Medical image dynamic labeling process recording system and method based on doctor labeled behavior semantic understanding

Abstract

The application relates to the technical field of medical care information science and discloses a medical image dynamic labeling process recording system and method based on doctor labeling behavior semantic understanding, which comprises the following steps: a labeling behavior tracking module acquires view behavior data and undo redo data in an interactive editing interface; a dynamic data analysis module utilizes an operator to generate a high-attention area label, and according to the spatial coincidence degree of a state machine migration sequence, a correction coefficient is calculated to down-regulate a mask confidence parameter, so as to calibrate ambiguous samples; and a state recording output module encapsulates the high-attention area label and low-confidence ambiguous samples into a behavior trajectory sequence and outputs the behavior trajectory sequence. The application converts a behavior space-time trajectory into a specific supervision signal, eliminates sample topological ambiguity in a labeling process, and establishes a closed-loop data flow path with feedback from data perception to model optimization.

Description

Technical Field

[0001] This invention belongs to the field of healthcare informatics technology, and in particular relates to a system and method for recording the dynamic annotation process of medical images based on semantic understanding of doctors' annotation behavior. Background Technology

[0002] Current lesion annotation systems mostly adopt interactive editing methods. Clinicians manually adjust the mask boundaries to correct errors based on the pre-segmented contours generated by the model. This has become the mainstream technique for obtaining high-quality supervision samples in clinical practice. However, with the increasing demands for refined lesion segmentation in clinical diagnosis, the inherent logical gaps and information loss problems of the above interactive editing method have become increasingly prominent when dealing with complex anatomical structures. When performing annotation tasks, clinicians often make real-time judgments and error correction decisions on the anatomical morphology of lesions. For example, when it is found that the model has mistakenly connected the tumor to the adjacent blood vessels, the clinician performs precise topological stripping by repeatedly undoing, redoing, and outlining the contours. The existing recording mechanism only stores the final pixel distribution after the interaction as the result in the database. The system discards the spatial topological decisions, the distribution of the area of ​​interest, and the sequence of operational intentions released by the expert during the error correction process. The downstream segmentation network can only obtain static geometric supervision and cannot perceive the process logic of the expert's identification of errors and implementation of anatomical reconstruction from the data level. As a result, the model is unable to effectively improve its fitting ability to the edges of complex lesions.

[0003] The static imaging interactive interface of the terminal has limitations in recording, and the control software algorithm during the interactive operation also has the defect of discarding process features. For example, Chinese invention patent application with publication number CN112396606A discloses a medical image segmentation method, system and device based on user interaction, which introduces a multi-iteration interactive correction mechanism. However, the input model is only the discrete modification results of the boundary by the user. In essence, it still relies on the final static segmentation mask to complete the model parameter update. This mode implicitly relies on the ideal working condition that each modification by the user is absolutely accurate. When facing complex anatomical variation areas in clinical practice and nonlinear trajectory deviations caused by random muscle tremors or cognitive hesitation of physicians, due to the lack of deep capture of the spatiotemporal context of interactive behavior, it is impossible to identify and filter out random noise at the physical operation level, and it is also impossible to quantify the ambiguous intervals of difficult samples. This makes it difficult for the downstream network to capture the procedural logic of clinical experts' identification errors and implement anatomical reconstruction.

[0004] Therefore, the technical problem to be solved by this invention is how to extract structured error correction trajectories with clinical semantics from the spatiotemporal sequence of topological evolution and behavioral interaction in the mask space, and use them as high-order supervision signals to construct an interactive learning system with anatomical cognitive closed loop. Summary of the Invention

[0005] This invention aims to address the problem that existing interactive annotation methods lack perception of underlying clinical diagnostic error correction logic, and that the mapping between annotation behavior trajectory and high-order anatomical semantics is discontinuous, making it difficult for downstream segmentation models to obtain procedural structured supervision information.

[0006] In this technical solution, a medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior includes:

[0007] The image display module is used to display medical images to be annotated and provide an interactive editing interface; The annotation behavior tracking module is equipped with a data acquisition interface that interfaces with the image display module to acquire non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface; The dynamic data analysis module is used to receive non-editable view behavior data and continuous undo / redo operation data, and includes the following actions during the calculation: Step S101, inputting the non-editable view behavior data into the visual attention weight operator to generate high attention region labels; Step S102, establishing the state machine transition sequence corresponding to the continuous undo / redo operation data; Step S103, when the state machine transition sequence meets the overlap threshold in spatial coordinates, calculating the duration correction coefficient to lower the confidence parameter of the corresponding mask, thereby calibrating low-confidence ambiguous samples; The status recording output module, connected to the dynamic data analysis module, is used to receive high-attention region labels and low-confidence ambiguous samples output by the dynamic data analysis module, encapsulate the high-attention region labels and low-confidence ambiguous samples into a structured behavior trajectory sequence according to the time stamp sequence relationship, and output the structured behavior trajectory sequence to the retraining pipeline of the deep learning model.

[0008] Preferably, the non-editable view behavior data includes: the scaling ratio parameter of the image display module, the translation speed vector, and the absolute dwell time of the mouse cursor in the local anatomical region; step S101 includes the following sub-steps: step S1011, when the scaling ratio parameter is greater than the preset scaling reference and the absolute dwell time exceeds the preset time threshold, a high attention area label is generated; step S1012, the high attention area label is bound to each labeled trajectory node in the structured behavior trajectory sequence using a data structure containing the center coordinates and the boundary range.

[0009] Preferably, the continuous undo / redo operation data is converted into a state machine transition sequence by converting continuous editing and undo / redo actions; in step S103, the state machine transition sequence satisfies the overlap threshold in spatial coordinates by detecting that the overlap of the discrete time sequence consisting of editing operation, continuous pause, undo operation and new editing operation in spatial coordinates is greater than a preset overlap ratio threshold.

[0010] Preferably, the structured behavior trajectory sequence encapsulated by the state recording output module includes: connection modification semantic labels, boundary modification semantic labels, short-term confirmation intent data, high attention spatial range data, and low confidence uncertainty index; the state recording output module converts the structured behavior trajectory sequence into a specific supervision signal, so that the downstream segmentation network in the retraining pipeline can update the weight parameters for topology misclassification data, structural adhesion region coordinates, and boundary correction features.

[0011] Preferably, after the dynamic data analysis module completes step S103, it calculates the latency loss of context switching in the retraining pipeline of the deep learning model, calculates the difference compensation factor based on the latency loss, and uses the difference compensation factor to correct the difference of timestamps in the structured behavior trajectory sequence.

[0012] Preferably, the medical images are medical image data containing samples of difficult lesions; the interactive editing interface corrects the topological region by manually adjusting the mask boundary based on the pre-segmented contour generated by the model; low-confidence ambiguous samples are input into the retraining pipeline to calibrate the inherent weak areas of the deep learning model and to isolate samples of difficult lesions.

[0013] Preferably, when the dynamic data analysis module transmits the data stream, it performs feature fusion of the original image signal, doctor interaction time series data, and confidence decline results, so that each behavioral event in the structured behavioral trajectory sequence corresponds to a closed-loop data flow path.

[0014] Preferably, the retraining pipeline of the deep learning model receives structured behavioral trajectory sequences to fine-tune the weight coefficients of the downstream lesion segmentation network, and converts high-interest region labels and low-confidence ambiguous samples into specific supervision signals, which are then output to the downstream lesion segmentation network.

[0015] Preferably, the system also includes a model update evaluation module; the input of the model update evaluation module is connected to the output of the state recording output module, and is used to obtain the structured behavior trajectory sequence and calculate the lesion segmentation accuracy of the deep learning model after incremental training. When the lesion segmentation accuracy reaches the preset performance index, the updated deep learning model is sent to the image display module to update the pre-segmented contour in the interactive editing interface.

[0016] A method for recording the dynamic annotation process of medical images based on semantic understanding of doctor annotation behavior, which runs in a system for recording the dynamic annotation process of medical images based on semantic understanding of doctor annotation behavior, includes the following steps: Step S101: Display the medical image to be annotated through the image display module and provide an interactive editing interface; Step S102: Obtain non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface through the data acquisition interface of the annotation behavior tracking module; Step S103: Input non-editable view behavior data into the visual attention weight operator through the dynamic data analysis module to generate high attention area labels, and establish a state machine transition sequence corresponding to continuous undo / redo operation data. When the state machine transition sequence meets the overlap threshold in spatial coordinates, calculate the duration correction coefficient to lower the confidence parameter of the corresponding mask, thereby calibrating low-confidence ambiguous samples. Step S104: Receive high-attention region labels and low-confidence ambiguous samples output by the dynamic data analysis module through the state recording output module, encapsulate the high-attention region labels and low-confidence ambiguous samples into a structured behavior trajectory sequence according to the time stamp sequence relationship, and output the structured behavior trajectory sequence to the retraining pipeline of the deep learning model.

[0017] Compared with existing technologies, the medical image dynamic annotation process recording system and method based on semantic understanding of doctor annotation behavior has the following advantages: 1. In the dynamic annotation process of medical images, by constructing the initial medical image mask topology and the target medical image mask topology, the number of independent connected components and the number of internal cavities are extracted. The changes in the number of connected components and the number of cavities are calculated. When the changes exceed the threshold, a first state arbitration signal is issued. When the changes are zero and the changes in the boundary curvature tensor are within the set threshold, a second state arbitration signal is issued. The mouse trajectory is classified into connection modification semantics and boundary modification semantics. This breaks away from the general pixel matrix addition and deletion method and translates the messy physical trajectory into a structure-level data stream that reflects anatomical cognition, providing high-order supervision information with clear structural topological relationships for the downstream segmentation network.

[0018] 2. Utilizing the first and second state arbitration signals, the user interface elements are adaptively rendered. When the first state arbitration signal, representing a topological change, is received, the display level of the two topological error correction factors, mislabeling and missing labeling, in the feedback window is increased. When the second state arbitration signal, representing edge fine-tuning, is received, the feedback window options are reconstructed and the display weight of the two geometric feature factors, boundary fine-tuning and contour calibration, is increased. This transforms the topological change of the underlying mask into interactive feedforward control, enabling transient prediction of clinical correction intentions. Under conditions of maintaining low operational latency, high-purity explicit supervision signals are collected to ensure efficient collaboration during the operation of the medical system.

[0019] 3. By periodically reading the scaling parameters, translation speed, and cursor dwell time in local areas of the image display module, a visual attention weight operator is input to calibrate high-attention areas. Continuous undo and redo operations are captured simultaneously, and a state machine transition sequence is constructed. When a discrete sequence consisting of edit, dwell, undo, and new edit is detected at a spatially overlapping position, a duration hedging coefficient is calculated to lower the confidence parameter of the corresponding mask. Low-confidence ambiguous samples are calibrated, and non-ideal deviations in the interaction process are actively quantified. Under the constraint of no additional external measurement equipment, the inherent weak areas of the segmentation model are identified in reverse, improving the automatic screening efficiency of difficult lesion samples. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of the medical image dynamic annotation process recording system of the present invention; Figure 2 This is a hardware structure diagram of the medical image dynamic annotation process recording system of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0022] A medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior includes: The image display module is used to display medical images to be annotated and provide an interactive editing interface; The annotation behavior tracking module is equipped with a data acquisition interface that interfaces with the image display module to acquire non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface; The dynamic data analysis module is used to receive non-editable view behavior data and continuous undo / redo operation data, and includes the following actions during the calculation: Step S101, inputting the non-editable view behavior data into the visual attention weight operator to generate high attention region labels; Step S102, establishing the state machine transition sequence corresponding to the continuous undo / redo operation data; Step S103, when the state machine transition sequence meets the overlap threshold in spatial coordinates, calculating the duration correction coefficient to lower the confidence parameter of the corresponding mask, thereby calibrating low-confidence ambiguous samples; The status recording output module, connected to the dynamic data analysis module, is used to receive high-attention region labels and low-confidence ambiguous samples output by the dynamic data analysis module, encapsulate the high-attention region labels and low-confidence ambiguous samples into a structured behavior trajectory sequence according to the time stamp sequence relationship, and output the structured behavior trajectory sequence to the retraining pipeline of the deep learning model.

[0023] Preferably, the non-editable view behavior data includes: the scaling ratio parameter of the image display module, the translation speed vector, and the absolute dwell time of the mouse cursor in the local anatomical region; step S101 includes the following sub-steps: step S1011, when the scaling ratio parameter is greater than the preset scaling reference and the absolute dwell time exceeds the preset time threshold, a high attention area label is generated; step S1012, the high attention area label is bound to each labeled trajectory node in the structured behavior trajectory sequence using a data structure containing the center coordinates and the boundary range.

[0024] Preferably, the continuous undo / redo operation data is converted into a state machine transition sequence by converting continuous editing and undo / redo actions; in step S103, the state machine transition sequence satisfies the overlap threshold in spatial coordinates by detecting that the overlap of the discrete time sequence consisting of editing operation, continuous pause, undo operation and new editing operation in spatial coordinates is greater than a preset overlap ratio threshold.

[0025] Preferably, the structured behavior trajectory sequence encapsulated by the state recording output module includes: connection modification semantic labels, boundary modification semantic labels, short-term confirmation intent data, high attention spatial range data, and low confidence uncertainty index; the state recording output module converts the structured behavior trajectory sequence into a specific supervision signal, so that the downstream segmentation network in the retraining pipeline can update the weight parameters for topology misclassification data, structural adhesion region coordinates, and boundary correction features.

[0026] Preferably, after the dynamic data analysis module completes step S103, it calculates the latency loss of context switching in the retraining pipeline of the deep learning model, calculates the difference compensation factor based on the latency loss, and uses the difference compensation factor to correct the difference of timestamps in the structured behavior trajectory sequence.

[0027] Preferably, the medical images are medical image data containing samples of difficult lesions; the interactive editing interface corrects the topological region by manually adjusting the mask boundary based on the pre-segmented contour generated by the model; low-confidence ambiguous samples are input into the retraining pipeline to calibrate the inherent weak areas of the deep learning model and to isolate samples of difficult lesions.

[0028] Preferably, when the dynamic data analysis module transmits the data stream, it performs feature fusion of the original image signal, doctor interaction time series data, and confidence decline results, so that each behavioral event in the structured behavioral trajectory sequence corresponds to a closed-loop data flow path.

[0029] Preferably, the retraining pipeline of the deep learning model receives structured behavioral trajectory sequences to fine-tune the weight coefficients of the downstream lesion segmentation network, and converts high-interest region labels and low-confidence ambiguous samples into specific supervision signals, which are then output to the downstream lesion segmentation network.

[0030] Preferably, the system also includes a model update evaluation module; the input of the model update evaluation module is connected to the output of the state recording output module, and is used to obtain the structured behavior trajectory sequence and calculate the lesion segmentation accuracy of the deep learning model after incremental training. When the lesion segmentation accuracy reaches the preset performance index, the updated deep learning model is sent to the image display module to update the pre-segmented contour in the interactive editing interface.

[0031] A method for recording the dynamic annotation process of medical images based on semantic understanding of doctor annotation behavior, which runs in a system for recording the dynamic annotation process of medical images based on semantic understanding of doctor annotation behavior, includes the following steps: Step S101: Display the medical image to be annotated through the image display module and provide an interactive editing interface; Step S102: Obtain non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface through the data acquisition interface of the annotation behavior tracking module; Step S103: Input non-editable view behavior data into the visual attention weight operator through the dynamic data analysis module to generate high attention area labels, and establish a state machine transition sequence corresponding to continuous undo / redo operation data. When the state machine transition sequence meets the overlap threshold in spatial coordinates, calculate the duration correction coefficient to lower the confidence parameter of the corresponding mask, thereby calibrating low-confidence ambiguous samples. Step S104: Receive high-attention region labels and low-confidence ambiguous samples output by the dynamic data analysis module through the state recording output module, encapsulate the high-attention region labels and low-confidence ambiguous samples into a structured behavior trajectory sequence according to the time stamp sequence relationship, and output the structured behavior trajectory sequence to the retraining pipeline of the deep learning model.

[0032] Example 1: In the case of interactive edge correction of tumor and adjacent blood vessel adhesion areas assisted by medical image segmentation system, due to the anatomical topological defects of the pre-segmented contour, the traditional recording method only saves the final mask pixel distribution after the interaction, which causes the spatial topological data, the distribution of the region of interest and the operation sequence to be broken during the interaction process. As a result, the downstream lesion segmentation network lacks process structured supervision information and cannot form a complete dataset for updating network weights.

[0033] When the image display module provides an interactive editing interface and the data acquisition interface captures a continuous interoperability command stream acting on the current medical image mask area, the annotation behavior tracking module acquires non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface. The dynamic data analysis module constructs the static image spatial morphology of the current time step into an initial medical image mask topology structure composed of spatial nodes and connected edges. When a single editing operation containing tool coordinates, scope of action, and tool type is completed, the modified target medical image mask topology structure is generated in real time. The topological invariant analytical operator is used to calculate the number of independent connected regions and the number of internal cavities of the initial medical image mask topology structure and the target medical image mask topology structure, respectively.

[0034] The dynamic data analysis module calculates the changes in the number of connected components and cavities. When the absolute value of at least one parameter in the changes in the number of connected components and cavities is greater than or equal to 1, the connectivity of the anatomical structure corresponding to a single editing operation undergoes topological splitting or critical path fusion. The data processing unit then issues a first-state arbitration signal and classifies it as a connection modification structured semantic. When both the changes in the number of connected components and cavities are equal to 0, the intrinsic topology of the anatomical structure corresponding to a single editing operation remains unchanged. The data processing unit calculates the Hausdorff distance feature between the initial mask boundary and the target mask boundary, as well as the change in the contour boundary curvature tensor, taking into account... The change in the curvature tensor of the contour boundary is mathematically and physically represented as a multi-dimensional high-order tensor matrix, while the preset spatial variation matrix threshold is represented as an isolated scalar value in engineering. To eliminate the dimensionality mismatch defect that exists when directly comparing the two, this invention introduces a clear dimensionality reduction mapping rule. Specifically, after obtaining the change matrix of the contour boundary curvature tensor, the data processing unit performs Frobenius norm calculation on it. That is, it first squares the feature elements at each position in the tensor matrix, sums up all the squared values, and finally takes the square root of the sum, thereby reducing the dimensionality of the multi-dimensional tensor matrix. The quantity matrix is ​​fully compressed and mapped to a single scalar score representing the severity of the overall curvature variation of the boundary. Through this dimensionality reduction mapping rule, the system achieves absolute unification of the dimensionality and logical dimension. This norm scalar score is directly compared with a predetermined spatial variation matrix threshold of 0.45. When the change in the contour boundary curvature tensor is less than the preset spatial variation matrix threshold, a second-state arbitration signal is issued and categorized as boundary modification structured semantics. This is used to adjust the rendering weight of the modification intent feedback window. Upon receiving the first-state arbitration signal, the display level of the mislabeling and missing label completion options is increased. Upon receiving the second-state arbitration signal, the window options... The geometric feature factors for boundary fine-tuning and contour calibration are adjusted to receive confirmation operations to collect intent supervision signals. Specifically, the rendering weight adjustment and adaptive update of the modified intent feedback window are executed by the UI control thread. When the data processing unit issues the first state arbitration signal indicating a sudden change in topological relationship, the UI control thread blocks the current rendering queue, dynamically reconstructs the menu of the modified intent feedback window, forcibly adjusts the vertical display coordinates of the two topology correction options, namely mislabeling disconnection and missing label completion, to the top layer of the pop-up menu, and increases their visual rendering opacity parameter from the default 60% to 100%, while allocating 70% of the system focus response weight.

[0035] Upon receiving the second-state arbitration signal indicating edge fine-tuning, the UI control thread clears the topology options, elevates the display priority of the two geometric feature factors, boundary fine-tuning and contour calibration, to the center primary position, bolds their rendered font, and assigns them 85% of the click event trigger weight. This control logic seamlessly maps the transient changes in the underlying mask topology to the dynamic rendering weights of the upper-layer UI elements, ensuring high-purity capture of clinical error correction intentions. In the behavioral trajectory sequence, the dynamic data analysis module introduces parameterized correlation analysis for non-editable view behaviors and continuous undo / redo behaviors, periodically reading the scaling parameters, translation velocity vector, and absolute dwell time of the mouse cursor in the local anatomical region from the image display module. When the scaling parameters are greater than the preset value... When the scaling reference and absolute dwell time exceed a preset time threshold, a high-interest region label is generated and bound to the current labeled trajectory node using a data structure containing center coordinates and boundary range. Simultaneously, continuous undo / redo operations are converted into a state machine transition sequence. When the overlap of a discrete temporal sequence containing edit, dwell, undo, and new edit operations in spatial coordinates exceeds a preset overlap ratio threshold, the confidence parameter of the corresponding mask is lowered by calculating the duration hedging coefficient of the loop sequence, and the spatial coordinate interval is labeled as a low-confidence ambiguous sample to identify weak areas in the segmentation model. In practical engineering implementation, the latency loss of context switching in the retraining pipeline of deep learning models essentially refers to interactive editing... When the UI process and the model inference background process perform asynchronous data exchange in video memory, mutex lock scheduling, and graphics context reconstruction, the cumulative physical blocking time generated at the system kernel level is typically stable within the range of 15 to 45 milliseconds during continuous system operation. Due to this underlying hardware-level latency, the timestamps of doctor mouse interactions collected by the behavior tracking module at the application layer will experience a slight backward drift relative to the actual physical moment when the image is refreshed and displayed on the screen, thus causing spatiotemporal continuity decoupling. To eliminate this cross-scale timing bias, the dynamic data analysis module uses a kernel-level general-purpose timer to measure the absolute time consumed by each context switch in real time and uses it as an input parameter in the behavior trajectory sequence. The equivalent millisecond-level delay value is subtracted from the original timestamp of the corresponding interaction event to complete the temporal baseline alignment. This ensures that the macroscopic human operation trajectory and the underlying mask topology evolution achieve absolute spatiotemporal self-consistency within a control precision of 1 millisecond. The state recording output module receives high-attention region labels and low-confidence ambiguous samples. It encapsulates the connection modification semantic labels, boundary modification semantic labels, confirmation intent data, high-attention spatial range data, and low-confidence uncertainty index into a structured behavior trajectory sequence according to the timestamp temporal relationship. The structured behavior trajectory sequence is then output as a supervision signal to the training loop of the segmentation network to update the weight parameters of the segmentation network for topological misclassification data, structural adhesion region coordinates, and boundary correction features.To eliminate topological ambiguity in the annotation process, specifically, the specific supervision signal achieves weight optimization in the retraining pipeline by constructing a multi-task loss function. During incremental training, the downstream segmentation network's total loss function is a weighted combination of pixel-level cross-entropy loss and spatiotemporal trajectory constraint loss. When the downstream segmentation network receives a connection modification semantic label containing coordinates of structurally bonded regions, the training loop extracts the target pixel set corresponding to the coordinates of the structurally bonded regions in real time by calling the position index matrix. In the backpropagation calculation of the loss function, the loss weight coefficient of each pixel in this target pixel set is multiplied and amplified to three times the original weight, thus generating a mandatory topological splitting penalty term. This drives the network weights to be updated in a targeted manner for topologically misclassified data. Simultaneously, for boundary modification semantic labels, the system extracts edge pixels within the high-attention spatial range data, calculates the Hausdorff distance between the current network-predicted contour and the manually adjusted boundary, and incorporates it as an additional boundary alignment loss term into the total loss function. This guides the downstream segmentation network to perform refined iteration of weights based on boundary correction features.

[0036] Under preset computing power constraints, a data flow channel is constructed to feed back from data perception to model optimization. When transmitting data streams and performing feature fusion, the system designs cross-modal alignment projection rules for the original image signal composed of three-dimensional or two-dimensional tensors, the doctor interaction time series data composed of one-dimensional discrete time series, and the confidence decline results composed of single-value scalars. The system uses a spatial coordinate transformation matrix to project and map the two-dimensional pixel coordinates of the mouse cursor in the one-dimensional interaction time series to the three-dimensional grid space corresponding to the original image signal, achieving strong alignment in the spatiotemporal dimensions. Then, the data processing unit performs full-space broadcast replication of the scalar value of the confidence decline result, expanding it into a scalar feature map with a spatial size completely consistent with the image signal. Finally, the system directly splices the original image feature map, the aligned interaction time series feature map, and the broadcast confidence feature map in the corresponding spatial channel dimension, thereby fusing heterogeneous multimodal data into a composite high-order feature tensor with a unified structure, so that each behavioral event in the structured behavioral trajectory sequence can correspond to a complete closed-loop data flow path.

[0037] Example 2: In the case of interactive edge correction of tumors and adjacent blood vessel adhesion areas assisted by a medical image segmentation system, a public dataset of multi-center clinical contrast-enhanced computed tomography scans was used as the experimental data source. The experiment was conducted on a hardware test platform consisting of a central processing unit with parallel floating-point tensor operation capabilities, an image display terminal with a resolution of 3840 x 2160 pixels, and an interactive input device with a sampling frequency of 125Hz. The determination of the preset time threshold for absolute dwell time was technically considered to balance the sensitivity of capturing clinically relevant regions with the system processing load. When the complexity of the corrected anatomical structure boundary increases with fluctuations in the local image grayscale gradient, to avoid false feature interference caused by frequent transient interruptions by clinical annotators, the preset time threshold value is positively correlated with the boundary complexity. In the specific operating conditions of this experiment, the preset time threshold was set to 800ms as the evaluation benchmark. To ensure visual attention weighting... The operator can accurately quantify the diagnostic focus of clinicians. This invention constructs a deterministic multidimensional feature mapping weight model. The operator periodically acquires the scaling parameter of the image display module, the reciprocal of the magnitude of the translation velocity vector, and the absolute dwell time of the mouse cursor in the local anatomical region as continuous independent variables. It multiplies the current scaling parameter with the absolute dwell time, and then divides the product result by the sum of the magnitude of the translation velocity vector and 1. This ensures that a high score is output under conditions of low speed, high magnification, and long dwell time. Then, the operator inputs the product combined score into a preset scalar regularization function, dynamically maps and normalizes it to a standard scalar range between 0.0 and 1.0. The normalized scalar value is used as the final visual attention weight. When the weight value exceeds the preset score threshold of 0.65 and the duration meets the above-mentioned 800 milliseconds, the system determines that the region has high diagnostic attention and triggers the generation of a high attention region label.

[0038] The experiment divided test subjects into a control group, a partially missing control group, and an experimental group. The control group used the traditional static recording method that only saved the final mask. The partially missing control group turned off the topological invariant parsing operator when recording the annotation behavior trajectory. The experimental group used the complete system. The experiment simulated the non-ideal annotation state by artificially injecting a pseudo-mask contour containing slight hand tremor bias and Gaussian white noise in the interactive editing interface. The degree of structural anatomical ambiguity in the tumor and vascular adhesion areas was divided into three problem intensity gradients: low, medium, and high. The test results showed that when the degree of ambiguity increased from low to high, the downstream segmentation network showed a flat decay in the Dessian similarity coefficient after retraining with its final mask dataset because the control group could not capture the evolution path of anatomical cognition. The average measured values ​​were only 0.724, 0.681, and 0.612, respectively, and its anatomical topological error rate was as high as 14.3%, showing significant performance degradation. Correspondingly, in the partially missing control group, the anatomical topological error rate was as high as 14.3%, showing significant performance degradation. In the missing control group, due to the lack of closed-loop topological analysis of the changes in the number of connected components and cavities, the system-level behavioral trajectory sequence could not correctly decouple positional bias and topological mutation, resulting in average Descein similarity coefficients of 0.782, 0.735, and 0.664 after retraining. In the experimental group using the complete scheme, the labeled behavior tracking module and the dynamic data analysis module formed a causal positive feedback. Even under strong noise interference, the topological invariant analysis operator could still translate the physical trajectory into a structured data stream reflecting anatomical cognition. The structured behavioral trajectory sequence output by the network served as a specific supervision signal to guide the network weight update. Under the three ambiguous gradients of low, medium, and high, the average Descein similarity coefficients of the downstream segmentation network after retraining were stably 0.915, 0.893, and 0.864, respectively, and the anatomical topological error rate of the final mask converged significantly to 0.8%, confirming the system-level synergistic effect generated by the complete feature combination.

[0039] The experiment also included test conditions with an out-of-range control group exceeding the value boundaries. When the spatial variation matrix threshold was set below the lower limit of the protection range, the algorithm became overly sensitive to minute measurement fluctuations in the change of the contour boundary curvature tensor. This caused the system to frequently misinterpret normal physiological boundary translations as topological or boundary modification structured semantics, thus frequently adjusting the rendering weight of the intention feedback window. This induced data congestion in the local bus transmission, causing the transmission delay of the intention data queue to jump to 142.6ms, demonstrating a nonlinear overload effect on the processing load. However, when the spatial variation matrix threshold was set within the protection range... When the range exceeds the upper limit, the data processing unit cannot identify local geometric calibration features, causing it to misjudge a large number of boundary fine-tuning behaviors as invalid redundant events with no intrinsic topological mutations. Consequently, it cannot issue the corresponding second-state arbitration signal, resulting in a significant loss of high-order semantics in the structured behavior trajectory sequence. This leads to a decrease in segmentation performance after retraining, with the Dessian similarity coefficient deteriorating to 0.751. The failure evolution data of the above-mentioned out-of-range control group provides the physical boundary basis for the working window of the parameter limit of this invention, confirming the stability and feasibility of the method system in the control loop of healthcare informatics.

[0040] Example 3: This example combines Figures 1 to 2 This document describes a system and method for recording the dynamic annotation process of medical images based on semantic understanding of physician annotation behavior. Figure 1 As shown, the medical image dynamic annotation system includes an image display module, an annotation behavior tracking module, a dynamic data analysis module, and a status recording output module. The image display module provides an interactive editing interface and connects to the data acquisition interface to link with the annotation behavior tracking module. The annotation behavior tracking module acquires behavior and undo / redo data and sends view behavior and operation data to the dynamic data analysis module. The dynamic data analysis module receives the view behavior and operation data to generate labels and label ambiguous samples. The dynamic data analysis module has steps S101, S102, and S103 during the operation. The dynamic data analysis module outputs region labels and ambiguous samples to the status recording output module. The status recording output module encapsulates a structured behavior trajectory sequence and outputs the behavior trajectory sequence. The output behavior trajectory sequence is transmitted to the deep learning model retraining pipeline so that the deep learning model retraining pipeline receives specific supervision signals.

[0041] like Figure 2As shown, the medical image dynamic annotation process recording system based on semantic understanding of physician annotation behavior involves equipment including an image display terminal, an interactive input device, a central processing unit (CPU), and a retraining pipeline. The image display terminal includes an image display module and an interactive editing interface. The interactive input device is connected to the interactive editing interface. The interactive editing interface is connected to the annotation behavior tracking module integrated in the CPU via a data acquisition interface. The annotation behavior tracking module is connected to the dynamic data analysis module integrated in the CPU. The dynamic data analysis module contains a baseline calibration loop and is connected to the state recording output module integrated in the CPU. The state recording output module outputs a structured behavior trajectory sequence to the retraining pipeline. The retraining pipeline includes a downstream lesion segmentation network to receive the structured behavior trajectory sequence.

[0042] Example 4: When anatomical variability deviations occur in the interactive edge correction assisted by the medical image segmentation system, the clinical annotator introduces physiological hand tremors when moving the input device and trajectory deviations due to cognitive hesitation. This causes the continuous interoperability command stream collected by the annotation behavior tracking module to exhibit nonlinear electrical signal fluctuations in the local geometric space. Traditional control systems rely solely on static empirical parameters to calculate mean filtering, which cannot separate the local geometric features belonging to the annotator's subjective calibration intention from the invalid physical displacements belonging to random muscle tremors in the interactive input with random noise. This causes the trigger boundary of the modification intention feedback window to bifurcate, resulting in the behavior trajectory sequence output by the state recording output module containing confusing noise. When the downstream segmentation network is retrained using a dataset containing incorrect semantic labels, the network weight update direction deviates and the anatomical topology error rate increases, constituting a technical problem that restricts the stability of the medical informatics system in clinical scenarios.

[0043] The spatial variation matrix threshold range of this invention is 0.35 to 0.65. Before image data analysis, the control loop imports a reference interaction trajectory containing standard anatomical contours and injected with Gaussian physiological tremor noise at a fixed frequency of 5Hz to 10Hz as a verification object at the input end. The dynamic data analysis module runs under the functional specification of the pre-processor meeting 1 billion floating-point operations per second, sequentially calling control states containing different parameter steps, calculating the change in contour boundary curvature tensor of the reference interaction trajectory under different values, and performing analysis on 10 discrete value points ranging from 0.1 to 1.0 with increments of 0.1. Monotonicity tests were completed. Experimental results showed that when the spatial variation matrix threshold was set below 0.35, the system oversaturated to the input tremor, frequently triggering the second-state arbitration signal and misclassifying physiological tremors as boundary modification structured semantics. When the spatial variation matrix threshold was set above 0.65, the system's signal interception performance deteriorated, causing the system to discard edge correction actions as invalid redundant events. By calculating the extreme values ​​of the misclassification distribution in the above operation path, 0.45 was determined as the preferred spatial variation matrix threshold, which was used as a closed-loop criterion to distinguish between subjective calibration actions and muscle tremors.

[0044] When the spatial variation matrix threshold is fixed at 0.45, the data processing unit intercepts and filters out 99.2% of unintentional input noise in the continuous operation flow, while retaining the effective boundary correction actions and converting them into boundary modification structured semantics. This drives the modification intent feedback window to adapt to the annotator's intent state, obtaining short-term confirmation intent data. The state recording output module encapsulates the structured behavioral trajectory sequence produced to remove physical deviations caused by physiological losses. The annotator's anatomical cognitive data, including confidence parameters, is transmitted to the downstream segmentation network's retraining pipeline for targeted updates of the weight parameters of topological error classification data and boundary correction features. This reduces pixel-level annotation ambiguity at lesion boundaries. The system forms a closed loop of data flow from motion capture in physical space to behavioral semantic understanding in anatomical space.

[0045] Example 5: When the system faces the situation of initial access to heterogeneous medical image network interfaces and initial synchronization of different hardware terminal specifications, the annotation behavior tracking module initiates a standardized de-identification self-test under the technical premise of explicit user authorization. The image display module loads a preset standardized anonymized example image, and the data acquisition interface periodically reads the screen refresh rate parameters, pixel pitch parameters, and initial physical jitter variance of the interactive input device of the current display terminal. The dynamic data analysis module inputs the coordinates of the mouse stationary trajectory for ten consecutive cycles to the baseline calibration loop. The hardware noise reference value is established by calculating the reciprocal of the initial physical jitter variance. On this basis, when the interactive input device is detected to have completed five standard circular operations in the anonymized area, the system completes the initial calibration of the motion damping coefficient based on the ratio of the absolute dwell time of the mouse cursor in the local area to the number of pixels traversed.

[0046] After establishing the hardware noise baseline, the dynamic data analysis module retrieves the spatial variation matrix threshold from the previous embodiment, superimposes the spatial variation matrix threshold with the hardware noise baseline, and automatically corrects the sensitivity gain of the topological invariant analytical operator based on the superimposed value. This is to avoid misjudgment of structured semantics of connection modification caused by hardware latency of different interactive devices. The data processing unit then locks the calibration parameter group in the current working mode and writes the anonymized self-test baseline data into the configuration register of the status record output module. This ensures that the system maintains the monotonic stability of the technical feature alignment relationship in the subsequent continuous operation of the healthcare informatics channel. Under this baseline calibration state, the mathematical mapping logic of the difference compensation factor... Specifically, the dynamic data analysis module takes the absolute latency loss scalar value of context switching in the deep learning model retraining pipeline, which is collected in real time, as the only independent variable input. Through the system's built-in time base mapping program, the latency loss is multiplied by an equal and inverse coefficient to generate a millisecond-level time base bias with completely equal values ​​but opposite signs. This bias is defined as the difference compensation factor. The configuration register outputs the difference compensation factor to the timing correction unit of the status recording output module in real time. This performs direct algebraic superposition correction on the current event timestamp in the structured behavior trajectory sequence to completely offset the hardware-level latency trailing caused by process switching and memory swapping at the underlying level, ensuring the accurate alignment of the self-test baseline data.

[0047] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior, characterized in that, include: The image display module is used to display medical images to be annotated and provide an interactive editing interface; The annotation behavior tracking module is equipped with a data acquisition interface that interfaces with the image display module to acquire non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface; The dynamic data analysis module is used to receive non-editable view behavior data and continuous undo / redo operation data, and includes the following actions during the calculation: Step S101, input the non-editable view behavior data into the visual attention weight operator to generate high attention area labels; Step S102: Establish the state machine transition sequence corresponding to the continuous undo / redo operation data; Step S103: When the state machine transition sequence meets the overlap threshold in spatial coordinates, calculate the duration correction coefficient to lower the confidence parameter of the corresponding mask, thereby calibrating the low confidence ambiguous sample. The status recording output module, connected to the dynamic data analysis module, is used to receive high-attention region labels and low-confidence ambiguous samples output by the dynamic data analysis module, encapsulate the high-attention region labels and low-confidence ambiguous samples into a structured behavior trajectory sequence according to the time stamp sequence relationship, and output the structured behavior trajectory sequence to the retraining pipeline of the deep learning model.

2. The medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, Non-editable view behavior data includes: the scaling parameter of the image display module, the translation speed vector, and the absolute dwell time of the mouse cursor in the local anatomical area; step S101 includes the following sub-steps: step S1011, when the scaling parameter is greater than the preset scaling reference and the absolute dwell time exceeds the preset time threshold, a high-attention area label is generated; step S1012, the high-attention area label is bound to each labeled trajectory node in the structured behavior trajectory sequence using a data structure containing the center coordinates and boundary range.

3. The medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, The continuous undo / redo operation data is converted into a state machine transition sequence by converting continuous editing and undo / redo actions; in step S103, the state machine transition sequence meets the overlap threshold in spatial coordinates, including: detecting that the overlap of the discrete time sequence composed of editing operation, continuous pause, undo operation and new editing operation in spatial coordinates is greater than the preset overlap ratio threshold.

4. The medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, The structured behavior trajectory sequence encapsulated by the state recording output module includes: connection modification semantic labels, boundary modification semantic labels, short-term confirmation intent data, high attention spatial range data, and low confidence uncertainty index. The state recording output module converts the structured behavior trajectory sequence into a specific supervision signal, which is used by the downstream segmentation network in the retraining pipeline to update the weight parameters for topology misclassification data, structural adhesion region coordinates, and boundary correction features.

5. A medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, After the dynamic data analysis module completes step S103, it calculates the latency loss of context switching in the retraining pipeline of the deep learning model, calculates the difference compensation factor based on the latency loss, and uses the difference compensation factor to correct the difference of timestamps in the structured behavior trajectory sequence.

6. The medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, Medical images are medical imaging data containing samples of difficult-to-identify lesions; the interactive editing interface corrects the topological regions by manually adjusting the mask boundaries based on the pre-segmented contours generated by the model; low-confidence ambiguous samples are input into the retraining pipeline to calibrate the inherent weak areas of the deep learning model and to isolate samples of difficult-to-identify lesions.

7. A medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, When the dynamic data analysis module transmits the data stream, it performs feature fusion on the original image signal, doctor interaction time series data and confidence decline results, so that each behavioral event in the structured behavioral trajectory sequence corresponds to a closed-loop data flow path.

8. A medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, The deep learning model retraining pipeline receives structured behavioral trajectory sequences to fine-tune the weight coefficients of the downstream lesion segmentation network, and converts high-interest region labels and low-confidence ambiguous samples into specific supervision signals, which are then output to the downstream lesion segmentation network.

9. A medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior according to claim 1, characterized in that, The system also includes a model update evaluation module; the input of the model update evaluation module is connected to the output of the state recording output module, which is used to obtain the structured behavior trajectory sequence and calculate the lesion segmentation accuracy of the deep learning model after incremental training. When the lesion segmentation accuracy reaches the preset performance index, the updated deep learning model is sent to the image display module to update the pre-segmented contour in the interactive editing interface.

10. A method for recording the dynamic annotation process of medical images based on semantic understanding of doctor annotation behavior, which operates in the medical image dynamic annotation process recording system based on semantic understanding of doctor annotation behavior as described in claim 1, characterized in that, Includes the following steps: Step S101: Display the medical image to be annotated through the image display module and provide an interactive editing interface; Step S102: Obtain non-editable view behavior data and continuous undo / redo operation data in the interactive editing interface through the data acquisition interface of the annotation behavior tracking module; Step S103: Input non-editable view behavior data into the visual attention weight operator through the dynamic data analysis module to generate high attention area labels, and establish a state machine transition sequence corresponding to continuous undo / redo operation data. When the state machine transition sequence meets the overlap threshold in spatial coordinates, calculate the duration correction coefficient to lower the confidence parameter of the corresponding mask, thereby calibrating low-confidence ambiguous samples. Step S104: Receive high-attention region labels and low-confidence ambiguous samples output by the dynamic data analysis module through the state recording output module, encapsulate the high-attention region labels and low-confidence ambiguous samples into a structured behavior trajectory sequence according to the time stamp sequence relationship, and output the structured behavior trajectory sequence to the retraining pipeline of the deep learning model.

Images