocean"), and a light blue area above the middle ("sky"). As previously discussed, the deviation in the sand part can instead lead to a graphical description of a "cove". Assuming that the sand area usually has spots of bright color (blank and umbrella-shaped) therein, if these spots are missing, the description will become "empty beach" (not just "missing spots"). Similarly, it can be expected that the sky area has several amorphous white spots ("clouds") therein. If there are too many of these white spots compared to the statistical mean of the category, the overall description can be changed to "stormy beach" (not just "too many white spots"). Finally, the sky area may typically show a pale yellow circular element ("sun"). If a yellowish circular element is present, but the wrong green color, then a text description "Beach scene with green sun" is generated (as opposed to "wrong color element in top area").
[0082] See also Figure 8 , which depicts a flow diagram of an example process 800 for generating a text description of an image using domain-independent anomaly analysis, according to an illustrative embodiment. In one or more embodiments, process 800 may be implemented in application 105. In other embodiments, process 800 may be implemented in client device 110 or device 132.
[0083] In block 802, the application 105 receives input image data containing one or more features or elements for which a domain-specific textual description of the one or more features or elements is desired. In block 804, the application 105 receives context information indicating a context of the input image data. In various embodiments, the context information includes text and / or metadata associated with the input image data. In various embodiments, the context of the input image data includes a morphology and a viewpoint of the input image data.
[0084] In block 806, the application 105 classifies the overall image represented by the input image data into domain-specific categories based on the contextual information. In block 808, the application 105 segments one or more elements in the image based on the domain-specific categories. In a particular embodiment, segmenting the one or more elements uses a segmentation process specific to the domain-specific categories. In another specific embodiment, segmenting the one or more elements uses a segmentation process specific to the elements. In another embodiment, segmenting the one or more elements aligns the generally segmented elements with a coarse model associated with the category.
[0085] In block 810, the application 105 compares each of the one or more elements to a domain-independent model to detect statistical anomalies in the one or more elements. In one embodiment, the application 105 compares each element using a contour-based shape model. In another embodiment, the application 105 compares each element using a texture-based internal model.
[0086] In block 812, the application 105 characterizes the detected statistical anomaly using one or more domain-independent text phrases. In an embodiment, characterizing the detected statistical anomaly includes labeling missing or under-represented elements. In an embodiment, characterizing the detected statistical anomaly includes additionally labeling or over-represented elements.
[0087] In block 814, the application 105 converts the one or more domain-independent text phrases into one or more domain-specific descriptions based on the domain-specific categories. In block 816, the application 105 outputs the one or more domain-specific descriptions. Thereafter, the process 800 ends.
[0088] Thus, in exemplary embodiments, a computer-implemented method, system or apparatus, and computer program product for generating a textual description of an image using domain-independent anomaly analysis and other related features, functions or operations is provided. Where embodiments or portions thereof are described with respect to a type of device, the computer-implemented method, system or apparatus, computer program product, or portions thereof, is adapted or configured for use with an appropriate and comparable representation of that type of device.
[0089] Where embodiments are described as being implemented in an application, delivery of the application in a Software as a Service (SaaS) model is contemplated within the scope of the illustrative embodiments. In the SaaS model, the ability to implement the application of the embodiments is provided to the user by executing the application in a cloud infrastructure. The user can access the application using a variety of client devices through a thin client interface such as a web browser (e.g., web-based email) or other lightweight client applications. The user does not manage or control the underlying cloud infrastructure including the network, servers, operating system, or storage of the cloud infrastructure. In some cases, the user may not even manage or control the capabilities of the SaaS application. In some other cases, the SaaS implementation of the application may allow for limited possible exceptions to user-specific application configuration settings.
[0090] The present invention may be a system, method and / or computer program product of any possible integrated technical detail level. The computer program product may include a computer-readable storage medium (or medium) having computer-readable program instructions thereon, the computer-readable program instructions being used to cause a processor to perform aspects of the present invention.
[0091] A computer-readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of computer-readable storage media includes the following: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device (such as a punch card) or a raised structure in a groove with instructions recorded thereon), and any suitable combination of the foregoing. As used herein, a computer-readable storage medium should not be interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.
[0092] The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to a corresponding computing / processing device, or downloaded to an external computer or external storage device via a network (e.g., the Internet, a local area network, a wide area network, and / or a wireless network). The network may include copper transmission cables, optical transmission fibers, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions to be stored in a computer-readable storage medium within the corresponding computing / processing device.
[0093] The computer-readable program instructions for performing the operation of the present invention can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, configuration data of integrated circuits, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, and procedural programming languages, such as "C" programming language or similar programming languages. The computer-readable program instructions can be executed completely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer, partially on a remote computer, or completely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any type of network (including a local area network (LAN) or a wide area network (WAN)), or can be connected to an external computer (for example, by using the Internet of an Internet service provider). In some embodiments, an electronic circuit (including, for example, a programmable logic circuit, a field programmable gate array (FPGA) or a programmable logic array (PLA)) can execute a computer-readable program instruction to personalize the electronic circuit by utilizing the state information of the computer-readable program instructions, so as to perform aspects of the present invention.
[0094] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each box of the flowchart and / or block diagram and the combination of boxes in the flowchart and / or block diagram can be implemented by computer-readable program instructions.
[0095] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine that, when executed by a processor of a computer or other programmable data processing apparatus, creates a device for implementing the functions / actions specified in one or more blocks of a flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that can instruct a computer, a programmable data processing apparatus, and / or other device to function in a particular manner, such that a computer-readable storage medium having instructions stored therein includes an article of manufacture that includes instructions for implementing various aspects of the functions / actions specified in one or more blocks of a flowchart and / or block diagram.
[0096] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other apparatus that causes a series of operating steps to be performed on a computer, other programmable apparatus, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.
[0097] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to different embodiments of the present invention.To this end, each box in the flow chart or block diagram can represent a part of a module, segment or instruction, which includes one or more executable instructions for realizing the logical function of the specification.In some alternative embodiments, the function marked in the box may not occur in the order marked in the figure.For example, depending on the function involved, the two boxes shown in succession can actually be performed substantially at the same time, or these boxes can sometimes be performed in reverse order.It will also be noted that each box in the block diagram and / or flow chart and the combination of the boxes in the block diagram and / or flow chart can be realized by a system based on special-purpose hardware, and the system based on special-purpose hardware performs a specified function or action or performs a combination of special-purpose hardware and computer instructions.
Claims
1. A computer-implemented method for generating a text description of an image, include: classifying images represented by the image data into domain-specific categories, wherein the domain comprises a medical field; segmenting one or more elements in the image data based on the domain-specific categories; comparing each of the one or more elements to a domain-independent model to detect one or more statistical anomalies in the one or more elements, wherein the comparing comprises retrieving a statistical model from a database, and identifying a change in the image by examining at least a portion of the image data based on the statistical model retrieved from the database; characterizing the one or more detected statistical anomalies using one or more domain-independent text phrases; as well as The one or more domain-independent text phrases are converted into one or more domain-specific descriptions based on the domain-specific categories.
2. The method according to claim 1, further comprising: include: The one or more domain-specific descriptions are output.
3. The method according to claim 1, further comprising: include: receiving the image data; as well as Context information indicating a context of the image data is received, wherein classifying the image into the domain-specific category is based on the context information.
4. The method according to claim 3, in, The context information includes metadata associated with the image data.
5. The method according to claim 3, in, The context includes the modality and viewpoint of the image data.
6. The method according to claim 1, in, The one or more elements are segmented using a segmentation process that is specific to the domain-specific category.
7. The method according to claim 1, in, Segmenting the one or more elements uses a segmentation process specific to the elements.
8. The method according to claim 1, in, Segmenting the one or more elements aligns the elements of the generic segmentation with a coarse model associated with the domain-specific categories.
9. The method according to claim 1, in, The domain-independent model is a contour-based shape model.
10. The method according to claim 1, in, The domain-independent model is a texture-based internal model.
11. The method according to claim 1, in, Characterizing the detected statistical anomalies includes annotating missing or under-represented elements.
12. The method according to claim 1, in, Characterizing detected statistical anomalies includes over-annotated or over-represented elements.
13. A computer usable program product comprising one or more computer readable storage devices and program instructions stored on at least one of the one or more storage devices, wherein the stored program instructions include: program instructions for classifying an image represented by the image data into a domain specific category, wherein the domain comprises a medical field; program instructions for segmenting one or more elements in the image data based on the domain-specific categories; program instructions for comparing each of the one or more elements to a domain-independent model to detect one or more statistical anomalies in the one or more elements, wherein the comparing comprises retrieving the statistical model from a database, and identifying a change in the image by examining at least a portion of the image data based on the statistical model retrieved from the database; program instructions for characterizing the one or more detected statistical anomalies using one or more domain-independent text phrases; as well as Program instructions for converting the one or more domain-independent text phrases into one or more domain-specific descriptions based on the domain-specific categories.
14. The computer usable program product of claim 13, further comprising: include: Program instructions for outputting the one or more domain-specific descriptions.
15. The computer usable program product of claim 13, further comprising: include: program instructions for receiving said image data; as well as Program instructions for receiving context information indicating a context of the image data, wherein classifying the image into the domain-specific category is based on the context information.
16. The computer usable program product according to claim 15, in, The context information includes metadata associated with the image data.
17. The computer usable program product according to claim 15, in, The context includes the modality and viewpoint of the image data.
18. The computer usable program product of claim 13, wherein the computer usable code is stored in a computer readable storage device in a data processing system, and wherein the computer usable code is transmitted from a remote data processing system over a network.
19. The computer usable program product of claim 13, in, The computer usable code is stored in a computer readable storage device in a server data processing system and wherein the computer usable code is downloaded over a network to a remote data processing system for use in a computer readable storage device associated with the remote data processing system.
20. A computer system comprising one or more processors, one or more computer-readable memories, and one or more computer-readable storage devices, and program instructions stored on at least one of the one or more storage devices for execution by at least one of the one or more processors via at least one of the one or more memories, the stored program instructions include: program instructions for classifying an image represented by the image data into a domain specific category, wherein the domain comprises a medical field; program instructions for segmenting one or more elements in the image data based on the domain-specific categories; program instructions for comparing each of the one or more elements to a domain-independent model to detect one or more statistical anomalies in the one or more elements, wherein the comparing comprises retrieving the statistical model from a database, and identifying a change in the image by examining at least a portion of the image data based on the statistical model retrieved from the database; program instructions for characterizing the one or more detected statistical anomalies using one or more domain-independent text phrases; and Program instructions for converting the one or more domain-independent text phrases into one or more domain-specific descriptions based on the domain-specific categories.
