Automatic data acquisition and standardization processing platform based on multi-mode ophthalmologic equipment
By employing high-precision clock synchronization and system delay compensation techniques, the timing synchronization problem in multimodal ophthalmic data acquisition was solved, achieving precise data alignment and normalization, and improving the training effect of deep learning models.
Patent Information
- Application Number
- CN202511904437.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-03-20
AI Technical Summary
In existing technologies, the timing synchronization accuracy is low and the system latency is not compensated during multimodal ophthalmic data acquisition, resulting in data distortion and affecting the ground truth basis for deep learning model training.
Employing a high-precision clock synchronization module, a delay error calculation module, an event point synchronization acquisition module, a timing correlation module, and a standardization processing module, the system ensures that multi-modal devices acquire data at the same instant through clock synchronization calibration, system delay compensation, event point identification, and data preprocessing, and performs data alignment and normalization processing.
It improves the timing synchronization accuracy and reliability of multimodal data acquisition, provides high-quality ground truth input, and enhances the training effect and generalization ability of deep learning models.
Smart Images
Figure CN121709190A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to an automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment. Background Technology
[0002] With the deep integration of medical information technology and artificial intelligence in ophthalmology, the use of multimodal data for precise diagnosis of ophthalmic diseases has become an industry trend. Common multimodal ophthalmic data include optical coherence tomography (OCT) images, fundus photographs, visual field test results, and structured clinical parameters. Fusion analysis of these heterogeneous data can significantly improve the accuracy and comprehensiveness of disease diagnosis.
[0003] However, existing technologies primarily rely on manual operation or post-processing algorithms based on data features to correlate data from different modalities. This approach suffers from low temporal accuracy and is susceptible to the effects of minute eye movements, blinking, or device operation delays, leading to inaccurate temporal alignment of the acquired data, introducing data distortion, and impacting the ground truth for deep learning model training. Even with calibration using a unified network time protocol, different brands and models of ophthalmic devices exhibit inherent, systematic differences in processing latency when receiving instructions, processing internally, and transmitting data. Existing solutions lack precise calculation and compensation mechanisms for these systemic latency issues, resulting in discrepancies in the actual timing of data acquisition actions despite simultaneous instruction issuance. Summary of the Invention
[0004] To overcome the above deficiencies, this invention provides an automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment, aiming to solve the technical problems of low timing synchronization accuracy, uncompensated system latency, and insufficient standardization in the existing multimodal ophthalmic data acquisition process.
[0005] This invention provides the following technical solution: an automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment, comprising: A high-precision clock synchronization module is used to establish a central master clock and perform clock synchronization calibration on all multimodal ophthalmic devices and data acquisition gateways connected to the platform; The delay error calculation module is used to send a test trigger signal and receive a response after clock synchronization calibration, so as to calculate and store the system delay compensation matrix of each ophthalmic device and data acquisition gateway relative to a preset reference time point; The event point synchronization acquisition module is used to identify event points in the predefined inspection process, and send a pre-compensation start time instruction to all multimodal ophthalmic devices according to the system delay compensation matrix. The acquisition gateway marks the received data with a unified event ID and a high-precision timestamp. The time-series correlation module is used to identify the synchronized dataset based on the preset time tolerance of the unified event ID and high-precision timestamp; A standardization processing module is used to perform modal geometry alignment and intensity normalization on the synchronous dataset, as well as preprocessing including at least retinal layer flattening; The unified tensor output module is used to encapsulate the standardized synchronous dataset into a set of multiple input tensors for fusion analysis by deep learning network models.
[0006] Preferably, the clock synchronization calibration step includes: The central master clock provides a millisecond-level time reference signal based on the PTP network protocol, and synchronously sends the time reference signal to the ophthalmic equipment and the data acquisition gateway. The ophthalmic equipment and acquisition gateway adjust and lock their local system clocks according to the time reference signal to maintain time consistency with the central master clock.
[0007] Preferably, the calculation steps for the system delay compensation matrix include: Using the central master clock as a reference, the central master clock synchronously sends test trigger commands to all ophthalmic devices and data acquisition gateways. Record the actual delay time for each device from receiving the test trigger command to its data being timestamped by the acquisition gateway; Calculate the delay compensation value for each device relative to the reference baseline, and combine all compensation values to form a system delay compensation matrix.
[0008] Preferably, the step of identifying event points in a predefined inspection process includes: Capture the start command or status change signal issued by the ophthalmic device before it begins acquisition; Capture time-unique physical signals during the operation of ophthalmic equipment, including but not limited to: patient gaze confirmation signal, light source activation signal, or scanning head movement start signal; The time point of the captured start command, status signal, or physical signal is used as a unified event point to trigger subsequent synchronous acquisition actions.
[0009] Preferably, the step of sending the pre-compensation start-up time instruction includes: After identifying the time of the unified event point, obtain the stored system latency compensation matrix; Calculate the pre-compensation start-up time for each ophthalmic device based on the system delay compensation matrix; Send a command containing pre-compensated startup time to each ophthalmic device to ensure that all devices are on a unified timeline of the master clock.
[0010] Preferably, the step of identifying the synchronized dataset includes: All original data items from different modalities with the same unified event ID are grouped into the same candidate dataset; Obtain the high-precision timestamps of all data items in the candidate dataset, select any two data items' timestamps, and determine whether the absolute value of their time difference is within the preset time tolerance. If the time difference of the timestamps of all data items in the candidate dataset is within the preset time tolerance, then the candidate dataset is determined to be a synchronous dataset.
[0011] Preferably, the modal geometry alignment and intensity normalization steps include: All image modalities in the synchronized dataset are resampled to a predetermined and uniform spatial resolution and size; Using preset anatomical markers as a reference, affine transformations are performed on all image modalities to ensure that anatomical structures have consistent spatial relationships in the image tensors of different modalities. Normalize the pixel intensity values of the image modality, scaling the pixel values to a fixed range. Feature scaling is performed on the non-image clinical parameters in the synchronized dataset.
[0012] Preferably, the pretreatment step of retinal layer planarization includes: An automatic segmentation algorithm is performed on optical coherence tomography images in a synchronous dataset to identify and extract the predetermined retinal reference layer interface; Select the reference layer interface as the reference plane, and calculate the vertical offset of each pixel on the reference layer interface relative to the reference plane. Vertical distortion correction is performed on all pixels in the optical coherence tomography image using a vertical offset to eliminate the influence of eyeball curvature.
[0013] Preferably, the step of encapsulating into a set of multiple input tensors includes: The standardized image modal data is converted into a three-dimensional tensor with uniform channel, height, and width dimensions; Standardized non-image clinical parameters are converted into one-dimensional or two-dimensional feature vector tensors; Using the unified event ID as a common index, the three-dimensional tensor and the feature vector tensor are combined and encapsulated into a multi-input tensor set.
[0014] The present invention has the following beneficial effects: 1. This invention, through a time delay error calculation module and a pre-compensation startup mechanism, not only solves the clock drift problem between devices, but also solves the complex system time delay problem, enabling all multimodal devices to start the acquisition action at the same instant, greatly improving the timing synchronization accuracy and reliability of data acquisition.
[0015] 2. The time compensation mechanism adopted in this invention eliminates time delay and timing deviation, so that the collected multimodal data can accurately reflect the patient's real eye state at a certain stable physiological moment, avoiding data distortion caused by traditional post-processing registration, and providing high-quality ground truth input for deep learning models.
[0016] 3. This invention introduces a key retinal layer flattening preprocessing, which eliminates the influence of eyeball curvature on OCT images, allowing the model to focus on learning the pathological features of the disease rather than the anatomical curvature, significantly improving the training effect and generalization ability of the multimodal fusion model. Attached Figure Description
[0017] Figure 1 This is an architecture diagram of the automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment proposed in this invention; Figure 2 This is a flowchart illustrating the implementation of the automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment proposed in this invention. Detailed Implementation
[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Example 1 In a first embodiment of the present invention, the present invention provides an automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment, such as... Figure 1 As shown, it includes: A high-precision clock synchronization module is used to establish a central master clock and perform clock synchronization calibration on all multimodal ophthalmic devices and data acquisition gateways connected to the platform; Preferably, the clock synchronization calibration step includes: The central master clock provides a millisecond-level time reference signal based on the PTP network protocol, and synchronously sends the time reference signal to the ophthalmic equipment and the data acquisition gateway. The ophthalmic equipment and acquisition gateway adjust and lock their local system clocks according to the time reference signal to maintain time consistency with the central master clock.
[0020] Specifically, a dedicated high-precision clock server is deployed within the data center or central control unit. This server is equipped with hardware modules supporting the IEEE 1588 Precision Time Protocol, serving as the PTP master clock for the entire platform. To ensure the absolute accuracy of the master clock, this master clock server connects to a stable and accurate external time source, such as GPS / BeiDou satellite timing signals, via a dedicated interface to synchronize and calibrate its own clock. All network switches, routers, and acquisition gateways on the platform use network devices that support PTP transparent clock functionality to minimize and compensate for latency jitter on the network transmission path.
[0021] The PTP master clock periodically broadcasts PTP synchronization messages and delay request / response messages on a dedicated synchronization network, providing millisecond- or even microsecond-level time reference signals to all ophthalmic devices and acquisition gateways in the network. Each data acquisition gateway located near an ophthalmic device is configured as a PTP slave clock. This gateway exchanges timestamps with the master clock via the PTP protocol, calculates and compensates for network transmission delays in real time, thereby precisely adjusting and locking its local system clock to the master clock time.
[0022] For multimodal ophthalmic devices that support either the PTP or NTP protocols, the platform configures them as a PTP slave clock client or a high-precision NTP client via its standard network communication interface, enabling automatic calibration of the local system clock. For traditional ophthalmic devices that do not directly support the PTP protocol, the data acquisition gateway handles the time synchronization task. When performing data acquisition, the acquisition gateway uses its high-precision local time, which is locked to the master clock, as the final timestamp, thereby indirectly achieving time consistency across all acquired data.
[0023] By deploying a dedicated PTP master clock and performing high-frequency timestamp exchange on a private network, the platform can converge the clock accuracy of all ophthalmic devices and acquisition gateways to the millisecond or microsecond level, providing a reliable time consistency basis for subsequent delay error calculation and event point synchronous acquisition.
[0024] The delay error calculation module is used to send a test trigger signal and receive a response after clock synchronization calibration, so as to calculate and store the system delay compensation matrix of each ophthalmic device and data acquisition gateway relative to a preset reference time point; Preferably, the calculation steps for the system delay compensation matrix include: Using the central master clock as a reference, the central master clock synchronously sends test trigger commands to all ophthalmic devices and data acquisition gateways. Record the actual delay time for each device from receiving the test trigger command to its data being timestamped by the acquisition gateway; Calculate the delay compensation value for each device relative to the reference baseline, and combine all compensation values to form a system delay compensation matrix.
[0025] Specifically, the time of the central master clock, calibrated via the PTP protocol, is used as the reference benchmark for calculating the latency of the entire system. All latency is measured relative to this uniform and highly accurate master clock time. A dedicated latency testing tool is deployed, running on the central control unit and communicating with the central master clock to acquire and mark high-precision time.
[0026] A unified test trigger command is synchronously sent from the central master clock or central control unit to all connected ophthalmic devices and data acquisition gateways. The central control unit records the precise time this command is sent. Upon receiving the command, each device immediately simulates a data acquisition action and sends a test data packet containing an identifier of that action to its corresponding acquisition gateway as quickly as possible. The acquisition gateway, upon receiving the test data packet from the device, marks the data packet with a precise timestamp using its locked, high-precision local time.
[0027] For each device, the actual latency is defined as the total time difference from when the central control sends an instruction to when the acquisition gateway receives and timestamps it. This time difference includes the sum of network transmission latency, device internal processing latency, and acquisition gateway processing latency. The minimum actual latency among all devices is determined, and the latency compensation value for each device is defined as the difference between its total latency and the minimum latency, i.e., the system latency of that device relative to the fastest device. The latency compensation values calculated for all devices are aggregated to form a system latency compensation matrix. This matrix is stored in the central control unit for subsequent pre-compensation start-up time calculations.
[0028] Through this calibration process, the platform accurately quantifies the inherent deviations in the actions performed by each device, laying a reliable calibration foundation for achieving action synchronization at the event point.
[0029] The event point synchronization acquisition module is used to identify event points in the predefined inspection process, and send a pre-compensation start time instruction to all multimodal ophthalmic devices according to the system delay compensation matrix. The acquisition gateway marks the received data with a unified event ID and a high-precision timestamp. Preferably, the step of identifying event points in a predefined inspection process includes: Capture the start command or status change signal issued by the ophthalmic device before it begins acquisition; Capture time-unique physical signals during the operation of ophthalmic equipment, including but not limited to: patient gaze confirmation signal, light source activation signal, or scanning head movement start signal; The time point of the captured start command, status signal, or physical signal is used as a unified event point to trigger subsequent synchronous acquisition actions.
[0030] Specifically, for ophthalmic devices with standardized communication interfaces, such as those supporting DICOM or custom APIs, a communication protocol listener is deployed within the data acquisition gateway. This listener continuously monitors the communication data flow between the device and the operating terminal. When a specific signal closely related to data acquisition is identified, such as an acquisition start command or a device status change signal, the time point of that signal is immediately marked as the master clock time, serving as a logical event point.
[0031] For ophthalmic devices with closed communication interfaces or significant software latency, a hardware capture mechanism is employed. Non-invasive sensors or signal acquisition modules are deployed on critical operating parts or signal transmission lines of the device. Physical moments highly consistent with the patient's physiological state or the device's optical actions are identified as event points and captured. For example, an opto-isolator connected to the patient's gaze cursor control circuit is used to capture low-level or high-level transition signals indicating the patient's gaze position; a photosensor is used to monitor the flash or activation moment of the fundus camera or OCT light source, and the start time of this light signal is used as the event point; a current or voltage sensor is used to monitor the drive current of the OCT scanning head or fundus camera shutter mechanism, and the starting point of the current change is used as the event point.
[0032] The platform prioritizes the captured event signals, typically using physical signals as the highest priority. Ultimately, the precise time points on the master clock timeline of the captured start commands, state change signals, or physical signals are determined as unified event points. These time points are then sent to the event point synchronization acquisition module for calculating and sending pre-compensated start time commands.
[0033] Through a combination of hardware and software mechanisms, the platform ensures that it can capture the "event point" that reflects the start of the acquisition action with extremely high accuracy, regardless of the device type, providing a triggering benchmark for cross-modal synchronous acquisition.
[0034] Preferably, the step of sending the pre-compensation start-up time command includes: After identifying the time of the unified event point, obtain the stored system latency compensation matrix; Calculate the pre-compensation start-up time for each ophthalmic device based on the system delay compensation matrix; Send a command containing pre-compensated startup time to each ophthalmic device to ensure that all devices are on a unified timeline of the master clock.
[0035] Specifically, when the event point synchronization acquisition module successfully identifies and determines the instant of occurrence of the unified event that triggered the acquisition, its time is precisely marked as the unified event point time. The module immediately retrieves the previously calibrated and stored system latency compensation matrix from the memory, which contains the latency compensation value of each device relative to the fastest device in the system.
[0036] For each ophthalmic device connected to the platform, the system calculates its unique pre-compensated startup time by subtracting the corresponding latency compensation value from the uniform event point time. Since the latency compensation value represents the inherent latency of the device, performing the subtraction operation is actually to send instructions to slower devices in advance to offset the system latency of those devices.
[0037] The event point synchronization acquisition module encapsulates a dedicated start acquisition command for each device, containing a calculated pre-compensation start time. The central control unit sends this command to the corresponding ophthalmic device. Upon receiving the command, the device does not immediately begin acquisition but continuously monitors the high-precision master clock. Only when the master clock reaches the pre-compensation start time specified in the command will the device immediately begin data acquisition.
[0038] Through this pre-compensation mechanism, the system latency differences of all ophthalmic devices are eliminated, thereby ensuring that all devices can start their physical acquisition actions at the same instant on the unified timeline of the master clock, which greatly improves the timing synchronization accuracy of multimodal data.
[0039] The time-series correlation module is used to identify the synchronized dataset based on the preset time tolerance of the unified event ID and high-precision timestamp; Preferably, the step of identifying the synchronized dataset includes: All original data items from different modalities with the same unified event ID are grouped into the same candidate dataset; Obtain the high-precision timestamps of all data items in the candidate dataset, select any two data items' timestamps, and determine whether the absolute value of their time difference is within the preset time tolerance. If the time difference of the timestamps of all data items in the candidate dataset is within the preset time tolerance, then the candidate dataset is determined to be a synchronous dataset.
[0040] Specifically, after the event point synchronization acquisition module triggers and completes the acquisition action, raw data from different modalities, such as OCT image data, fundus photograph data, and visual field examination data, will be successively transmitted to the data storage server. All data carries two key characteristics: a unified event ID and a high-precision timestamp. The time-series correlation module first scans the data repository and logically merges all raw data items with the same unified event ID into a candidate dataset.
[0041] The module extracts the corresponding high-precision timestamp from each data item in the candidate dataset. The platform pre-sets a preset time tolerance, which is determined based on the system's clock synchronization accuracy and the residual error after delay compensation, typically on the order of microseconds or sub-microseconds. Then, it arbitrarily selects two different data items from the candidate dataset and calculates the absolute time difference between their respective timestamps.
[0042] The module determines whether the absolute time difference meets the condition of being less than or equal to a preset time tolerance. If the absolute value of the time difference calculated from all possible combinations of data items in the candidate dataset is within the preset time tolerance range, it proves that all data in the candidate dataset has been collected through strict time-series synchronization. At this point, the time-series correlation module formally determines the candidate dataset as a synchronized dataset.
[0043] Through this logical judgment process, the platform ensures that only data that meets extremely high time consistency requirements can be recognized as valid multimodal fusion input data, thereby guaranteeing the truth basis for model training.
[0044] A standardization processing module is used to perform modal geometry alignment and intensity normalization on the synchronous dataset, as well as preprocessing including at least retinal layer flattening; Preferably, the modal geometry alignment and intensity normalization steps include: All image modalities in the synchronized dataset are resampled to a predetermined and uniform spatial resolution and size; Using preset anatomical markers as a reference, affine transformations are performed on all image modalities to ensure that anatomical structures have consistent spatial relationships in the image tensors of different modalities. Normalize the pixel intensity values of the image modality, scaling the pixel values to a fixed range. Feature scaling is performed on the non-image clinical parameters in the synchronized dataset.
[0045] Specifically, the platform determines a unified target spatial resolution and image size based on the input requirements of a predetermined deep learning network model. Resampling is performed on all image modalities in the synchronized dataset to adjust them to the unified spatial resolution and size. A pre-trained anatomical structure detection algorithm is used to automatically identify key anatomical landmarks in each image, such as the optic disc center, macular center, or fovea. One of these anatomical landmarks is selected as a reference point. The translation vector, rotation angle, and necessary scaling factor of this reference point relative to the center point of the target tensor are calculated for each image. The calculated affine transformation matrix is applied to all image modalities to ensure that key anatomical structures such as the optic disc and macula occupy consistent and predetermined spatial positions in the image tensors of different modalities in the final output tensor.
[0046] The minimum and maximum pixel intensities of all image modalities in the entire training dataset are calculated. Min-Max normalization is then performed on all image pixel values in the synchronous dataset, linearly scaling the pixel intensity values to [0,1] to eliminate pixel intensity differences caused by different devices, lighting conditions, and gain settings, enabling the model to stably learn the structural features of the images. Structured non-image clinical parameters, such as patient age, intraocular pressure, and visual acuity, are extracted from the synchronous dataset. Feature scaling is then performed on these parameters, for example, scaling them to the range of [0,1] to ensure that these numerical features do not dominate feature weights due to excessively large values in the input of the deep learning model, maintaining fairness among features.
[0047] Through the aforementioned rigorous geometric alignment and intensity normalization steps, the platform ensures that multimodal data is unified in both spatial and numerical scales before being input into the deep learning network, greatly improving the training efficiency of the model and the accuracy of fusion diagnosis.
[0048] Preferably, the pretreatment step of retinal layer planarization includes: An automatic segmentation algorithm is performed on optical coherence tomography images in a synchronous dataset to identify and extract the predetermined retinal reference layer interface; Select the reference layer interface as the reference plane, and calculate the vertical offset of each pixel on the reference layer interface relative to the reference plane. Vertical distortion correction is performed on all pixels in the optical coherence tomography image using a vertical offset to eliminate the influence of eyeball curvature.
[0049] Specifically, two-dimensional OCT images from the synchronized dataset are input into the preprocessing module. First, basic noise reduction and contrast enhancement are performed. Then, an automatic segmentation algorithm is executed using a deep learning segmentation model. A predetermined layer interface with clinical stability and representativeness is selected from the segmentation results as a reference layer interface. Preferably, this reference layer interface is the retinal pigment epithelium or the vitreoretinal interface, and it is extracted as a pixel coordinate curve.
[0050] The average height of the reference layer interface in the vertical direction is defined as the virtual reference plane. The vertical offset of the vertical coordinates at each horizontal position on the reference layer interface relative to the virtual reference plane is calculated. This offset describes the change in retinal curvature in the horizontal direction.
[0051] Vertical deformation correction is performed on all pixels in the OCT image, resulting in new vertical coordinates. By performing this operation on all pixels, the entire retinal layer structure is shifted vertically, ultimately aligning the reference layer interface with a horizontal virtual reference plane.
[0052] This process effectively eliminates the effects of eyeball curvature and scanning tilt in the original OCT images, making the thickness of the retinal layer and the vertical position of the lesion highly spatially comparable in different patients or different scans.
[0053] The unified tensor output module is used to encapsulate the standardized synchronous dataset into a set of multiple input tensors for fusion analysis by deep learning network models.
[0054] Preferably, the step of encapsulating into a set of multiple input tensors includes: The standardized image modal data is converted into a three-dimensional tensor with uniform channel, height, and width dimensions; Standardized non-image clinical parameters are converted into one-dimensional or two-dimensional feature vector tensors; Using the unified event ID as a common index, the three-dimensional tensor and the feature vector tensor are combined and encapsulated into a multi-input tensor set.
[0055] Specifically, all geometrically aligned and intensity-normalized image modal data is converted from the traditional pixel matrix format to a three-dimensional tensor format. All image tensors are forcibly converted to a uniform dimensional format, typically: For example, the tensor size of a color fundus photograph might be... The tensor size of a grayscale OCT image may be... .
[0056] Standardized non-image-based clinical parameters, such as scaled age, intraocular pressure, and gender, are encoded into machine-readable numerical forms. For categorical features, one-hot encoding is used. All encoded numerical features are arranged in a predetermined order and assembled into a one-dimensional or two-dimensional feature vector tensor. For example, if there are a total of If there are 1 clinical feature, then the feature vector tensor size is 1. .
[0057] A unified event ID, common to the dataset, is used as the core index. This event ID ensures that all image tensors and feature vector tensors correspond to each other in time and semantics. The unified tensor output module combines and encapsulates all transformed and aligned tensors into a single tensor set. The aligned tensors include fundus image tensors, OCT image tensors, visual field tensors, clinical feature vector tensors, etc. This tensor set is output through a predetermined data interface.
[0058] Through this encapsulation process, the platform completes the transformation from raw, heterogeneous ophthalmic device data to a set of aligned, multi-path input tensors that are directly usable by the model, enabling data-driven multimodal fusion analysis.
[0059] Example 2 In the follow-up of wet macular degeneration, lesions such as exudate or neovascularization are dynamically changing, requiring doctors or AI to precisely align OCT and fundus photography data acquired at the same time. However, existing acquisition systems lack high-precision clock synchronization and system latency compensation mechanisms, resulting in microsecond-level or even millisecond-level temporal discrepancies between images captured by different devices. This discrepancy is sufficient to cause minute eye movements in patients, making it impossible to achieve precise spatial registration between multimodal images. Ultimately, this leads to AI models learning incorrect lesion associations and affects the accurate judgment of changes in lesion volume and location.
[0060] To address the aforementioned problems, this invention employs an automated data acquisition and standardized processing platform based on multimodal ophthalmic equipment. The platform's workflow is as follows: Figure 2 As shown, it includes: Establish a central master clock to synchronize and calibrate the clocks of all multimodal ophthalmic devices and data acquisition gateways connected to the platform; After clock synchronization calibration, a test trigger signal is sent and a response is received to calculate and store the system delay compensation matrix of each ophthalmic device and data acquisition gateway relative to a preset reference time point; The system identifies event points in the predefined inspection process and sends a pre-compensation start time instruction to all multimodal ophthalmic devices according to the system delay compensation matrix. The acquisition gateway then marks the received data with a unified event ID and a high-precision timestamp. Based on the preset time tolerance of the unified event ID and high-precision timestamp, and through the preset time tolerance, the synchronized dataset is identified; Modal geometry alignment and intensity normalization are performed on the synchronous dataset, as well as preprocessing including at least retinal layer flattening; The standardized synchronous dataset is encapsulated into a set of multi-path input tensors for fusion analysis by deep learning network models.
[0061] This platform ensures action synchronization of data from different modalities with microsecond-level precision, greatly improving the reliability of the truth value of the synchronized dataset. Simultaneously, combined with refined standardization processes such as retinal layer flattening, this invention constructs a multi-path input tensor set that is highly consistent in both space and time for ophthalmic AI models, significantly improving the accuracy, robustness, and automation level of deep learning networks in the diagnosis and follow-up of dynamic lesions such as wet macular degeneration.
[0062] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A data automatic acquisition and standardized processing platform based on multimodal ophthalmic equipment, characterized in that: include: A high-precision clock synchronization module is used to establish a central master clock and perform clock synchronization calibration on all multimodal ophthalmic devices and data acquisition gateways connected to the platform; The delay error calculation module is used to send a test trigger signal and receive a response after clock synchronization calibration, so as to calculate and store the system delay compensation matrix of each ophthalmic device and data acquisition gateway relative to a preset reference time point; The event point synchronization acquisition module is used to identify event points in the predefined inspection process, and send a pre-compensation start time instruction to all multimodal ophthalmic devices according to the system delay compensation matrix. The acquisition gateway marks the received data with a unified event ID and a high-precision timestamp. The time-series correlation module is used to identify the synchronized dataset based on the preset time tolerance of the unified event ID and high-precision timestamp; A standardization processing module is used to perform modal geometry alignment and intensity normalization on the synchronous dataset, as well as preprocessing including at least retinal layer flattening; The unified tensor output module is used to encapsulate the standardized synchronous dataset into a set of multiple input tensors for fusion analysis by deep learning network models.
2. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The clock synchronization calibration steps include: The central master clock provides a millisecond-level time reference signal based on the PTP network protocol, and synchronously sends the time reference signal to the ophthalmic equipment and the data acquisition gateway. The ophthalmic equipment and acquisition gateway adjust and lock their local system clocks according to the time reference signal to maintain time consistency with the central master clock.
3. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The calculation steps for the system delay compensation matrix include: Using the central master clock as a reference, the central master clock synchronously sends test trigger commands to all ophthalmic devices and data acquisition gateways. Record the actual delay time for each device from receiving the test trigger command to its data being timestamped by the acquisition gateway; Calculate the delay compensation value for each device relative to the reference baseline, and combine all compensation values to form a system delay compensation matrix.
4. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The steps for identifying event points in a predefined inspection process include: Capture the start command or status change signal issued by the ophthalmic device before it begins acquisition; Capture time-unique physical signals during the operation of ophthalmic equipment, including but not limited to: patient gaze confirmation signal, light source activation signal, or scanning head movement start signal; The time point of the captured start command, status signal, or physical signal is used as a unified event point to trigger subsequent synchronous acquisition actions.
5. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The steps for sending the pre-compensation start-up time command include: After identifying the time of the unified event point, obtain the stored system latency compensation matrix; Calculate the pre-compensation start-up time for each ophthalmic device based on the system delay compensation matrix; Send a command containing pre-compensated startup time to each ophthalmic device to ensure that all devices are on a unified timeline of the master clock.
6. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The steps for identifying synchronized datasets include: All original data items from different modalities with the same unified event ID are grouped into the same candidate dataset; Obtain the high-precision timestamps of all data items in the candidate dataset, select any two data items' timestamps, and determine whether the absolute value of their time difference is within the preset time tolerance. If the time difference of the timestamps of all data items in the candidate dataset is within the preset time tolerance, then the candidate dataset is determined to be a synchronous dataset.
7. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The modal geometry alignment and intensity normalization steps include: All image modalities in the synchronized dataset are resampled to a predetermined and uniform spatial resolution and size; Using preset anatomical markers as a reference, affine transformations are performed on all image modalities to ensure that anatomical structures have consistent spatial relationships in the image tensors of different modalities. Normalize the pixel intensity values of the image modality, scaling the pixel values to a fixed range. Feature scaling is performed on the non-image clinical parameters in the synchronized dataset.
8. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The pretreatment steps for retinal layer flattening include: An automatic segmentation algorithm is performed on optical coherence tomography images in a synchronous dataset to identify and extract the predetermined retinal reference layer interface; Select the reference layer interface as the reference plane, and calculate the vertical offset of each pixel on the reference layer interface relative to the reference plane. Vertical distortion correction is performed on all pixels in the optical coherence tomography image using a vertical offset to eliminate the influence of eyeball curvature.
9. The automatic data acquisition and standardized processing platform based on multimodal ophthalmic equipment according to claim 1, characterized in that, The steps to encapsulate it into a set of multiple input tensors include: The standardized image modal data is converted into a three-dimensional tensor with uniform channel, height, and width dimensions; Standardized non-image clinical parameters are converted into one-dimensional or two-dimensional feature vector tensors; Using the unified event ID as a common index, the three-dimensional tensor and the feature vector tensor are combined and encapsulated into a multi-input tensor set.