US20220164951A1

US20220164951A1 - Systems and methods for using ai to identify regions of interest in medical images

Info

Publication number: US20220164951A1
Application number: US17/531,177
Authority: US
Inventors: Haili Chui; Nikolaos Gkanatsios; Zhenxue Jing; Ashwini Kshirsagar; Samir Parikh; Venky Vaddineni
Original assignee: Hologic Inc
Current assignee: Hologic Inc
Priority date: 2020-11-20
Filing date: 2021-11-19
Publication date: 2022-05-26
Also published as: US20220164586A1; US12530860B2

Abstract

Examples of the present disclosure describe systems and methods for using AI to identify regions of interest (ROI) in medical images. In aspects, medical reports and images may be provided to a first environment. The first environment may use the medical report data/medical images to train a natural language processing (NLP)-based algorithm to identify the location in images of ROI described in the medical report data. The output of the NLP-based algorithm may be stored in an ROI repository in the first environment. After the NLP-based algorithm has been trained, a request to train a user-specific model may be received in a second environment. Data objects for the requested user-specific model may be provided to the first environment, which uses the ROI repository to train the model. The trained model may be provided to the second environment, where the trained user-specific model/algorithm may be tested and stored.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This application claims priority to and the benefit of U.S. Provisional Application No. 63/116,162, filed on Nov. 20, 2020, entitled “Systems and Methods for Using AI to Identify Regions of Interest in Medical Images,” the disclosure of which is hereby incorporated by reference in its entirety.

BACKGROUND

In the realm of artificial intelligence (AI), deep learning enables systems to automatically discover the information required to perform feature detection or classification using raw data. Deep learning requires access to large amounts of accurately labeled data. Typically, the data labeling is primarily a manual process, which can be prohibitively costly in terms of time and human/financial resources. Moreover, data privacy concerns are often a consideration.
It is with respect to these and other general considerations that the aspects disclosed herein have been made. Also, although relatively specific problems may be discussed, it should be understood that the examples should not be limited to solving the specific problems identified in the background or elsewhere in the present disclosure.

SUMMARY

Examples of the present disclosure describe systems and methods for using AI to identify regions of interest (ROI) in medical images. In aspects, medical report data and/or corresponding medical images may be provided to a first service or application in a first environment. The first service/application may use the medical report data/medical images to train a natural language processing (NLP)-based algorithm to identify within the medical images the location of findings described in the medical report data. The output of the NLP-based algorithm may be stored in an ROI repository in the first environment. After the NLP-based algorithm has been trained, a request to train a user-specific model or an algorithm may be received by a second service or application in a second environment. In response to the request, one or more data objects for the requested user-specific model/algorithm may be provided to the first service/application in the first environment. The first service/application may use data in the ROI repository to populate the data objects and train the user-specific model/algorithm. The trained user-specific model/algorithm may then be provided to the second service/application in the second environment, where the trained user-specific model/algorithm may be tested, stored, and/or provided to the user.
This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. Additional aspects, features, and/or advantages of examples will be set forth in part in the description which follows and, in part, will be apparent from the description, or may be learned by practice of the disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

Non-limiting and non-exhaustive examples are described with reference to the following figures.

FIG. 1 illustrates an overview of an example system for using AI to identify ROI in medical images, as described herein.

FIG. 2 is a diagram of an example process flow for using AI to identify ROI in medical images, as described herein.

FIG. 3 illustrates an example method for training an NLP-based model to generate medical images comprising labeled ROI, as described herein

FIG. 4 illustrates an example method for using AI to identify ROI in medical images, as described herein.

FIG. 5 illustrates one example of a suitable operating environment in which one or more of the present embodiments may be implemented.

DETAILED DESCRIPTION

Medical imaging has become a widely used tool for identifying and diagnosing abnormalities, such as cancers or other conditions, within the human body. Medical imaging processes such as mammography and tomosynthesis are particularly useful tools for imaging breasts to screen for, or diagnose, cancer or other lesions with the breasts. Tomosynthesis systems are mammography systems that allow high resolution breast imaging based on limited angle tomosynthesis. Tomosynthesis, generally, produces a plurality of X-ray images, each of discrete layers or slices of the breast, through the entire thickness thereof. In contrast to conventional two-dimensional (2D) mammography systems, a tomosynthesis system acquires a series of X-ray projection images, each projection image obtained at a different angular displacement as the X-ray source moves along a path, such as a circular arc, over the breast. In contrast to conventional computed tomography (CT), tomosynthesis is typically based on projection images obtained at limited angular displacements of the X-ray source around the breast. Tomosynthesis reduces or eliminates the problems caused by tissue overlap and structure noise present in 2D mammography imaging.
In recent times, artificial intelligence (AI) has been increasingly used to evaluate the image data generated using medical imaging. In particular, machine learning methods, such as deep learning, provide powerful tools for evaluating image data. Although such tools are highly accurate and efficient, these tools must be trained to perform specific tasks. The training requires access to a large amount of accurately labeled data. Generally, the data labeling process is performed manually. For example, a clinical professional must read medical report documents (physician notes, radiology reports, biopsy reports, etc.) to identify ROI associated with a patient. The clinical professional labels the identified ROI on medical images associated with the medical documents. Often, the quality of the labeling varies among clinical professionals based on various factors, such as experience, ability, fatigue, etc. The labeled medical images are provided as input to an AI component. Based on the input, the AI component is trained to identify the labeled ROI in medical images subsequently provided to the trained AI component. When the clinical professional intends to train the AI component to identify a new ROI or a new aspect of a ROI for which the AI component was previously trained, the entire process must be repeated. Thus, the data labeling process is often time-consuming, cumbersome, expensive, and potentially inaccurate.
In addition, the large amount of accurately labeled data includes patient records and other sensitive physical information that is protected by various laws and regulations including data security and protection and confidential handling of protected health information. As such, to comply with the laws and regulations, the data for purposes of labeling must first be deidentified by removing identification of a particular patient prior to the export of data from a medical facility. Such deidentification is time consuming and often done manually.
To address such issues with data labeling for AI training, the present disclosure describes systems and methods for using AI to identify ROI in medical images. In aspects, a first computing environment may comprise sensitive physical and/or electronic data, such as the medical report data, medical images, patient records, and other hospital information system (HIS) data. The first computing environment may correspond to a healthcare facility or a section or deport intent of a healthcare facility. At least a portion of the medical report data and/or medical images may be provided as input to a first service or application in the first computing environment. The first service or application may use the input to train an AI model or algorithm to identify ROI within the medical images based on the medical report data. In at least one example, the model or algorithm may use NLP techniques to identify language that describes the locations of findings in the medical report data. The model or algorithm may use the identified language to provide output including image overlays for the medical images or annotated versions of the medical images that include labeled locations of the findings identified by the identified language. The labeled locations may include textual labels, numerical values, highlighting, encircling (and/or other types of content enclosing), arrows or pointers, font or style modifications, etc. The output of the model or algorithm may be stored in at least one data repository in the first computing environment. The data repository may also store one or more portions of the medical report data and/or the patient records.
In aspects, a second computing environment may include a second service or application for training and storing user-requested models or algorithms. The second computing environment may be physically and/or logically separate from the first computing environment. In response to receiving a request to train a user-requested model or algorithm, the second service or application may provide data objects and/or training requirements for the requested user-specific model or algorithm to a training component in the first computing environment. The training component may search the data repository to identify information relevant to the requested user-specific model or algorithm. The training component may use the identified information to train the requested user-specific model or algorithm. The trained user-specific model or algorithm may be provided to the second service or application in the second computing environment without allowing the second computing environment access to the sensitive data in the first computing environment. Thus, the integrity and security of the sensitive data may be maintained throughout the training process. Upon receiving the trained user-specific model or algorithm, the second service or application may evaluate the model to determine a set of performance metrics. The set of performance metrics may represent the accuracy or effectiveness of the trained user-specific model or algorithm. In at least one aspect, the second service or application may use the set of performance metrics to iteratively tune/train the trained user-specific model or algorithm.
Accordingly, the present disclosure provides a plurality of technical benefits including but not limited to: training an NLP-based model to detect text relating to ROI locations, using NLP-based model output to train specific AI models, improving data security/privacy during model creation, improving the accuracy of labeled data, improving the efficiency of generating labeled data, enabling self-learning AI systems within client or sensitive environments.
FIG. 1 illustrates an overview of an example system for using AI to identify regions of interest (ROI) in medical images as described herein. Example system 100 as presented is a combination of interdependent components that interact to form an integrated system. Components of system 100 may be hardware components or software components implemented on and/or executed by hardware components of the system. System 100 may provide one or more operating environments for software components to execute according to operating constraints, resources, and facilities of system 100. In one example, the operating environment(s) and/or software components may be provided by a single processing device, as depicted in FIG. 6. In another example, the operating environment(s) and software components may be distributed across multiple devices. For instance, input may be entered on a user device and information may be processed or accessed using other devices in a network, such as one or more network devices and/or server devices.
As one example, system 100 comprises environments 101 and 121 and network 110. One of skill in the art will appreciate that the scale of systems such as system 100 may vary and may include more or fewer environments and/or components than those described in FIG. 1. For instance, in some examples, at least a portion of the functionality and components of environments 101 and 121 may be integrated into a single environment, processing system, or device. Alternately, the functionality and components of environments 101 and/or 121 may be distributed across multiple environments or processing systems.
Environment 101 may comprise user devices 102A, 102B, and 102C (collectively “user devices 102”), server device 104, and data store(s) 106. In at least one aspect, environment 101 may represent a cloud-based or distributed computing environment. User devices 102 may be configured to receive or collect input from one or more users or alternate devices. Examples of user devices 102 include, but are not limited to, personal computers (PCs), server devices, mobile devices (e.g., smartphones, tablets, laptops, personal digital assistants (PDAs)), and wearable devices (e.g., smart watches, smart eyewear, fitness trackers, smart clothing, body-mounted devices). User devices 102 may include sensors, applications, and/or services for receiving or collecting input. Example sensor include microphones, touch-based sensors, keyboards, pointing/selection tools, optical/magnetic scanners, accelerometers, magnetometers, gyroscopes, etc. The collected input may include, for example, voice input, touch input, text-based input, gesture input, video input, and/or image input.
Server device 104 may be configured to receive collected input from user devices 102. Examples of server device 104 include, but are not limited to, application servers, web servers, file servers, database servers, and mail servers. Upon receiving collected input, server device 104 may provide access to data and one or more services/applications. The data and services/applications may be stored remotely from server device 104 and accessed by server device 104 via network 110. Alternately, the data and services/applications may be stored and accessed locally on server device 104 using a data store, such as data store(s) 106. Examples of data store(s) 106 include, but are not limited to, databases, file systems, directories, flat files, and email storage systems. In some aspects, data store(s) 106 may comprise data objects and/or sets of instructions for one or more algorithms and/or models. A model, as used herein, may refer to a predictive or statistical utility or program that may be used to determine a probability distribution over one or more character sequences, classes, objects, result sets or events, and/or to predict a response value from one or more predictors. A model may be based on, or incorporate, one or more rule sets, machine learning (ML), a neural network, or the like. In at least one aspect, the algorithms and/or models may be proprietary and/or subject to trade secret protections by the owners of the algorithms and/or models.
The algorithms and/or models in data store(s) 106 may be used to perform one or more specific tasks, such as identifying a type of cancer, a category of disease, image anomalies, etc. Although reference to specific types of tasks are described herein, it is contemplated that the algorithms and/or models may be used to perform alternate types of tasks and used in alternate types of environments. In response to receiving the collected input, server device 104 may collect or receive one or more data objects and/or sets of instructions relating to a specific task or set of tasks from data store(s) 106. Server device 104 may identify a task and/or corresponding data objects/instructions based on one or more terms in or associated with the collected input. For example, server device 104 may parse the collected input to identify query terms or input terms. The identified terms may be used to search the data (e.g., algorithm names, data object text, instruction text) in data store(s) 106 for similar or matching terms using search techniques, such as pattern matching, regular expressions, fuzzy matching, etc. When one or more matches are identified, the corresponding algorithm(s)/model(s) may be selected and server device 104 may collect or receive one or more data objects and/or sets of instructions relating to the selected algorithm(s)/model(s). Server device 104 may provide one or more data objects and/or sets of instructions to environment 121 based on the collected input.
Server device 104 may be further configured to evaluate response data received from environment 121. The response data may be provided by environment 121 in response to one or more data objects and/or sets of instructions provided to environment 121. In aspects, server device 104 may comprise or provide access to an execution environment (not pictured). The execution environment may comprise or utilize functionality for evaluating the response data. In at least one aspect, the response data corresponds to a trained user-requested model or algorithm. The evaluated response data may be stored in one or more data stores, such as data store(s) 106. The response data may be provided to a user in response to receiving the collected input.
Environment 121 may comprise server device 124, data store(s) 126, and feature store(s) 128. In at least one aspect, environment 121 may represent a computing environment comprising sensitive data, such as a healthcare computing environment comprising patient data. Server device 124 may be configured to collect data from the one or more data sources, such as data store(s) 126 and/or feature store(s) 128. Examples of data store(s) 126 and feature store(s) 128 include, but are not limited to, databases, file systems, directories, flat files, and email storage systems. In at least one aspect, the collected data may correspond to medical report data, medical images, patient records, and/or other sensitive medically related information. The collected data may be used to train an NLP-based algorithm or model (not pictured). At least a portion of the output of the trained NLP-based algorithm or model and/or the collected data may be stored in feature store(s) 128.
Server device 124 may be further configured to receive one or more data objects and/or sets of instructions from environment 101. Server device 124 may identify a specific task associated with the received one or more data objects and/or sets of instructions. The identified specific task may be used to search feature store(s) 128 for stored data relevant to performing the specific task. In at least one aspect, the stored data may correspond to labeled or annotated image data, text terms or phrases from medical report data, and/or feature data associated with image data or medical report data. Stored data identified to be relevant may be provided to a training component (not pictured) within environment 121. The training component may be a hardware device, a software component within server device 124, or a software component within a separate hardware device of environment 121. In examples, the training component may be implemented as a black box that provides separation between environment 101 and environment 121. The separation may prevent environment 101 (and other environments external to environment 121) from accessing the sensitive data of environment 121 from outside of environment 121. The separation may also prevent environment 121 from unauthorized access of the models and/or algorithms stored in data store(s) 106. For instance, as the models and/or algorithms may be proprietary to owners who are third parties with respect to environment 101, it may be desirable for the owners to keep the algorithms secure from users in environment 101.
The training component may be configured to train a user-requested model or algorithm. In examples, the stored data identified to be relevant may be provided to the training component. The training component may use the relevant stored data to train a user-requested model or algorithm that is operable to perform the identified specific task. The trained user-requested model or algorithm may then be provided as response data to environment 101. In aspects, the user-requested model or algorithm may be trained and provided to environment 101 such that sensitive data in environment 121 is not exposed to environment 101. As such, the patient data that is located in any sensitive medically related information used to train the trained user-requested model or algorithm does not need to be de-identified because it is not removed from the environment 101 and stays on site in the environment 101. This results in saving significant time from gathering, processing, exporting and storing information, which previously may have been done manually by a highly skilled medical technician.
FIG. 2 is a diagram of an example process flow for using AI to identify regions of interest (ROI) in medical images, as described herein. Process flow 200, as presented, comprises environments 201 and 221. In examples, environment 201 may represent a healthcare facility, such as a hospital, an imaging and radiology center, an urgent care facility, a medical clinic or medical offices, an outpatient surgical facility, a physical rehabilitation center, etc. Environment 201 may comprise sensitive or private information associated with a healthcare facility, healthcare patients, and/or healthcare personnel. Environment 221 may represent a web-based, cloud-based, or distributed computing environment associated with environment 201. Environment 221 may be publicly or selectively accessible and may implement security procedures to enable the secure access of environment 201. Generally, environment 221 may not store or have access to sensitive or private information comprised by environment 201. Environments 201 and 221 may be physically and/or logically separated. In addition, the environment 201 and 221 may be separated by firewalls and authentication protocols ensuring safe handling of the sensitive medical information comprised in the environment 201.
Environment 201 may comprise ROI analysis engine 202, medical data 204A, 204B, and 204C (collectively “medical data 204”), ROI repository 206, orchestration engine 214, and training engine 216. Environment 221 may comprise user(s) 208, application 210, algorithm repository 212, and model repository 218. One of skill in the art will appreciate that the number and type of environments and/or components associated with environment 201, environment 221, and process flow 200 may vary from those described in FIG. 2.
In aspects, ROI analysis engine 202 may be provided, or may have access to, medical data associated with one or more patients, such as medical data 204. ROI analysis engine 202 may be configured to identify ROI associated with medical data 204. Examples of medical data 204 include, but are not limited to, medical report data 204A (e.g., radiology reports, biopsy reports, audio reports, healthcare professional notes and documents), medical image data 204B (e.g., X-ray images, CT images, MRI images, ultrasound images), and electronic medical record (EMR) data 204C (e.g., patient records, medical and treatment history information, patient health data). Although specific references to medical data and procedures are described, it is contemplated that the systems and methods described herein may be implemented with alternate types of data, procedures, and environments.
Upon receiving the medical data, ROI analysis engine 202 may use medical data 204 to train an AI model/algorithm (not pictured) within environment 201. The AI model/algorithm may be stored by ROI analysis engine 202 or elsewhere within environment 201. The AI model/algorithm may be configured to identify ROI within the medical image data based on corresponding medical report data. For example, the AI model/algorithm may implement NLP techniques to identify text and/or speech in medical report data that describes the locations of one or more findings within the patient. The AI model/algorithm may use the identified text and/or speech to identify the findings in corresponding medical image data. The AI model/algorithm may label the identified finding within the medical image data by generating image overlays or annotated versions of the medical image data. The medical image data labeled by the AI model/algorithm, image feature data relating to the medical image data, and the corresponding identified text and/or speech may be stored in a data store, such as ROI repository 206.
After the AI model/algorithm had been trained, a user in or interfacing with environment 221, such as user(s) 208, may access application 210. Examples of user(s) 208 may include one or more manufacturers of algorithms designed to detect different types of medical conditions or abnormalities, such as cancers which may be diagnosed by healthcare professionals from medical images. Application 210 may be configured to receive, store, and/or process user requests to train a user-specific algorithm to perform a specific task. Upon receiving a request from user(s) 208 to train a new user-specific algorithm, application 210 may access algorithm repository 212. Algorithm repository 212 may be configured to store and provide various algorithms relating to environment 201. The algorithms of algorithm repository 212 may relate to various topics, concepts, or areas. For example, a first algorithm may be used to identify a first type of cancer, a second algorithm may be used to identify a second type of cancer, and a third algorithm may be used to identify images having poor image quality. Algorithm repository 212 may be configured to store and provide data objects and/or instructions for training the stored algorithms. Algorithms in the algorithm repository 212 may be proprietary and subject to trade secret protections. It may be desirable for the owners of the algorithms to keep the algorithms secure. As discussed above, environments 221 and 201 may be physically and logically separated and protected by firewalls and other security measures. By separating the environments 221 and 201 access to the algorithms is secured and can be managed by the owners as they reside in the environments subject to the owners' control.
Application 210 may use terms and keywords in the request from user(s) 208 to identify a context (e.g., a topic, a concept, or an area) associated with the request. Application 210 may use the identified context to search algorithm repository 212 for relevant algorithms. When a relevant algorithm is identified in algorithm repository 212, the identified algorithm, one or more data objects, and/or instructions for training the identified algorithm may be provided to orchestration engine 214. In some examples, orchestration engine 214 may be configured to monitor environment 221 and/or application 210 to detect when a user request to train a user-specific algorithm is received by application 210. The monitoring may include the implementation of monitoring services or software used to transmit periodic queries to application 210, receive notifications from application 210, intercept messages between users(s) 208 and application 210, etc. When a user request to train a user-specific algorithm is detected, orchestration engine 214 may cause algorithm repository 212 to provide the identified algorithm, one or more data objects, and/or instructions for training the identified algorithm to orchestration engine 214 and/or training engine 216. For example, orchestration engine 214 may request the access path and/or credentials for algorithm repository 212. Orchestration engine 214 may use the access path and/or credentials to retrieve the identified algorithm, data objects, and/or instructions. Alternately, orchestration engine 214 may provide the access path and/or credentials to training engine 216 and training engine 216 may use the access path and/or credentials to retrieve the identified algorithm, data objects, and/or instructions.
Orchestration engine 214 and/or training engine 216 may also be configured to initiate the training of the identified algorithm within environment 201. Orchestration engine 214 may provide the identified algorithm, one or more data objects, and/or instructions for training the identified algorithm to training engine 216. Alternately or additionally, orchestration engine 214 may provide a command (including parameters) for initiating the training of the identified algorithm to the training engine 216. Training engine 216 may be configured to search ROI repository 206 for data (e.g., medical image data, image feature data, identified text and/or speech) associated with the identified algorithm, and to train a model based on the data. In aspects, training engine 216 may be implemented in a manner that provides separation between environment 201 and environment 221. For example, training engine 216 may prevent users and devices in environment 221 (and other environments external to environment 201) from accessing the sensitive or secure data of environment 201, such as medical data 204, from outside of environment 201. Further, training engine 216 may prevent users and devices in environment 201 (and other environments external to environment 221) from directly accessing the algorithms stored in algorithm repository 212. For instance, training engine 216 may implement security features or policies that prevent users and devices in environment 201 and environment 221 from viewing or accessing the data (e.g., ROI repository 206 data or algorithm repository 212) received by training engine 216.
Upon receiving the identified algorithm, one or more data objects, instructions for training the identified algorithm, and/or command (including parameters) for initiating the training of the identified algorithm, training engine 216 may train a model based on the identified algorithm. When the model has been trained, orchestration engine 214 or training engine 216 may provide the trained model to model repository 218. Alternately, orchestration engine 214 or training engine 216 may provide the trained model to application 210 and application 210 may provide the trained model to model repository 218. Model repository 218 may be configured to store various trained models and associated data, such as creation/modification data, a description of the model, testing data, result accuracy data, keywords or terms associated with the model, version/iteration number, etc.
In aspects, after the trained model has been provided to model repository 218 and/or application 210, user(s) 208 may interact with the trained model using application 210. For example, application 210 may also be configured to provide a testing environment (not pictured) to test the trained model. The testing environment may implement tools for evaluating the performance metrics for the trained model. In examples, the performance metrics may relate to receiver operating characteristics (ROCs) and/or free-response receiver operating characteristics (FROCs), such as sensitivity, specificity, precision, hit rate, accuracy, etc. Evaluating the performance metrics for the trained model may include using the trained model to perform a specific task intended by the user and/or comparing the performance metrics for the trained model to a set of baseline performance metrics. For example, the trained model may be used to identify image data or aspects thereof. Based on the performance metrics for the trained model, the trained model may be provided to training engine 216, as described above, to be refined/retrained. A set of training parameters may for refining/retraining may also be provided to training engine 216. Training engine 216 may refine/retrain the trained model based on the set of training parameters. The refined/retrained model may be provided to application 210 and/or model repository 218. The testing environment of application 210 may be used to evaluate the performance metrics of the refined/retrained model. In some aspects, the performance metrics of trained model and the refined/retrained model may be compared to determine whether the trained model or the refined/retrained model is more accurate. Based on the comparison, the trained model and/or the refined/retrained model may be stored or removed from the model repository 218. Additionally, the refined/retrained model may be further refined/retrained using the process described above.
Having described a system and process flow that may employ the techniques disclosed herein, the present disclosure will now describe one or more methods that may be performed by various aspects of the present disclosure. In aspects, methods 300 and 400 may be executed by a system, such as system 100 of FIG. 1. However, methods 300 and 400 are not limited to such examples. In other aspects, methods 300 and 400 may be performed by a single device comprising multiple computing environments. In at least one aspect, methods 300 and 400 may be executed (e.g., computer-implemented operations) by one or more components of a distributed network, such as a web service/distributed network service (e.g. cloud service).
FIG. 3 illustrates an example method for training an NLP-based model to generate images comprising labeled ROI, as described herein. Example method 300 begins at operation 302, where medical data may be received. In aspects, an analysis component, such as ROI analysis engine 202, may be located in a secure environment, such as environment 201. The secure environment may correspond to a client environment of a healthcare facility or of another location comprising sensitive data. The analysis component may receive or have access to medical data from one or more sources, such as data store(s) 106. The medical data may include medical report data, medical image data, EMR data, and other HIS data.
At operation 304, text describing the location of ROI may be identified. In aspects, the analysis component may apply one or more NLP techniques to the medical data. Example NLP techniques include, but are not limited to, named entity recognition, sentiment analysis, tokenization, sentence segmentation, and stemming and lemmatization. The NLP techniques may be used to identify significant terms and/or phrases in text data of the medical data. The significant terms and/or phrases may correspond to terms and/or phrases of a standardized (or semi-standardized) lexicon used for reporting the outcomes of image review. As one example, the NLP techniques may be applied to medical report data (e.g., radiology reports and/or biopsy reports) to identify text describing one or more findings or ROI (e.g., lesions, asymmetric breast tissue, macrocalcifications, asymmetry density, distortion mass, or adenopathy) resulting from a mammographic exam. The text may include features of the findings or ROI, such as size, location, texture, density, symmetry, etc. As a specific example, the NLP techniques may identify a sentence in a radiology report that indicates a lesion was detected in the superior medial portion of a patient's left breast. The NLP techniques may also identify another sentence in the radiology report that indicates the size and density of the lesion and the approximate location of the lesion with in the superior medial quadrant. The text associated with each sentence may be extracted by the analysis component. The extracted text may be labeled (e.g., superior medial lesion) and stored with text relating to similar findings. For instance, all text describing findings or ROI in the superior medial quadrant of a breast may be stored under the category “Superior Medial Findings.”
At operation 306, an NLP-based model may be trained. In aspects, the significant terms and/or phrases identified in the text data of the medical data (and in other medical data) may be provided as input to an NLP-based model located within the secure environment. The NLP-based model may be generated and/or maintained by the analysis component or by another component within the secure environment. Image data corresponding to the identified significant terms and/or phrases may also be provided as input to an NLP-based model. The input may be used to train the NLP-based model to match the identified significant terms and/or phrases to corresponding locations of ROI in the image data. Matching the identified significant terms and/or phrases to the corresponding locations may include generating labeled image data comprising labels and/or annotations of the ROI. For example, various text strings from a radiology report and one or more corresponding tomosynthesis computer-aided design (CAD) images may be provided to an NLP-based model. In response to the text string “a lesion was detected in the superior medial portion of a patient's left breast,” the NLP-based model may evaluate the CAD image(s) to identify images of the patient's left breast. For each identified CAD image of the patient's left breast, the NLP-based model may evaluate the superior medial quadrant of the breast in the CAD image to identify ROI corresponding to the text string. The evaluation may include the use of unified vectors of location features. For each identified ROI, the NLP-based model may label the ROI on the CAD image and/or create a labeled version of the CAD image. For instance, the NLP-based model may generate an overlay in which the identified ROI is encircled or otherwise highlighted.
At operation 308, the NLP-based model data may be stored in a repository. In aspects, content generated or output by the NLP-based model may be stored in a data repository, such as ROI repository 206. The content may include, labeled or otherwise annotated image data, unlabeled/unannotated image data, feature vectors, terms and/or phrases, and/or medical data available to the analysis component. The data repository may also be located in the secure environment such that the NLP-based model may be trained, and the content generated therefrom may be stored without exposing sensitive data to entities outside of the secure environment.
FIG. 4 illustrates an example method for using AI to identify ROI in medical images as described herein. In aspects, an application or service providing component (“application component”), such as application 210 may be located in a first computing environment, such as environment 221. The first computing environment may correspond to a cloud-based or web-based environment that may be publicly or selectively accessible. The first computing environment may provide access to a second computing environment, such as environment 201. The second computing environment may correspond to a secure, client environment of a healthcare facility or of another location comprising private or sensitive data.
Example method 400 begins at operation 402, where a request to train a user-selected algorithm is detected. In aspects, a user in or accessing the first computing environment may access a user interface provided by the application component. The user interface may provide the user with an option to identify an algorithm to be trained to perform one or more tasks. Identifying the algorithm may comprise selecting an algorithm from a list of algorithms in an algorithm store, such as algorithm repository 212. Alternately, identifying an algorithm may comprise providing one or more algorithm characteristics (e.g., intended function or type/category) to the user interface. In response to a user identifying an algorithm, the application component may provide the identified algorithm, one or more data objects associated with the identified algorithm, and/or instructions for training the identified algorithm (collectively, “algorithm container”).
At operation 404, the algorithm container may be provided to the second computing environment. In some aspects, the application component may send the algorithm container to one or more components in the second computing environment in response to receiving the user request. For example, the application component may send the algorithm container to an algorithm training orchestration component, such as orchestration engine 214, or to an algorithm training component, such as training engine 216. In other aspects, the algorithm training orchestration component of the second computing environment may monitor the application component in the first computing environment. Upon detecting a request to train a user-selected algorithm has been received by the application component, the orchestration component or the algorithm training component may request the algorithm container from the application component or the algorithm store. In response to the request by the orchestration component, the application component may provide the algorithm container to the orchestration component or the algorithm training component. Alternately, the application component may provide information for the algorithm container (e.g., identifier, location/path, access credentials) to the orchestration component or algorithm training component. The orchestration component or algorithm training component may use the information for the algorithm container to retrieve the algorithm container.
At operation 406, the algorithm container may be used to identify content related to the user-selected algorithm. In aspects, one or more identifiers (e.g., terms, phrases, topics, contexts) associated with the received algorithm container may be identified. The identifiers may be used to search a data repository, such as ROI repository 206, for content related (e.g., relevant) to the algorithm container. The content in the data repository may include, for example, labeled or otherwise annotated image data, unlabeled/unannotated image data, ROI feature vectors, terms and/or phrases describing ROI, and/or other medical data available in the second computing environment. Searching the data repository may include using pattern matching techniques, such as regular expressions, fuzzy logic, pattern recognition models, etc. Any content determined to be related to the algorithm container may be identified and extracted from the data repository. As on specific example, an algorithm container for detecting metastatic breast cancer may be titled as or comprise the term “Metastatic.” Based on identifying the term “metastatic” in/for the algorithm container, image data comprising ROI that include instances of metastatic breast cancer may be identified.
At operation 408, content related to the algorithm container may be used to train a model. In aspects, the algorithm container and/or the content related to the algorithm container may be provided as input to the training component in the second computing environment. The training component may use the input to train a model corresponding to the algorithm container. For example, the training component may use overlay image data in the related content to populate or otherwise configure one or more data objects in the algorithm container according to a set of instructions and/or parameters in the algorithm container. The populated/configured data objects may be used to construct a model representing the algorithm the user requested to be trained. In examples, the model may be trained such that data used to train the model in the second computing environment is not exposed to the first computing environment.
At operation 410, the trained model may be provided to the first computing environment. In aspects, the orchestration component may receive or collect the trained model from the training component. The orchestration component may provide the trained model to the first computing environment. For example, the orchestration component may provide the trained model to the application component and/or to a model store of the first computing environment, such as model repository 218. Alternately, the training component may provide the trained model to the first computing environment. The trained model may be stored in the model store and/or presented to the user using the user interface. The user interface may enable the user to execute, modify, or otherwise interact with the trained model.
At operation 412, the trained model may be evaluated. In aspects, the first computing environment or a component thereof, such as the application component, may comprise a test operating environment. The test operating environment may provide one or more tools for evaluating the trained model. The evaluation may include identifying performance metrics for the trained model and/or comparing the identified performance metrics to a set of baseline or default performance metrics. In some aspects, the test operating environment may enable the iterative training of a model. For example, after evaluating a trained model in the test operating environment, an updated algorithm container may be manually or automatically selected from the algorithm store or may otherwise be acquired. The updated algorithm container may be selected by, for example, the application component based on predefined testing constraints or according to a test script or executable test file for the selected algorithm or algorithm type. The trained model and the updated algorithm container may be provided to the training component in the second computing environment. The updated algorithm container may comprise an updated set of instructions and/or parameters for training the trained model. Based on the updated algorithm container, the training component may update/(re)train the trained model. The updated trained model may be provided to the first computing environment. The test operating environment may be used to evaluate performance metrics for the updated trained model. The performance metrics for the trained model and the performance metrics for the updated trained model may then be compared to determine the which model (e.g., trained model or updated trained model) is more accurate. Based on the comparison, the most accurate model may be selected, and a newly updated algorithm container may be selected or obtained. The process may continue as described above until a set of performance metrics meeting or exceeding a threshold value/level is acquired, or until a defined set of criteria is met.
FIG. 5 illustrates an exemplary suitable operating environment for the automating clinical workflow decision techniques described in FIG. 1. In its most basic configuration, operating environment 500 typically includes at least one processing unit 502 and memory 504. Depending on the exact configuration and type of computing device, memory 504 (storing, instructions to perform the techniques disclosed herein) may be volatile (such as RAM), non-volatile (such as ROM, flash memory, etc.), or some combination of the two. This most basic configuration is illustrated in FIG. 5 by dashed line 506. Further, environment 500 may also include storage devices (removable, 508, and/or non-removable, 510) including, but not limited to, magnetic or optical disks or tape. Similarly, environment 500 may also have input device(s) 514 such as keyboard, mouse, pen, voice input, etc. and/or output device(s) 516 such as a display, speakers, printer, etc. Also included in the environment may be one or more communication connections 512, such as LAN, WAN, point to point, etc. In embodiments, the connections may be operable to facility point-to-point communications, connection-oriented communications, connectionless communications, etc.
Operating environment 500 typically includes at least some form of computer readable media. Computer readable media can be any available media that can be accessed by processing unit 502 or other devices comprising the operating environment. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transitory medium which can be used to store the desired information. Computer storage media does not include communication media.
Communication media embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, microwave, and other wireless media. Combinations of the any of the above should also be included within the scope of computer readable media.
The operating environment 500 may be a single computer operating in a networked environment using logical connections to one or more remote computers. The remote computer may be a personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above as well as others not so mentioned. The logical connections may include any method supported by available communications media. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.
The embodiments described herein may be employed using software, hardware, or a combination of software and hardware to implement and perform the systems and methods disclosed herein. Although specific devices have been recited throughout the disclosure as performing specific functions, one of skill in the art will appreciate that these devices are provided for illustrative purposes, and other devices may be employed to perform the functionality disclosed herein without departing from the scope of the disclosure.
This disclosure describes some embodiments of the present technology with reference to the accompanying drawings, in which only some of the possible embodiments were shown. Other aspects may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments were provided so that this disclosure was thorough and complete and fully conveyed the scope of the possible embodiments to those skilled in the art.
Although specific embodiments are described herein, the scope of the technology is not limited to those specific embodiments. One skilled in the art will recognize other embodiments or improvements that are within the scope and spirit of the present technology. Therefore, the specific structure, acts, or media are disclosed only as illustrative embodiments. The scope of the technology is defined by the following claims and any equivalents therein.

Claims

What is claimed is:

1. A method for using AI to identify ROI in medical images, the method comprising:

detecting, by an orchestration engine in a first environment, a request to train an algorithm to identify the ROI, wherein the request is provided to a second environment external to the first environment;

receiving an algorithm container for the algorithm, wherein the algorithm container comprises one or more data objects for the algorithm;

identifying, in a ROI repository in the first environment, content related to the algorithm container, wherein the content comprises one or more images of the ROI;

training, by a training engine in the first environment, a model using the content and

providing the trained model to the second environment.

2. A system for using AI to identify ROI in medical images, the system comprising:

a processor; and

memory coupled to the processor, the memory comprising computer executable instructions that, when executed, perform a method comprising:

providing the trained model to the second environment.

3. A system for using AI to identify ROI in medical images, the system comprising:

a processor; and

receiving, by an application in a first environment, a request to train an algorithm to identify the ROI;

providing, to a second environment, an algorithm container for the algorithm, wherein the algorithm container comprises one or more data objects for the algorithm;

receiving a trained model from the second environment, wherein the trained model is based on the algorithm container; and

evaluating performance metrics of the trained model using a test operating environment of the second environment.