Autonomous medical imaging system with ai agent integration

Abstract

A computerized medical image management system integrates artificial intelligence (AI) agents into clinical workflows. The system includes a dynamic agent selector that autonomously selects and manages AI agents based on real-time user interactions and contextual parameters. The system may include a learning module that autonomously learns user preferences and improves customization over time.

Description

SNTH.013WO PATENTAUTONOMOUS MEDICAL IMAGING SYSTEM WITH Al AGENT INTEGRATIONBACKGROUND

[0001] Customized workflows in clinical medical image management, reporting, and workflow systems have heretofore required conventional programs. For example, Reicher et al received U.S. Patent No 5,452,416 in 1995 for providing user preferences for display and output in a PACs. The Reicher patent is hereby incorporated by reference herein for all purposes. The problems with such programs is that users must select customized options from complex menus. User training is difficult, and configuration dialogs are easily misunderstood. In addition, systems do not autonomously learn a user’s preferences, so that the customization does not improve over time in response to passive user actions. Furthermore, there is no learning from one user to another. For example, historical systems could not automatically suggest or apply the preferences of an efficient reading physician to a less efficient physician. Another example: A reading physician may rearrange the order or position of the displayed medical images series to optimize comparison and may set rules for which images series are displayed where, but once these rules are set, they must be associated with specific medical imaging exam criteria so that the proper rules act on the appropriate exams.SUMMARY

[0002] The systems, methods, and devices described herein each have several aspects, no single one of which is solely responsible for its desirable attributes. Without limiting the scope of this disclosure, several non-limiting features will now be described briefly.

[0003] The systems and methods discussed herein may be implemented in a computerized medical image management system. These systems and methods integrate artificial intelligence (“Al”) agents into clinical workflows. For example, a dynamic agent selector may autonomously select and manage Al agents (that each perform specific tasks) based on real-time user interactions and contextual parameters, continuously updating its internal model to adapt to changing conditions. The system may include a learning module that autonomously learns user preferences and improvesSNTH.013WO PATENT customization over time, such as based on passive user actions. Additionally, the system may include a constraint module that allows Al agents to be constrained based on reference material, improving accuracy and relevance.

[0004] The system may automatically improve or update rules or exam associations in response to user actions, benefiting not only reading physicians but also technologists, referring doctors, and patients. Furthermore, the system may address the issue of Al services producing false or erroneous information (hallucinations) by automatically or semi-automatically restricting the inputs to the Al service, thereby mitigating this problem.

[0005] The following description discusses various processes and components that may perform artificial intelligence (“Al”) processing or functionality. Al generally refers to the field of creating computer systems that can perform tasks that typically require human intelligence. This includes understanding natural language, recognizing objects in images, making decisions, and solving complex problems. Al systems can be built using various techniques, like neural networks, rule-based systems, or decision trees, for example. Neural networks learn from vast amounts of data and can improve their performance over time. Neural networks may be particularly effective in tasks that involve pattern recognition, such as image recognition, speech recognition, or Natural Language Processing.

[0006] Al processing may include training, fine-tuning, and inference. Training may use optimization algorithms such as stochastic gradient methods and may involve distributed strategies like data parallelism, model parallelism, or pipeline parallelism. Hardware platforms can include GPUs, TPUs, or specialized accelerators with high- bandwidth memory and interconnects. Data pipelines may include tokenization, deduplication, and sharding. Inference can be optimized using techniques like quantization, caching, and speculative decoding. Non-neural approaches may include probabilistic models or gradient-boosted decision trees. For vision tasks, architectures may include convolutional neural networks or vision transformers; for speech, recurrent or conformer models; and for time-series, temporal convolution or transformer-based models. Evaluation may use metrics such as accuracy, F1 score, BLEU, or word error rate, and deployment can occur on-device, at the edge, or in the cloud, with options for federated learning or secure enclaves.SNTH.013WO PATENT

[0007] Natural Language Processing (NLP) is an area of artificial intelligence (Al) that focuses on teaching computers to understand, interpret, and generate human language. By combining techniques from computer science, machine learning, and / or linguistics, NLP allows for more intuitive and user-friendly communication with computers. NLP pipelines may include text normalization, tokenization, part-of-speech tagging, parsing, semantic role labeling, and coreference resolution. Sentiment analysis may classify text into emotional categories, while machine translation may use encoderdecoder architectures with attention mechanisms. Named-entity recognition may apply tagging schemes like BIO or IOB, and summarization can be extractive or abstractive. Speech recognition systems may process audio into spectrograms and use models such as CTC or attention-based architectures, with decoding strategies like beam search. Question answering can be extractive, generative, or retrieval-augmented, often leveraging vector search for context retrieval. Natural Language Understanding (NLU) focuses on comprehension, including idioms, metaphors, and pragmatic context, and may integrate knowledge graphs or retrieval systems. NLP systems can be multilingual or cross-lingual and often include privacy measures such as Pll detection and differential privacy.

[0008] A Language Model is any algorithm, rule, model, and / or other programmatic instructions that can predict the probability of a sequence of words. A language model may, given a starting text string (e.g., one or more words), predict the next word in the sequence. A language model may calculate the probability of different word combinations based on the patterns learned during training (based on a set of text data from books, articles, websites, audio files, etc.). A language model may generate many combinations of one or more next words (and / or sentences) that are coherent and contextually relevant. Thus, a language model can be an advanced artificial intelligence algorithm that has been trained to understand, generate, and manipulate language. A language model can be useful for natural language processing, including receiving natural language prompts and providing natural language responses based on the text on which the model is trained. A language model may include an n-gram, exponential, positional, neural network, and / or other type of model. Neural language models may use architectures such as recurrent networks, convolutional networks, or transformers, withSNTH.013WO PATENT positional encoding to capture word order. Decoding strategies may include greedy search, beam search, top-k sampling, or nucleus sampling, and systems may apply penalties or constraints to improve diversity and relevance. Language models can also support structured outputs, function calling, and streaming generation for real-time applications.

[0009] A Large Language Model (“LLM”) is any type of language model that has been trained on a larger data set and has a larger number of training parameters compared to a regular language model. An LLM can understand more intricate patterns and generate text that is more coherent and contextually relevant due to its extensive training. Thus, an LLM may perform well on a wide range of topics and tasks. An LLM may comprise a NN trained using self-supervised learning. An LLM may be of any type, including a Question Answer (“QA”) LLM that may be optimized for generating answers from a context, a multimodel LLM / model, and / or the like. An LLM (and / or other models of the present disclosure), may include, for example, attention-based and / or transformer architecture or functionality. LLMs can be extremely useful for natural language processing, including receiving natural language prompts and providing natural language responses based on the text on which the model is trained.

[0010] LLMs may be based on transformer architectures that use multi-head attention, feed-forward layers, normalization, and / or residual connections. Efficiency improvements may include sparse attention, memory optimization, and quantization. Training may involve self-supervised learning on diverse collections, followed by fine- tuning and alignment techniques such as supervised instruction tuning or reinforcement learning from human feedback. LLMs can be multimodal, processing text alongside images, audio, or other data types. They may also integrate retrieval-augmented generation for grounding responses in external knowledge and support tool use for executing functions or accessing APIs. Deployment strategies may include model parallelism, caching, and hardware acceleration to meet latency and cost requirements.

[0011] As used herein, references to specific uses and / or implementations of Al, NLP, NLU, or LLM should be interpreted to include any other implementations, including any of those discussed above. For example, references to NLP herein should be interpreted to include NLU also. Thus, these terms may encompass training, fine- tuning, and inference phases; monolingual, multilingual, and multimodal variants; and / orSNTH.013WO PATENT standalone or retrieval-augmented configurations. Implementations may run on singlenode or distributed systems, on-premises, at the edge, or in the cloud. References to language or text include tokenized sequences and structured representations such as JSON or XML. Features described in connection with one embodiment may be combined with others unless context dictates otherwise.

[0012] A system of one or more computers can be configured to perform the below example operations or actions by virtue of having software, firmware, hardware, or a combination of them installed on the system that in operation causes the system to perform the actions. One or more computer programs can be configured to perform particular operations or actions by virtue of including instructions that, when executed by data processing apparatus, cause the apparatus to perform the actions.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a diagram illustrating an example of differences between non- agentic and agentic workflows in the context of LLM-based (Large Language Modelbased) agents.

[0014] Figure 2 is a diagram illustrating an example of agentic reasoning design patterns.

[0015] Figure 3 illustrates an example of an agentic framework flow within a medical image management platform.

[0016] Figure 4 is a block diagram illustrating an example system architecture for a clinical imaging platform that incorporates an agent selector to dynamically manage and deploy Al agents based on user interactions, contextual parameters, and other inputs.

[0017] Figure 5 is an example user interface of the medical image management platform, showing various functionalities designed to assist users in managing and interpreting medical images and reports.

[0018] Figure 6 illustrates another example user interface that allows a user to det preferences for agents, tasks, and triggers.DETAILED DESCRIPTION

[0019] Embodiments of the invention will now be described with reference to the accompanying figures, wherein like numerals refer to like elements throughout. TheSNTH.013WO PATENT terminology used in the description presented herein is not intended to be interpreted in any limited or restrictive manner, simply because it is being utilized in conjunction with certain specific embodiments. Furthermore, embodiments of the invention may include several novel features, no single one of which is solely responsible for its desirable attributes or which is essential to practicing the inventions herein described.

[0020] Although certain preferred embodiments and examples are disclosed below, inventive subject matter extends beyond the specifically disclosed embodiments to other alternative embodiments and / or uses and to modifications and equivalents thereof. Thus, the scope of the claims appended hereto is not limited by any of the particular embodiments described below. For example, in any method or process disclosed herein, the acts or operations of the method or process may be performed in any suitable sequence and are not necessarily limited to any particular disclosed sequence. Various operations may be described as multiple discrete operations in turn, in a manner that may be helpful in understanding certain embodiments; however, the order of description should not be construed to imply that these operations are order dependent. Additionally, the structures, systems, and / or devices described herein may be embodied as integrated components or as separate components. For purposes of comparing various embodiments, certain aspects and advantages of these embodiments are described. Not necessarily all such aspects or advantages are achieved by any particular embodiment. Thus, for example, various embodiments may be carried out in a manner that achieves or optimizes one advantage or group of advantages as taught herein without necessarily achieving other aspects or advantages as may also be taught or suggested herein.

[0021] The systems and methods discussed herein may be performed by various computing systems, which are referring to herein generally as a computing system.Overview

[0022] An Al agent is a software entity that performs tasks using artificial intelligence techniques. IBM describes an Al agent as “a system or program that is capable of autonomously performing tasks on behalf of a user or another system by designing its workflow and utilizing available tools. These agents can be deployed in various applications to solve complex tasks in various enterprise contexts from softwareSNTH.013WO PATENT design and IT automation to code-generation tools and conversational assistants. They may use the advanced natural language processing techniques of large language models (LLMs) to comprehend and respond to user inputs step-by-step and determine when to call on external tools.

[0023] Al agents are used in various applications, from virtual assistants like Siri and Alexa to more complex systems like self-driving cars and recommendation engines.Examples of Agents:• Virtual Assistants: Siri, Alexa, and Google Assistant help users with tasks like setting reminders, answering questions, and controlling smart home devices.• Chatbots: Customer service bots on websites that assist with inquiries, bookings, and troubleshooting.• Recommendation Systems: Netflix and Amazon use Al agents to suggest movies, shows, and products based on user preferences.• Autonomous Vehicles: Self-driving cars from companies like Tesla and Waymo use Al to navigate and make driving decisions.• Robotic Process Automation (RPA): Software bots that automate repetitive tasks in business processes, such as data entry and invoice processing.• Game Al: Non-player characters (NPCs) in video games that adapt to player actions and provide a challenging experience.• Healthcare Al: Systems that assist in diagnosing diseases, recommending treatments, and managing patient care.Example Agent Types

[0024] Al agents can be of various types, each with unique characteristics. Some of these agent types, each of which may be included in agents that are usable by the Al system discussed here, may include:1 . Simple reflex agents:• Preprogrammed to perform actions that correspond to certain conditions being met. Example: a thermostatSNTH.013WO PATENT2. Model-based reflex agents:Use both their current perception and memory to maintain an internal model of the world. As the agent continues to receive new information, the model is updated. Example: Robotic vacuum cleaner3. Goal-based agents:• Have an internal model of the world and also a goal or set of goals. Search for action seguences that reach their goal and plan these actions before acting on them. Example: A navigation system that recommends the fastest route to your destination.4. Utility-based agents:• Utility-based agents select the sequence of actions that reach the goal and maximize utility or reward. Utility may be calculated using a utility function, measuring the usefulness of an action or how “happy” it will make the agent, to each scenario based on a set of fixed criteria. These agents may be useful in cases where multiple scenarios achieve a desired goal, and an optimal one must be selected. Example: A navigation system that recommends the route to your destination that optimizes fuel efficiency and minimizes the time spent in traffic and the cost of tolls.5. Learning agents• Unique in their ability to learn. New experiences are added to their initial knowledge base, which occurs autonomously. This learning enhances the agent’s ability to operate in unfamiliar environments. Learning agents may be utility or goal-based in their reasoning and are comprised of four main elements: a. Learning: This improves the agent’s knowledge by learning from the environment through its precepts and sensors. b. Critic: This provides feedback to the agent on whether the quality of its responses meets the performance standard. c. Performance: This element is responsible for selecting actions upon learning.SNTH.013WO PATENT d. Problem generator: This creates various proposals for actions to be taken.• Example: learning agents used to personalize radiologists’ workflow. These agents may track user activity and preferences in their memory. For example, a radiologist may frequently adjust the display settings for certain types of MRI scans or prefer specific tools for analyzing certain types of scans (e.g., chest X-rays). The learning agent captures these preferences and uses this information to recommend customized settings and tools for future cases. The cycle repeats each time new recommendations are made. The radiologist's activity is continuously stored for learning purposes. In doing so, the agent improves its accuracy over time.Example Agent Selector

[0025] An "agent selector" (also referred to herein as an "agent orchestrator” or “Admin agent”) may itself be an agent. For example, an agent selector may be classified as a model-based reflex agent with elements of a goal-based agent. This classification may be due to its ability to use current perceptions and memory to maintain an internal model of the world, as well as its capability to achieve specific goals by selecting the most appropriate Al agents based on contextual and task-specific factors.

[0026] For example, as a model-based reflex agent, the agent selector may use its internal model to evaluate user actions and contextual parameters in real-time. It may update this model as it receives new information, allowing it to make more informed decisions about which Al agents to activate. For example, when a radiologist opens a new set of brain MRI images, the agent selector (or PACS / admin agent) may consider the user, the exam modality, and the body region to activate the most relevant Al agent(s).

[0027] Additionally, an agent selector may be configured to handle various administrative tasks such as merging and splitting cases, resending cases, and troubleshooting interface issues. For example, if a technologist frequently merges multiple imaging exams of the same patient into a single record, the agent selector can learn this preference and offer to automate the merging process for future cases. Similarly, if the technologist often splits a single imaging exam into multiple records based on different body regions or modalities, the agent selector can learn this pattern and suggest it forSNTH.013WO PATENT future exams. The agent selector can also troubleshoot interface issues by using its memory of past troubleshooting steps to suggest solutions, ensuring that the most appropriate actions are taken to resolve issues efficiently. By leveraging its internal model, the agent selector enhances the efficiency and accuracy of administrative tasks, ultimately improving the overall performance of the medical imaging system. Additionally, the agent selector may exhibit characteristics of a goal-based agent because it aims to achieve specific objectives, such as optimizing workflow efficiency and providing contextually relevant assistance.

[0028] In some implementations, a primary function of an agent selector is to determine the optimal agent or set of agents to handle specific tasks by analyzing input data, context, and system conditions, for example. An agent selector dynamically selects one or more Al agents based on implicit and / or explicit user actions, based on predefined or dynamically adjustable parameters, for example. An agent selector may operate in real-time, ensuring that the most appropriate agents are deployed to assist the user. The agent selector can consider a wide range of factors, including user preferences, facility protocols, exam modalities, body regions, and patient characteristics. Additionally, the agent selector can incorporate broader contextual factors such as demographics, payer information, weather, seasonal factors, and / or the like. For example, types of injuries may vary according to the season, as do disease processes like RSV pneumonia in children during winter months. The agent selector can also take into account disease prevalence in a family or region, enabling it to provide more contextually relevant assistance. For example, if a particular region has a high prevalence of Lyme disease, the agent selector can prioritize Al agents that are specialized in identifying and diagnosing this condition. Similarly, if payer information indicates that certain imaging procedures are more likely to be approved for specific patient demographics, the agent selector can activate Al agents that optimize the selection and scheduling of these procedures. As another example, during flu season the agent selector might prioritize Al agents that assist in diagnosing respiratory conditions, while in summer, it might focus on agents that help identify sports- related injuries.

[0029] In some implementations, agents can work on parameters that are not predefined. For example, an agent could draw conclusions based on a series of inputs derived from other variable inputs. Consider an agent that determines which radiologistSNTH.013WO PATENT should review a specific exam. The inputs could include radiologist preference, radiologist skill, number of exams in the queue for that radiologist, past performance metrics, and / or the quality of prior interpretations. While these inputs can be predefined, the agent could also use dynamic inputs to determine if a radiologist is getting tired, such as cadence of voice, heart rate, mouse clicks, intentionality of mouse movement, and / or eye tracking.

[0030] Al agents can be associated with specific parameters to ensure they are activated under the right circumstances. These parameters can include users, such as individual radiologists, technologists, referring physicians, laboratory technicians, pathologists, patients, payers, regulators, etc. who can have personalized Al agents that cater to their specific needs and preferences. Different healthcare facilities may have unique protocols and workflows, and Al agents can be configured to adhere to these facility-specific guidelines. Al agents can be specialized for different imaging modalities, such as MRI, CT, ultrasound, or X-ray, ensuring that they provide relevant assistance based on the type of exam. Similarly, Al agents can be tailored for laboratory workflows, such as blood tests, tissue biopsies, and genetic screenings, providing specialized analysis and recommendations based on the type of laboratory test. Al agents can be tailored to specific body regions, such as the brain, chest, abdomen, or musculoskeletal system, providing specialized analysis and recommendations. Additionally, Al agents can consider patient-specific factors, such as age, medical history, and known allergies, to provide personalized and safe recommendations.

[0031] An agent selector may provide one or more of the example functions to a medical imaging system:• Real-Time Contextual Evaluation: The agent selector may operate in real-time, continuously monitoring user actions, system events, and contextual parameters to determine the most appropriate Al agents to activate. For example, when a radiologist rearranges the image series for an MR exam, the agent selector may detect this implicit action and activate an Al agent that offers to automate the rearrangement process for future brain MR exams. In some implementations, the Al agent can remember the user’s preferences, automatically analyze image series, and display them correctly in the future. Activation may be triggered through multiple input modalities, such as voice commands, keyboard shortcuts, on-screen buttons, haptic input (touch), virtualSNTH.013WO PATENT manipulation (such as tracking eye movement or finger movement with devices like Apple Vision goggles), or even by neural communication. The agent selector may also log the triggering event, context, and rationale for activation for audit and explainability purposes.• Dynamic Adaptation: The agent selector may dynamically adapt to changing conditions by updating its internal model based on new information, such as user feedback, workflow changes, or system performance metrics. This allows the system to improve its performance over time and provide increasingly accurate and relevant assistance. For example, if a technologist uploads a scanned document related to an ultrasound exam, the agent selector may detect this implicit action and activate an Al agent that extracts discrete data from the document and enters it into the data repository. Adaptation may include learning user-specific preferences, applying site-level policies, or running new agents in “shadow mode” to collect telemetry before enabling them by default.• Multi-Agent Coordination: The agent selector can coordinate the actions of multiple Al agents, ensuring that they work together seamlessly to achieve the desired objectives. For example, if a radiologist opens a new set of brain MRI images, the agent selector can activate one Al agent to segment brain regions and another to identify potential abnormalities and suggest relevant differential diagnoses. Coordination may include managing dependencies between agents, sequencing tasks, resolving conflicts (e.g., competing hanging protocol suggestions), and / or enforcing resource constraints such as GPU availability or latency budgets.• User-Specific Customization: The agent selector can be customized to address specific needs and preferences of individual users or user groups. This ensures that the Al agents provide personalized assistance that aligns with the unique workflows and requirements of each user. For example, a referring physician reviewing a radiology report for a patient with a history of pancreatic cysts can benefit from the agent selector, which activates an Al agent that extracts key findings from the report, correlates them with the patient's medical history, and provides a summary of relevant clinical guidelines for follow-up. CustomizationSNTH.013WO PATENT may include role-based profiles, modality-specific settings, site-level overrides, and / or the like.• Parameter-Based Activation: The agent selector can activate Al agents based on a wide range of parameters, including user preferences, facility protocols, exam modalities, body regions, patient characteristics, demographics, weather, and / or any other contextual factors. Parameters may include, for example, DICOM metadata (e.g., SeriesDescription, BodyPartExamined), patient age and sex, clinical indication, prior imaging availability, and / or system conditions such as network bandwidth or compute load. This ensures that the Al agents provide contextually relevant assistance that is tailored to the specific circumstances of each scenario. The agent selector may maintain a registry of available agents with metadata describing their capabilities, input / output requirements, preconditions for activation, and the like.• Enhanced Learning and Error Mitigation: The agent selector can enhance the learning of the Al system by ensuring that Al agents consider only relevant parameters when learning custom protocols. Additionally, it can deploy secondary Al agents or rule-based validators to monitor and validate the outputs of primary Al agents, reducing the risk of erroneous outputs (hallucinations) and ensuring accuracy and relevance. For example, if a primary Al agent suggests a custom hanging protocol, a secondary Al agent can verify the suggestion against predefined rules and guidelines to ensure accuracy and relevance before applying it. The system may also maintain an audit log of all agent activations, validation results, and user actions (acceptance or override) for quality assurance, regulatory compliance, and / or as feedback to a learning module that updates one or more of the agents or user preferences.

[0032] As one example implementation, consider a scenario where a radiologist is reviewing a series of brain MRIs. The agent selector may evaluate the context and activate the most appropriate Al agents. For example, when the radiologist opens a new set of brain MRI images, the agent selector considers the user (radiologist), the exam modality (MRI), and the body region (brain). Based on these parameters, the agent selector activates an Al agent specialized in neuroimaging. This agentSNTH.013WO PATENT automatically segments brain regions, identifies potential abnormalities, and suggests relevant differential diagnoses. If the radiologist rearranges the image series to optimize their workflow, the agent selector may detect this implicit action and activate another Al agent that offers to automate the rearrangement process for future brain MR exams.

[0033] As another example, consider a technologist that is responsible for acquiring ultrasound images. When the technologist completes an ultrasound exam of the abdomen, the agent selector considers the user (technologist), the exam modality (ultrasound), and the body region (abdomen). The agent selector may then activate an Al agent specialized in abdominal ultrasound. This agent may analyze the images for quality assurance, ensuring that all required views are captured and flagging any issues that may require a repeat scan. If the technologist uploads a scanned document related to the exam, the agent selector may detect this implicit action and activate another Al agent that extracts discrete data from the document and enters it into the data repository.

[0034] In additional embodiments, Al agents may supervise, correct, or act on forms completed by technologists during the performance of medical imaging exams. For example, when a technologist completes a pelvic ultrasound and documents a thickened endometrium, an Al agent may provide additional instructions based on patient-specific parameters such as age, ethnicity, menstrual history, or other relevant factors. If a uterine mass is reported, the agent may prompt the technologist to obtain further characteristics such as precise location, vascularity, and calcifications, and may modify the technologist form to derive a lesion classification based on a respected staging system. Similarly, in obstetrical ultrasound, when a technologist measures fetal head size, current systems may use hard-coded formulas to calculate expected fetal age and percentile. In contrast, an agentic workflow may customize these calculations using patient demographics, clinical history, prior exams, and / or maternal / paternal biometric data. These example implementations demonstrate how agentic workflows can enhance not only the interpretation of exams by reading physicians but also the acquisition and documentation processes performed by technologists.

[0035] As another example, a referring physician reviewing a radiology report can also benefit from the agent selector. For example, when the physician opens a radiology report for a patient with a history of pancreatic cysts, the agent selector may consider the user (referring physician), the patient characteristic (history of pancreaticSNTH.013WO PATENT cysts), and the facility's protocol for follow-up. The agent selector may then select and activate an Al agent (or multiple agents) that extracts key findings from the report, correlates them with the patient's medical history, and provides a summary of relevant clinical guidelines for follow-up. If the physician asks a clinical question about the best imaging modality for evaluating pancreatic cysts, the agent selector may detect this implicit action and activate another Al agent that provides evidence-based recommendations based on clinical guidelines and literature.

[0036] A patient who has access to her own medical images, may wish to read a report in more simplistic language. An agent can be activated. This agent could translate the report to nonmedical terminology and draw an illustration that accurately depicts the pathology. It could even simulate a video of the surgical procedure that could be used to treat this patient’s condition.

[0037] In another example, a patient who has access to her own medical images may wish to read a report in more simplistic language. In this scenario, an Al agent can be selected by the agent selector and activated to assist the patient in understanding the medical information. This agent could translate the complex medical terminology in the report into nonmedical, layman's terms, making it easier for the patient to comprehend the findings and recommendations. Additionally, the agent selector may select an agent that generates visual aids to further enhance understanding. For example, an agent may draw an illustration that accurately depicts the pathology described in the report. This visual representation can help the patient visualize the condition, providing a clearer understanding of what is happening inside their body. The agent selector may select another agent that simulates a video of the surgical procedure that might be used to treat the patient's condition. This video could include a step-by-step explanation of the procedure, highlighting key aspects and potential outcomes. The agent select may select another agent that offers educational resources tailored to the patient's specific condition, such as articles, videos, and interactive tools that explain the condition, potential treatments, and lifestyle changes that could improve the patient's health. In other implementations, other agents may be selected.

[0038] By incorporating an agent selector that dynamically selects Al agents based on implicit and / or explicit user actions, the medical image management platform can provide a highly responsive and adaptive user experience. This approach ensuresSNTH.013WO PATENT that the Al agents operate in a manner that aligns with the unique needs and preferences of each user, ultimately improving the quality and efficiency of medical imaging workflows.Example System Features

[0039] The systems and methods discussed herein may result in one or more of the technical features or advantages:• Individual users or user groups may deploy specified agents under specified circumstances. For example, a user may indicate to the agent selector which specific agent(s) to execute.• Agents can be activated by explicit user actions (such as asking a question or clicking a button) or may launch as a result of user actions (such as rearranging image series, completing a form, dictating a recommendation, asking a clinical question). For example, a user might rearrange medical image series, then say, “do it this way for all MR exam types,” which causes the agent selector to select an agent configured to implement the requested user preference. During the next exam, the user may again rearrange image series, and the Al service may ask, “Do you want me to do this for you next time for brain MR exams or all MR exams?” The learning of the system may be enhanced or constrained from hallucinations by adding conventionally programmed constraint or another Al agent. For example, the system may be instructed to only take into account the body region and modality but not other exam parameters in learning the custom hanging protocol. Agents can be activated in other manners, such as by pressing a keyboard shortcut or a button, a haptic input (touch), virtual manipulation (such as tracking eye movement or finger movement with devices like Apple Vision goggles), or even by neural communication.• Agents may be associated with certain users, facilities, exam modalities, exam body regions, patient characteristics or other such parameters to further specify by preference which agents act when.• Agents may also be associated with other agents that inform their activities. For example, predictive agents could forecast temporal events, such as increased demand in a hospital system based on real-time traffic updates related to inclement weather or news of a mass casualty event. This predictiveSNTH.013WO PATENT agent could then activate another agent to recruit more radiologists within a network to read exams within a specific time period following a detected mass casualty event nearby. By coordinating with other agents, the system can dynamically adjust resources and workflows to respond effectively to changing conditions and emergent situations.• Agents may respond using an LLM to generate text or speech context. For example, when a user states a diagnosis for a medical imaging finding, the Al agent may respond by presenting a differential diagnosis. When a user states a recommendation, the Al agent may respond by presenting extracted information from a published guideline. Here again, the agent selector is configure to customize the use of reference data to allow speed learning in a healthcare setting, while also reducing hallucinations. For example, the user or facility may provide a published guideline for follow up of pancreatic cystic lesions so that the Agent’s “advice” is restricted to being based on a specified source.• The agents selected by the agent selector may be customized to certain tasks based on the user class. For example, when a technologist provides a scanned document, the agent may analyze the document to enter discrete data in a data repository and even respond to the user as to what additional data may be required or useful.• Agents may make autonomous decisions as to how to best store and present data. For example, suppose a patient has 10 various radiology reports. An agent many extract various diagnoses and build an active problem list vs. resolved problem list and store at the patient level. Suppose a patient has an allergic reaction in response to IV contrast given for a CT scan... the agent may extract this information and store at the patient level so that a warning can be issued when IV contrast is contemplated for this patient in the future.Example of Medical Imaging Agents

[0040] The medical image management platform can deploy a variety of specialized Al agents to enhance different aspects of radiology workflows. These agentsSNTH.013WO PATENT are designed to perform specific tasks, ranging from administrative functions to advanced clinical decision support. Below are detailed examples of such agents:• Radiology-Specific Summary Creation: This Al agent can automatically generate a concise summary of radiology reports, highlighting key findings, recommendations, and relevant clinical guidelines. For example, after a radiologist completes a diagnostic report, the agent can extract essential information and create a summary that can be easily reviewed by referring physicians.• RIS PACS Admin Functions: This set of Al agents can handle various administrative tasks within the Radiology Information System (RIS) and Picture Archiving and Communication System (PACS). Examples include: o Merging Exams: An Al agent can merge multiple imaging exams of the same patient into a single record, ensuring that all relevant images and data are consolidated. o Splitting Exams: An Al agent can split a single imaging exam into multiple records if the exam covers different body regions or modalities, facilitating more accurate reporting and analysis. o Unlocking Exams: An Al agent can unlock exams that have been inadvertently locked, allowing radiologists and technologists to make necessary edits or additions.• Clinical Workflow Enhancements: These Al agents can streamline various clinical workflows, improving efficiency and accuracy, such as: o Autocorrection: An Al agent can automatically correct common errors in radiology reports, such as typographical mistakes or incorrect terminology, ensuring that the final report is accurate and professional. o Recommendation Engines: An Al agent can provide evidence-based recommendations for follow-up tests, treatments, or additional imaging studies based on the initial findings and patient history. o Generating Diagrams to Represent Pathology: An Al agent can create visual diagrams that represent identified pathologies, helping radiologists and referring physicians better understand and communicate the findings.SNTH.013WO PATENT• Exam Auto-assignment: An Al agent can automatically assign imaging exams to radiologists based on their specialization, workload, and availability, ensuring that exams are reviewed promptly and by the most qualified professionals.• Intelligent Caching: An Al agent can predict which images or data will be needed next and cache them in advance, reducing loading times and improving workflow efficiency.• Predictive Scheduling for Customers: An Al agent can analyze historical data and predict staffing needs for radiologists and technologists, optimizing scheduling to ensure adequate coverage and minimize wait times for patients.• LLM Search Functions: These Al agents can leverage large language models (LLMs) to perform advanced search and data retrieval tasks: For example, this agent can search for all ultrasound exams with specific findings, such as bursal surface tears of the supraspinatus tendon, and create a folder with these exams. The agent can then share the folder with relevant staff, such as ultrasound technologists, for further review.• Creating Differential Diagnoses: An Al agent can generate a list of potential differential diagnoses based on the initial findings, helping radiologists consider all possible conditions and make more informed decisions.• Clinical Correlation: An Al agent can integrate with electronic health record (EHR) systems, such as Epic, to retrieve relevant clinical data and correlate it with imaging findings. For example, if clinical correlation is required for a specific diagnosis, the agent can automatically pull the necessary data from the EHR and present it to the radiologist.• Agent Creation: An Al agent can assist users in creating their own custom Al agents. This capability would empower users to tailor the system even more precisely to their specific needs and workflows. For example, a radiologist could create an agent to find all biopsy-positive cases of thyroid nodules. This agent could then be used to develop an image analysis tool that helps predict which nodules are likely to be positive. The output of this tool could be automatically displayed whenever a thyroid ultrasound is performed, providing real-time decision support.SNTH.013WO PATENT• Report Assistant: An Al agent may be activated by an agent selector when a radiologist completes or begins finalizing a diagnostic report. Parameters used to trigger this agent may include user role (radiologist), report status (draft or finalized), exam modality, and body region. Customized inputs to the agent may include the full text of the report, structured metadata (e.g., exam type, clinical indication), and site-specific reporting conventions. The report assistant agent may include multiple coordinated sub-functions. First, a syntax correction module may analyze the report for grammatical accuracy, formatting consistency, and adherence to institutional style guidelines. This ensures that the report is professionally written and free of typographical or structural errors. Second, a summary generation module may extract key findings and automatically generate an impression section, tailored to the specific exam type and clinical context. This summary may be optimized for clarity and relevance, particularly for referring physicians. Third, a recommendation engine may analyze the report content and generate evidence-based follow-up suggestions. These recommendations may be informed by clinical guidelines, patient history, and prior imaging, and may be customized to reflect facility-specific protocols. Fourth, a specialized algorithm module may run the report against diagnostic algorithms tailored to the specific exam type. For example, in a chest CT, the agent may apply a lung nodule risk stratification model to validate or augment the radiologist’s conclusions.The report assistant agent may also include a report macro function that allows the radiologist to insert hints or annotations into the report. These hints may serve as cues to the agent selector, indicating which Al agents should be triggered and what type of response or output is expected. For example, a radiologist may include “recommendation needed for pancreatic cyst follow-up” in a report or comments about an exam, prompting the agent selector to activate an agent constrained to published guidelines for that condition.• Educational support agent: an Al agent may be deployed to assist trainees in generating high-quality radiology reports by providing real-time supervision and feedback. Activation of this agent may be triggered by parameters such as user role (e.g., trainee), report status (e.g., draft or submitted for review), exam modality (e.g., CT, MRI, ultrasound), and body region (e.g., abdomen, chest, pelvis).SNTH.013WO PATENTAdditional parameters may include institutional training protocols or tagging of the report as part of a learning module. Customized inputs to the agent may include the full text of the trainee-generated report, structured metadata (e.g., exam type, clinical indication), and institutional style guidelines. These guidelines may specify grammar conventions (e.g., US vs. UK spelling), preferred sentence structures (e.g., “There is an area of consolidation...” vs. “Consolidation is present...”), and formatting rules. The agent may also receive access to diagnostic checklists or disease-specific reporting templates. The agent may output a set of suggested modifications to the report, including corrections to grammar and formatting, alignment with institutional style, and identification of missing pertinent findings. For example, in a case of pancreatitis, the agent may prompt the trainee to include key elements such as the extent of pancreatic non-enhancement, characterization of peripancreatic fluid collections, presence of ascites, pleural effusions, and bowel edema, each of which may be relevant for disease grading and prognosis. The agent may also provide references from licensed clinical resources or web-based literature to support the inclusion of specific findings or terminology. In addition to report-level feedback, the agent may prompt the trainee to tag and save the exam for future reference, contributing to longitudinal learning and consistency in reporting. These tagged cases may be used to build personalized learning sets or to support future algorithm development. The agent may also log feedback interactions for review by supervising radiologists or educators, enabling oversight and refinement of training protocols.

[0041] The table below provides examples of trigger parameters, inputs, and outputs to these example agents. These are examples only and do not limit the scope of the trigger parameters, inputs, or outputs of these example agents.SNTH.013WO PATENT

[0042] As one example of how an agent selector may impact a medical imaging workflow, consider a case where a radiologist opens an abdominal CT scan for a patient with a history of liver disease. The agent selector may evaluate contextual parameters such as user role (radiologist), exam modality (CT), body region (abdomen), patient history (liver disease), and facility-specific protocols to determine which Al agents toSNTH.013WO PATENT activate. Based on these parameters, the agent selector may activate a clinical workflow enhancement agent functioning as a recommendation engine. Customized inputs to this agent may include the lesion characteristics extracted from the CT scan, the patient's prior imaging history, and relevant clinical guidelines. The agent may suggest evidencebased follow-up actions, such as recommending a contrast-enhanced MRI and a three- month follow-up interval.

[0043] After the radiologist completes the diagnostic report, the agent selector may activate a radiology-specific summary creation agent. Parameters used for activation may include the completion of a report, the user role, and the presence of structured findings. Customized inputs to this agent may include the full text of the radiologist’s report, structured metadata such as exam type and body region, and facility-specific formatting preferences. This agent may extract key findings and generate a concise summary tailored for the referring physician.

[0044] To support clinical decision-making, the agent selector may activate a clinical correlation agent. Activation may be triggered by the radiologist’s request for lab data or by the presence of findings that typically require correlation. Parameters may include patient history, body region, and exam modality. Customized inputs may include the patient’s medical record number, recent liver function test results, and diagnostic codes. This agent may retrieve relevant laboratory data, such as ALT and AST levels, from the electronic health record and correlate these results with the imaging findings.

[0045] Additionally, the agent selector may activate an agent that generates diagrams to visually represent the pathology. Parameters used may include the presence of a lesion, body region, and user role. Customized inputs may include lesion dimensions, anatomical location, and imaging slices. This agent may produce a visual illustration of the liver lesion in relation to surrounding structures, enhancing communication with referring physicians and surgical teams. These coordinated actions may improve diagnostic accuracy, streamline documentation, and enhance interdisciplinary communication, while adapting to the specific workflow and preferences of the radiologist.

[0046] As another example of how an agent selector may impact a medical imaging workflow, consider a case where a technologist completes an obstetrical ultrasound exam. The agent selector may evaluate contextual parameters such as user role (technologist), exam modality (ultrasound), body region (pelvis), patientSNTH.013WO PATENT characteristics (pregnancy status), and facility-specific documentation protocols to determine which Al agents to activate. Based on these parameters, the agent selector may activate a clinical workflow enhancement agent configured to supervise and correct technologist-completed forms. Customized inputs to this agent may include the technologist’s documentation of fetal head measurements, gestational age estimates, and maternal biometric data. The agent may analyze these inputs and suggest refinements to the expected fetal age calculation by incorporating prior imaging, demographic factors, and clinical history.

[0047] Following the documentation process, the agent selector may activate a syntax correction agent to review the technologist’s notes for grammatical consistency and adherence to institutional formatting standards. Parameters used for activation may include completion of a structured form and user role. Customized inputs may include the raw text of the technologist’s report and site-specific grammar conventions.

[0048] To support longitudinal learning and quality assurance, the agent selector may also activate an agent that tags and saves the exam for future reference. Parameters used may include the presence of key findings, user role, and facility learning protocols. Customized inputs may include the exam metadata, annotated measurements, and diagnostic codes. This agent may store the exam in a curated training set for use in educational modules or algorithm development.

[0049] Additionally, the agent selector may activate a predictive scheduling agent to analyze staffing needs based on current exam volume and historical trends. Parameters may include time of day, modality usage rates, and technologist availability. Customized inputs may include real-time queue data, facility staffing logs, and seasonal patient volume patterns. This agent may forecast increased demand and recommend adjustments to technologist schedules to maintain workflow efficiency. These coordinated actions may enhance documentation accuracy, support educational initiatives, and optimize resource allocation, while adapting to the specific workflows and responsibilities of the technologist.Example Autonomous Learning of User Preferences

[0050] The medical image management platform may be configured to autonomously learn a user’s preferences, allowing for continuous customization and improvement over time. This feature leverages advanced machine learning algorithms toSNTH.013WO PATENT observe and analyze passive user actions, thereby tailoring the system to better meet individual user needs without requiring explicit input or configuration.

[0051] In some embodiments, the platform continuously monitors user interactions with the system, such as how a radiologist arranges image series, the types of annotations they frequently make, or the specific tools they use most often. These interactions may be logged and analyzed to identify patterns and preferences, such as by an agent selector. For example, if a radiologist consistently rearranges brain MRI images in a particular order, the system may recognize this pattern and offer to automate the arrangement for future brain MRI exams.

[0052] As the system gathers more data on user behavior, it updates its internal models to reflect these preferences. This real-time adaptation ensures that the system becomes increasingly efficient and user-friendly over time. For example, if a technologist frequently adjusts the brightness and contrast settings for abdominal ultrasound images, the system can learn to apply these adjustments automatically whenever similar images are reviewed.

[0053] The autonomous learning feature also allows for enhanced customization at both the individual and group levels. Individual users, such as radiologists, technologists, and referring physicians, can benefit from a system that adapts to their unique workflows and preferences. Similarly, user groups within a healthcare facility can have the system tailored to their collective needs, ensuring that the platform supports the specific protocols and workflows of the facility.

[0054] By autonomously learning a user’s preferences, the medical image management platform offers a highly personalized and efficient user experience. This feature not only enhances the usability of the system but also significantly improves workflow efficiency, making it an invaluable tool for radiologists, technologists, and referring physicians. The ability to adapt and improve over time ensures that the platform remains a cutting-edge solution for medical imaging management.

[0055] As one example, the system can observe how a radiologist interacts with the outputs of an algorithm and adjust thresholds or weightings of various algorithms. For example, an agent may analyze whether a particular radiologist prefers more detailed information, or if they prefer a more summarized view. For example, a "lumper" radiologist tends to group or "lump" together findings under a broader, more generalSNTH.013WO PATENT diagnosis. They may emphasize the commonalities between different findings, leading to a more inclusive, overarching interpretation. In contrast, a "splitter" radiologist prefers to "split" or differentiate findings into more specific categories or distinct diagnoses. They focus on the differences and may provide a more detailed analysis of each individual finding. By analyzing a radiologist’s interactions, the agent can adjust the algorithm's output to match the radiologist's preferences. For instance, if the agent detects that a radiologist frequently expands sections of a report to view more details, it can infer that the radiologist prefers a more detailed output. Conversely, if the radiologist often collapses sections or skips over detailed information, the agent can infer a preference for a more summarized view. This adaptive learning process ensures that the system provides outputs that are tailored to the individual preferences of each radiologist, enhancing their efficiency and satisfaction. The agent can continuously refine its understanding of the radiologist's preferences by incorporating new data from ongoing interactions, ensuring that the system remains responsive to any changes in the radiologist's workflow or preferences.

[0056] Moreover, this approach can be extended to other users, such as technologists and referring physicians, ensuring that the system provides customized outputs that align with the specific needs and preferences of each user.Example Shared Learning

[0057] Once the system identifies effective workflows and preferences from experienced users, it can transfer this knowledge to other users. For example, if an experienced radiologist consistently uses a particular sequence of tools to analyze brain MRI images, the agent selector can activate an Al agent that suggests this sequence to other radiologists who are less familiar with the process. This knowledge transfer can be done automatically or with user approval, depending on the system's configuration.

[0058] These knowledge sharing features may be particularly beneficial for training and onboarding new users. By learning from the actions of experienced users, the system can provide new users with guided workflows and recommendations that align with best practices. For example, a new technologist can receive automated suggestions on how to adjust image settings based on the preferences of more experienced technologists, thereby reducing the learning curve and improving the quality of their work.SNTH.013WO PATENTThe agent selector can activate Al agents that provide these guided workflows and recommendations.

[0059] Additionally, learning from one user to another may also help maintain consistency in workflows and decision-making across the user base. This may be especially important in healthcare settings where standardized procedures and protocols are crucial for patient safety and care quality. By adopting the best practices of experienced users, the system ensures that all users follow consistent and effective workflows. The agent selector can activate Al agents that monitor and enforce these standardized procedures.

[0060] In some embodiments, Al agents may provide supervision and educational support to trainees by reviewing trainee-generated reports and suggesting modifications. These suggestions may be based on a specific radiologist’s reporting style, which could reflect institutional preferences such as grammar conventions (e.g., US vs. Canadian vs. UK spelling), sentence structure (e.g., “There is an area of consolidation in the left lower lobe” vs. “Consolidation in the left lower lobe”), and formatting. Agents may also identify missing pertinent facts that are important for accurate diagnosis and prognosis. For example, in cases of pancreatitis, agents may prompt inclusion of key findings such as the extent of pancreatic non-enhancement, presence and characterization of peripancreatic fluid collections, ascites, pleural effusions, and bowel edema. These elements are useful for disease grading and prognostication, yet are often omitted by trainees. Additionally, agents may provide references from licensed resources or web-based literature to support report content and prompt trainees to tag and save exams for future reference, thereby facilitating longitudinal learning and consistency in reporting.Example Diagrams and User Interfaces

[0061] Figure 1 is a diagram illustrating an example of differences between non-agentic and agentic workflows in the context of LLM-based (Large Language Modelbased) agents. The figure contrasts two approaches to completing a task, such as writing an essay on a given topic.

[0062] First, a non-agentic workflow (zero-shot) approach is illustrated on the left side of the figure. This approach involves a straightforward, linear process where theSNTH.013WO PATENT prompt instructs the LLM to type out an essay on topic X from start to finish without any intermediate steps or iterative improvements.

[0063] On the right side of the figure, the agentic workflow is depicted. This approach involves a more complex, iterative process that leverages the capabilities of LLM-based agents to assist the user at various stages. This iterative process may be repeated to provide incremental improvements and refinement. The agentic workflow allows for multiple stages of feedback, research, and revision, enabling the user to produce a more polished and well-researched final product.

[0064] Figure 2 is a diagram illustrating an example of agentic reasoning design patterns. In this example, reflection and tool use are classified under robust technology, representing well-established and reliable methods in agentic reasoning. Reflection may involve, for example, the agent's ability to reflect on its actions and decisions, learning from past experiences to improve future performance. Tool use may refer to the agent's capability to utilize various tools, such as other agents, to accomplish tasks, enhancing its functionality by leveraging external resources. Planning and multiagent collaboration are categorized under emerging technology, representing methods in agentic reasoning that are still evolving. Planning may involve the agent's ability to strategize and plan its actions to achieve specific goals, enabling it to anticipate future scenarios and make informed decisions. Multi-agent collaboration may involve the ability of multiple agents to work together to achieve common objectives, coordinating their actions and sharing information to enhance overall performance.

[0065] Figure 3 illustrates an example of an agentic framework flow within a medical image management platform. In this implementation, an agent selector performs real-time evaluation of incoming inputs and orchestrates the activation of one or more Al agents based on contextual parameters and task requirements. Inputs to the agent selector may include voice commands, natural language text, operational parameters (OPRs), Al algorithm outputs, medical images, structured metadata such as exam modality, body region, and user role, and / or any other available data.

[0066] The processing stage may involve a combination of API calls, prompts to large language models (LLMs), sequential logic, command execution, and agent activation. For example, when a radiologist opens a chest CT scan and begins dictating a report, the agent selector may process the voice input and report status to determineSNTH.013WO PATENT which agents to activate. Based on these inputs, the agent selector may trigger a report assistant agent to perform syntax correction, generate a summary impression, and run the report against a lung nodule risk stratification algorithm. Simultaneously, a clinical correlation agent may be activated to retrieve relevant lab data from the electronic health record, and a recommendation engine agent may provide evidence-based follow-up suggestions.

[0067] Each agent performs a specialized function and contributes to the overall processing workflow. Outputs from the system may include corrected report text, summarized impressions, diagnostic recommendations, segmented image overlays, and visual diagrams, each tailored to the user’s workflow and clinical context. These outputs may be returned to the user interface for interaction or forwarded to downstream systems for further analysis, documentation, or decision support.

[0068] This agentic framework enables dynamic, context-aware coordination of Al agents, ensuring that each task is handled by the most appropriate agent based on real-time conditions and user-specific parameters.

[0069] Figure 4 is a block diagram illustrating an example system architecture for a clinical imaging platform that incorporates an agent selector to dynamically manage and deploy Al agents based on user interactions, contextual parameters, and other inputs. In this example, a user 405 interacts with a user interface module 410, which may be implemented as software code that generates tailored user interfaces for different user roles, such as radiologists, technologists, referring physicians, or patients.

[0070] The user interface module 410 may include an image viewer for displaying and manipulating medical images (e.g., MRIs, CT scans, X-rays), with tools for zooming, panning, adjusting brightness and contrast, and annotating images using lines, circles, arrows, and text notes. The module may also support report generation, allowing users to dictate or type findings, recommendations, and conclusions. Integrated Al functionality may provide real-time assistance, such as suggesting differential diagnoses, highlighting potential abnormalities, and offering evidence-based recommendations for follow-up imaging or treatment. A report viewer may also be included, enabling users to search, filter, and review diagnostic reports along with associated images and annotations.SNTH.013WO PATENT

[0071] The agent selector 415 may evaluate user actions and contextual parameters in real-time, including inputs from the internal data source 430, the user interface module 410, and other system components, to identify tasks requested by the user, either explicitly (e.g., voice command or button click) or implicitly (e.g., rearranging image series or completing a form). Based on this evaluation, the agent selector may activate one or more Al agents from the agent pool 425, each designed to perform specific tasks. The agent selector may continuously update its internal model based on new information, allowing it to adapt to changing conditions and improve performance over time. Depending on the implementation, the agent selector may operate locally on the user device or remotely in the cloud.

[0072] The agent pool 425 includes a collection of Al agents 455A-455N that may be invoked by the agent selector. These agents may include those for generating radiology-specific summaries, performing RIS / PACS administrative functions (e.g., merging or splitting exams), enhancing clinical workflows (e.g., autocorrection, recommendation engines, diagram generation), executing advanced search and retrieval tasks, and more. Each agent may be associated with specific activation parameters, such as user role, exam modality, body region, or patient characteristics.

[0073] To complete assigned tasks, the agents 455 may access various data sources. For example, an external data source 470 may include third-party content such as published medical guidelines, clinical studies, and research articles. External Al algorithms 440 may provide additional processing capabilities, such as advanced image analysis, machine learning models, or natural language processing pipelines. The synthetic data source 430 may include data generated or processed by the agent pool or external sources, which may be used for training, validation, or simulation purposes. A large language model (LLM) 450 may interact with the agents and / or the agent selector to support natural language understanding, summarization, recommendation generation, and other language-based tasks.

[0074] Figure 5 is an example user interface of the medical image management platform, showing various functionalities designed to assist users in managing and interpreting medical images and reports. In this example, an agent dialog, e.g., a chat window, has been launched, which allows the user to interact with an Al assistant named "Rae." The chat window displays a conversation between the user andSNTH.013WO PATENT the Al assistant, with the user requesting support for missing images and the Al assistant responding with relevant actions. The chat interface includes options for sending messages, viewing conversation history, and accessing additional tools.

[0075] In this scenario, the reading physician converses with the agent to indicate that some images are missing. This invention customizes the use of agents so that the agent may only appear when the user is a reading physician and not a patient or referring physician. In the future, the agent may also warn the technologist when images are missing based on what is learned from the reading physician. This tailored interaction ensures that the Al assistant provides contextually relevant support, enhancing the efficiency and accuracy of medical imaging workflows.

[0076] Figure 6 illustrates another example user interface that allows a user to det preferences for agents, tasks, and triggers. In this particular example, the pane is designed to configure and customize the Al assistant, Rae, to perform specific tasks and workflows based on user-defined protocols and instructions.Example Implementations

[0077] Examples of the implementations of the present disclosure can be described in view of the following example clauses. The features recited in the below example implementations can be combined with additional features disclosed herein. Furthermore, additional inventive combinations of features are disclosed herein, which are not specifically recited in the below example implementations, and which do not include the same features as the specific implementations below. For sake of brevity, the below example implementations do not identify every inventive aspect of this disclosure. The below example implementations are not intended to identify key features or essential features of any subject matter described herein. Any of the example clauses below, or any features of the example clauses, can be combined with any one or more other example clauses, or features of the example clauses or other features of the present disclosure.

[0078] Clause 1 . A computerized image management system, comprising: one or more Al agents configured to perform specific tasks related to medical image management, a configuration module that allows preferences to be set for which Al agents are activated based on various parameters, wherein the parameters include one or more of user parameters, exam parameters, patient parameters, or facility parameters;SNTH.013WO PATENT an agent selector that dynamically selects and manages the Al agents based on real-time user interactions and contextual parameters, continuously updating its internal model to adapt to changing conditions and improve performance over time; and a learning module that enables the system to autonomously learn user preferences and improve customization over time, based on passive user actions and interactions with the system.

[0079] Clause 2. The system of clause 1 , wherein the agents include perform functions including one or more of: generating radiology-specific summaries, handling RIS PACS admin functions, providing clinical workflow enhancements, performing advanced search, or data retrieval tasks.

[0080] Clause 3. The system of clause 1 , wherein user-specific parameters are associated with one or more of: a radiologist, technologist, referring physician, laboratory technician, pathologist, or patient.

[0081] Clause 4. The system of clause 1 , wherein exam-specific parameters are associated with one or more of: an imaging modality (e.g., MRI, CT, ultrasound, X- ray) or laboratory test (e.g., blood test, tissue biopsy, genetic screening).

[0082] Clause 5. The system of clause 1 , wherein patient-specific parameters are associated with one or more of: age, medical history, known allergies, and other relevant characteristics.

[0083] Clause 6. The system of clause 1 , wherein facility-specific parameters are associated with one or more of: unique protocols or workflows.

[0084] Clause 7. The system of clause 1 , further comprising: a constraint module that allows the Al agents to be constrained based on reference material, wherein the reference material is includes one or more of the following: user-specific reference material, such as published guidelines, clinical protocols, or user-defined rules; examspecific reference material, such as imaging protocols, laboratory test guidelines, or diagnostic criteria; patient-specific reference material, such as personalized treatment plans, medical history records, or genetic information; or facility-specific reference material, such as institutional policies, standard operating procedures, or compliance requirements.

[0085] Clause 8. The system of clause 1 , wherein the agent selector includes a coordination module configured to manage communication of information between agents.SNTH.013WO PATENT

[0086] Clause 9. A computer-implemented method for managing medical image data in a clinical system, the method comprising: presenting a user interface displaying a plurality of medical images, each associated with user preference settings; receiving implicit or explicit user actions indicating preferences for display and output of medical images; detecting patterns in these user actions to learn and autonomously improve user-specific customization over time; dynamically activating Al agents based on one or more of contextual parameters, user interactions, or other inputs to perform specific tasks tailored to a workflow; continuously updating an internal model based on new information to adapt to changing conditions and improve performance over time; and providing personalized and efficient assistance to users in managing and interpreting medical images.

[0087] Clause 10. The system of clause 1 , wherein the agent selector maintains a registry of available Al agents, each associated with metadata describing capabilities, input / output requirements, activation conditions, and dependencies.

[0088] Clause 11. The system of clause 1 , wherein the learning module autonomously adjusts agent activation thresholds and preferences based on observed user behavior patterns.

[0089] Clause 12. The system of clause 1 , further comprising a validation module that monitors outputs of selected Al agents and applies rule-based or Al-based verification to reduce erroneous outputs.

[0090] Clause 13. The system of clause 1 , wherein the agent selector is triggered by implicit user actions including rearranging image series, completing forms, uploading documents, or interacting with specific interface elements.

[0091] Clause 14. The system of clause 1 , wherein the Al agents include a report assistant agent comprising sub-functions for syntax correction, impression generation, recommendation generation, and diagnostic algorithm execution.

[0092] Clause 15. The system of clause 14, wherein the report assistant agent includes a macro function that allows the user to insert hints into the report metadata to guide agent selection.

[0093] Clause 16. The system of clause 1 , wherein the Al agents include an educational support agent configured to review trainee-generated reports, suggestSNTH.013WO PATENT modifications based on institutional style guidelines, identify missing diagnostic elements, and provide references to clinical literature.

[0094] Clause 17. The system of clause 1 , wherein the Al agents include a patient-facing explanation agent configured to translate diagnostic reports into layperson language, generate visual illustrations of pathology, and simulate treatment procedures for patient education.

[0095] Clause 18. The system of clause 1 , wherein the agent selector includes a coordination module that manages task sequencing, resource allocation, and conflict resolution among concurrently activated agents.

[0096] Clause 19. The system of clause 1 , wherein the agent selector incorporates broader contextual factors including one or more of demographics, payer information, weather, seasonal factors, or disease prevalence to inform agent selection.

[0097] Clause 20. The system of clause 1 , wherein the agent selector is configured to activate predictive agents that forecast temporal events and trigger resource allocation agents in response.

[0098] Clause 21. The system of clause 1 , wherein the Al agents autonomously determine optimal data storage and presentation formats based on user interaction history and clinical relevance.

[0099] Clause 22. The system of clause 1 , wherein the agent selector and Al agents are configured to operate in a distributed architecture across cloud and edge devices to support real-time responsiveness and scalability.

[0100] Clause 23. The system of clause 1 , wherein the Al agents include a radiology-specific summary creation agent configured to trigger based on one or more of: report status, user role, or exam modality; wherein the agent selector provides one or more of: full report text, structured findings, or clinical indication as inputs to the radiologyspecific summary creation agent; and the radiology-specific summary creation agent outputs one or more of: a summary impression, key findings, or guideline references.

[0101] Clause 24. The system of clause 1 , wherein the Al agents include a RIS PACS merging agent configured to trigger based on one or more of: multiple exams for the same patient or overlapping body regions; wherein the agent selector provides one or more of: patient ID, exam metadata, or timestamps as inputs to the RIS PACS mergingSNTH.013WO PATENT agent; and the RIS PACS merging agent outputs a unified exam record with consolidated images.

[0102] Clause 25. The system of clause 1 , wherein the Al agents include a RIS PACS splitting agent configured to trigger based on one or more of: a single exam covering multiple body regions or modalities; wherein the agent selector provides one or more of: image series tags or exam metadata as inputs to the RIS PACS splitting agent; and the RIS PACS splitting agent outputs separate exam records per region or modality.

[0103] Clause 26. The system of clause 1 , wherein the Al agents include a RIS PACS unlocking agent configured to trigger based on one or more of: locked exam status or user role; wherein the agent selector provides one or more of: exam ID or lock reason as inputs to the RIS PACS unlocking agent; and the RIS PACS unlocking agent outputs an unlocked exam ready for editing.

[0104] Clause 27. The system of clause 1 , wherein the Al agents include an autocorrection agent configured to trigger based on one or more of: report draft status or user role; wherein the agent selector provides one or more of: report text or grammar rules as inputs to the autocorrection agent; and the autocorrection agent outputs a corrected report with standardized language.

[0105] Clause 28. The system of clause 1 , wherein the Al agents include a recommendation engine agent configured to trigger based on one or more of: report completion or presence of findings; wherein the agent selector provides one or more of: report findings, prior imaging, or clinical guidelines as inputs to the recommendation engine agent; and the recommendation engine agent outputs one or more of: follow-up imaging or treatment recommendations.

[0106] Clause 29. The system of clause 1 , wherein the Al agents include a diagram generation agent configured to trigger based on one or more of: pathology identification, body region, or user role; wherein the agent selector provides one or more of: lesion dimensions, anatomical location, or image slices as inputs to the diagram generation agent; and the diagram generation agent outputs a visual diagram of the pathology.

[0107] Clause 30. The system of clause 1 , wherein the Al agents include an exam auto-assignment agent configured to trigger based on one or more of: new exam arrival, radiologist availability, or specialization; wherein the agent selector provides oneSNTH.013WO PATENT or more of: radiologist profiles, workload metrics, or exam metadata as inputs to the exam auto-assignment agent; and the exam auto-assignment agent outputs an assigned exam to an appropriate radiologist.

[0108] Clause 31. The system of clause 1 , wherein the Al agents include an intelligent caching agent configured to trigger based on one or more of: user activity or image access patterns; wherein the agent selector provides one or more of: prior access logs or system latency data as inputs to the intelligent caching agent; and the intelligent caching agent outputs preloaded image series for faster access.

[0109] Clause 32. The system of clause 1 , wherein the Al agents include a predictive scheduling agent configured to trigger based on one or more of: time of day, modality usage, or staffing levels; wherein the agent selector provides one or more of: historical volumes, staff schedules, or seasonal trends as inputs to the predictive scheduling agent; and the predictive scheduling agent outputs one or more of: staffing recommendations or shift adjustments.

[0110] Clause 33. The system of clause 1 , wherein the Al agents include an LLM search agent configured to trigger based on one or more of: user query, exam type, or finding keywords; wherein the agent selector provides one or more of: natural language query or structured exam data as inputs to the LLM search agent; and the LLM search agent outputs a folder of matching exams shared with relevant staff.

[0111] Clause 34. The system of clause 1 , wherein the Al agents include a differential diagnosis agent configured to trigger based on one or more of: report findings, body region, or modality; wherein the agent selector provides one or more of: imaging features, clinical context, or diagnostic models as inputs to the differential diagnosis agent; and the differential diagnosis agent outputs a ranked list of possible diagnoses.

[0112] Clause 35. The system of clause 1 , wherein the Al agents include a clinical correlation agent configured to trigger based on one or more of: diagnosis requiring lab data or user request; the agent selector provides one or more of: patient ID, EHR access, or lab results as inputs to the clinical correlation agent; and the clinical correlation agent outputs correlated clinical data presented with imaging.

[0113] Clause 36. The system of clause 1 , wherein the Al agents include an agent creation agent configured to trigger based on a user request to build a custom agent, wherein the agent selector provides one or more of: user-defined criteria, historicalSNTH.013WO PATENT cases, or imaging features as inputs to the agent creation agent; and the agent creation agent outputs a new agent with defined task and activation rules.

[0114] Clause 37. The system of clause 1 , wherein the Al agents include a patient-facing explanation agent, configured to trigger based on one or more of: patient access to report or user role; wherein the agent selector provides one or more of: report text or pathology description as inputs to the patient-facing explanation agent; and the patient-facing explanation agent outputs one or more of: a layman summary, illustrations, or video simulation.

[0115] Clause 38. A computerized clinical imaging platform comprising: a user interface module configured to generate one or more user interfaces tailored to a user role, the user interface module comprising: an image viewer configured to display and manipulate medical images, including tools for one or more of: zooming, panning, adjusting brightness and contrast, or annotating images; a report generation interface configured to receive dictated or typed diagnostic findings, recommendations, or conclusions; a report viewer configured to display diagnostic reports and associated image annotations; an Al integration interface configured to present outputs from one or more Al agents, including one or more of: differential diagnoses, abnormality highlights, or evidence-based recommendations; an agent selector configured to: evaluate real-time user interactions and contextual parameters, including one or more of: user role, exam modality, body region, patient history, or system conditions; identify a task requested by the user, either explicitly or implicitly; select and activate one or more Al agents from an agent pool based on the evaluation; provide input data to the selected Al agents, including one or more of: report text, image data, operational parameters, structured metadata, or external data sources; receive and coordinate outputs from the selected Al agents; continuously update an internal model based on new information to adapt agent selection and improve performance over time.

[0116] Clause 39. The platform of clause 38, further comprising: an agent pool comprising a plurality of Al agents, each configured to perform a specific task related to medical image management, wherein each agent is associated with metadata describing one or more of: capabilities, input / output requirements, activation conditions, or dependencies; one or more data sources, including: an internal data source comprising structured or unstructured clinical data; an external data source comprising third-partySNTH.013WO PATENT information including one or more of: medical guidelines, research articles, or clinical studies; a synthetic data source comprising data generated or processed by the agent pool or external sources; one or more external Al algorithms configured to enhance agent capabilities through one or more of: advanced image processing, machine learning, or natural language processing techniques.

[0117] Clause 40. The platform of clause 39, further comprising: a large language model (LLM) configured to interact with the agent selector and one or more Al agents to support one or more of: natural language understanding, summarization, recommendation generation, or other language-based tasks.

[0118] Clause 41 . The platform of clause 40, wherein the agent selector and Al agents are configured to operate in a distributed architecture across cloud or edge devices to support real-time responsiveness and scalability.Additional Implementation Details and Embodiments

[0119] Various embodiments of the present disclosure may be a system, a method, and / or a computer program product at any possible technical detail level of integration. The computer program product may include a computer readable storage medium (or mediums) having computer readable program instructions thereon for causing processor to carry out aspects of the present disclosure.

[0120] For example, the functionality described herein may be performed as software instructions are executed by, and / or in response to software instructions being executed by, one or more hardware processors and / or any other suitable computing devices. The software instructions and / or other executable code may be read from a computer readable storage medium (or mediums).

[0121] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0122] These computer readable program instructions may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that theSNTH.013WO PATENT instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart(s) and / or block diagram(s) block or blocks.

[0123] The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks. For example, the instructions may initially be carried on a magnetic disk or solid-state drive of a remote computer. The remote computer may load the instructions and / or modules into its dynamic memory and send the instructions over a telephone, cable, or optical line using a modem. A modem local to a server computing system may receive the data on the telephone / cable / optical line and use a converter device including the appropriate circuitry to place the data on a bus. The bus may carry the data to a memory, from which a processor may retrieve and execute the instructions. The instructions received by the memory may optionally be stored on a storage device (e.g., a solid-state drive) either before or after execution by the computer processor.

[0124] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the Figures. For example, two blocks shown in succession may, in fact, beSNTH.013WO PATENT executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. In addition, certain blocks may be omitted in some implementations. The methods and processes described herein are also not limited to any particular sequence, and the blocks or states relating thereto can be performed in other sequences that are appropriate.

[0125] As described above, in various embodiments certain functionality may be accessible by a user through a web-based viewer (such as a web browser), or other suitable software program. In such implementations, the user interface may be generated by a server computing system and transmitted to a web browser of the user (e.g., running on the user’s computing system). Alternatively, data (e.g., user interface data) necessary for generating the user interface may be provided by the server computing system to the browser, where the user interface may be generated (e.g., the user interface data may be executed by a browser accessing a web service and may be configured to render the user interfaces based on the user interface data). The user may then interact with the user interface through the web-browser. User interfaces of certain implementations may be accessible through one or more dedicated software applications. In certain embodiments, one or more of the computing devices and / or systems of the disclosure may include mobile computing devices, and user interfaces may be accessible through such mobile computing devices (for example, smartphones and / or tablets).

[0126] Many variations and modifications may be made to the abovedescribed embodiments, the elements of which are to be understood as being among other acceptable examples. All such modifications and variations are intended to be included herein within the scope of this disclosure. The foregoing description details certain embodiments. It will be appreciated, however, that no matter how detailed the foregoing appears in text, the systems and methods can be practiced in many ways. As is also stated above, it should be noted that the use of particular terminology when describing certain features or aspects of the systems and methods should not be taken to imply that the terminology is being re-defined herein to be restricted to including any specific characteristics of the features or aspects of the systems and methods with which that terminology is associated.

[0127] Conditional language, such as, among others, “can,” “could,” “might,” or “may,” unless specifically stated otherwise, or otherwise understood within the contextSNTH.013WO PATENT as used, is generally intended to convey that certain embodiments include, while other embodiments may not include, certain features, elements, and / or steps. Thus, such conditional language is not generally intended to imply that features, elements and / or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without user input or prompting, whether these features, elements and / or steps are included or are to be performed in any particular embodiment.

[0128] The term “substantially” when used in conjunction with the term “real- time” forms a phrase that will be readily understood by a person of ordinary skill in the art. For example, it is readily understood that such language will include speeds in which no or little delay or waiting is discernible, or where such delay is sufficiently short so as not to be disruptive, irritating, or otherwise vexing to a user.

[0129] Conjunctive language such as the phrase “at least one of X, Y, and Z,” or “at least one of X, Y, or Z,” unless specifically stated otherwise, is to be understood with the context as used in general to convey that an item, term, etc. may be either X, Y, or Z, or a combination thereof. For example, the term “or” is used in its inclusive sense (and not in its exclusive sense) so that when used, for example, to connect a list of elements, the term “or” means one, some, or all of the elements in the list. Thus, such conjunctive language is not generally intended to imply that certain embodiments require at least one of X, at least one of Y, and at least one of Z to each be present.

[0130] The term “a” as used herein should be given an inclusive rather than exclusive interpretation. For example, unless specifically noted, the term “a” should not be understood to mean “exactly one” or “one and only one”; instead, the term “a” means “one or more” or “at least one,” whether used in the claims or elsewhere in the specification and regardless of uses of quantifiers such as “at least one,” “one or more,” or “a plurality” elsewhere in the claims or specification.

[0131] The term “comprising” as used herein should be given an inclusive rather than exclusive interpretation. For example, a general purpose computer comprising one or more processors should not be interpreted as excluding other computer components, and may possibly include such components as memory, input / output devices, and / or network interfaces, among others.SNTH.013WO PATENT

[0132] While the above detailed description has shown, described, and pointed out novel features as applied to various embodiments, it may be understood that various omissions, substitutions, and changes in the form and details of the devices or processes illustrated may be made without departing from the spirit of the disclosure. As may be recognized, certain embodiments of the inventions described herein may be embodied within a form that does not provide all of the features and benefits set forth herein, as some features may be used or practiced separately from others. The scope of certain inventions disclosed herein is indicated by the appended claims rather than by the foregoing description. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.

Claims

SNTH.013WO PATENTWHAT IS CLAIMED IS:1 . A computerized image management system, comprising: one or more Al agents configured to perform specific tasks related to medical image management, a configuration module that allows preferences to be set for which Al agents are activated based on various parameters, wherein the parameters include one or more of user parameters, exam parameters, patient parameters, or facility parameters; an agent selector that dynamically selects and manages the Al agents based on real-time user interactions and contextual parameters, continuously updating its internal model to adapt to changing conditions and improve performance over time; and a learning module that enables the system to autonomously learn user preferences and improve customization over time, based on passive user actions and interactions with the system.

2. The system of claim 1 , wherein the agents include perform functions including one or more of: generating radiology-specific summaries, handling RIS PACS admin functions, providing clinical workflow enhancements, performing advanced search, or data retrieval tasks.

3. The system of claim 1 , wherein user-specific parameters are associated with one or more of: a radiologist, technologist, referring physician, laboratory technician, pathologist, or patient.

4. The system of claim 1 , wherein exam-specific parameters are associated with one or more of: an imaging modality (e.g., MRI, CT, ultrasound, X-ray) or laboratory test (e.g., blood test, tissue biopsy, genetic screening).

5. The system of claim 1 , wherein patient-specific parameters are associated with one or more of: age, medical history, known allergies, and other relevant characteristics.

6. The system of claim 1 , wherein facility-specific parameters are associated with one or more of: unique protocols or workflows.

7. The system of claim 1 , further comprising:SNTH.013WO PATENT a constraint module that allows the Al agents to be constrained based on reference material, wherein the reference material is includes one or more of the following: user-specific reference material, such as published guidelines, clinical protocols, or user-defined rules; exam-specific reference material, such as imaging protocols, laboratory test guidelines, or diagnostic criteria; patient-specific reference material, such as personalized treatment plans, medical history records, or genetic information; or facility-specific reference material, such as institutional policies, standard operating procedures, or compliance requirements.

8. The system of claim 1 , wherein the agent selector includes a coordination module configured to manage communication of information between agents.

9. The system of claim 1 , wherein the agent selector maintains a registry of available Al agents, each associated with metadata describing capabilities, input / output requirements, activation conditions, and dependencies.

10. The system of claim 1 , wherein the Al agents include a report assistant agent comprising sub-functions for syntax correction, impression generation, recommendation generation, and diagnostic algorithm execution.11 . The system of claim 11 , wherein the report assistant agent includes a macro function that allows the user to insert hints into the report metadata to guide agent selection.

12. A computerized clinical imaging platform comprising: a user interface module configured to generate one or more user interfaces tailored to a user role, the user interface module comprising: an image viewer configured to display and manipulate medical images, including tools for one or more of: zooming, panning, adjusting brightness and contrast, or annotating images; a report generation interface configured to receive dictated or typed diagnostic findings, recommendations, or conclusions; a report viewer configured to display diagnostic reports and associated image annotations;SNTH.013WO PATENT an Al integration interface configured to present outputs from one or more Al agents, including one or more of: differential diagnoses, abnormality highlights, or evidence-based recommendations; an agent selector configured to: evaluate real-time user interactions and contextual parameters, including one or more of: user role, exam modality, body region, patient history, or system conditions; identify a task requested by the user, either explicitly or implicitly; select and activate one or more Al agents from an agent pool based on the evaluation; provide input data to the selected Al agents, including one or more of: report text, image data, operational parameters, structured metadata, or external data sources; receive and coordinate outputs from the selected Al agents; continuously update an internal model based on new information to adapt agent selection and improve performance over time.

13. The platform of claim 12, further comprising: an agent pool comprising a plurality of Al agents, each configured to perform a specific task related to medical image management, wherein each agent is associated with metadata describing one or more of: capabilities, input / output requirements, activation conditions, or dependencies; one or more data sources, including: an internal data source comprising structured or unstructured clinical data; an external data source comprising third-party information including one or more of: medical guidelines, research articles, or clinical studies; a synthetic data source comprising data generated or processed by the agent pool or external sources; one or more external Al algorithms configured to enhance agent capabilities through one or more of: advanced image processing, machine learning, or natural language processing techniques.

14. The platform of claim 13, further comprising:SNTH.013WO PATENT a large language model (LLM) configured to interact with the agent selector and one or more Al agents to support one or more of: natural language understanding, summarization, recommendation generation, or other languagebased tasks.

15. The platform of claim 14, wherein the agent selector and Al agents are configured to operate in a distributed architecture across cloud or edge devices to support real-time responsiveness and scalability.

Images