Medical data labeling model training method and medical data labeling method

Abstract

The present application belongs to the technical field of medical data labeling, and discloses a training method of a medical data labeling model and a medical data labeling method. The method comprises: obtaining multiple expert labeling data corresponding to medical images and labeling behavior characteristics thereof; generating behavior characteristic representations of each expert based on the labeling behavior characteristics, and further obtaining an implicit diagnosis intention vector; dividing the expert labeling data into a labeling group with consistent intentions and a labeling group with divergent intentions, generating fusion labeling for the labeling group with consistent intentions, and respectively retaining multiple intention sub-labels for the labeling group with divergent intentions; dynamically adjusting a supervision weight based on the implicit diagnosis intention vector, and training a labeling model using the fusion labeling and the intention sub-labels. The labeling model trained by the present application can simultaneously and stably learn expert consensus labeling results and multiple systematic diagnosis intention preferences, thereby effectively solving the technical problems of labeling signal contradiction and unstable model boundary caused by differences in clinical strategies.

Description

Technical Field

[0001] This invention relates to the field of medical data annotation technology, specifically to a training method for a medical data annotation model and a medical data annotation method. Background Technology

[0002] Medical data processing platforms can provide end-to-end services, from image data access and high-quality dataset construction to intelligent annotation, model training and validation, and model deployment and version management, facilitating the monetization of medical data. The training of medical data annotation models typically relies on a large amount of expert-annotated data. To improve annotation quality and overcome single-expert bias, actual annotation is often completed collaboratively by multiple doctors.

[0003] However, medical image annotation is highly subjective and influenced by varying clinical strategies, often resulting in systematic biases in the annotations of different experts. For example, some physicians tend to adopt more aggressive early screening strategies, thus expanding the boundaries of suspected lesions; while others adopt a more conservative follow-up strategy, with more conservative annotation boundaries. In such cases, the discrepancies between annotations from multiple experts are not simply random noise, but rather stem from different implicit diagnostic and treatment intentions.

[0004] Existing methods for training medical data annotation models typically treat disagreements among experts as random errors, obtaining a single merged label through pixel-level averaging, voting, or confidence fusion. However, when annotation differences stem from differences in clinical strategies, simple fusion of multi-expert annotations can disrupt semantic consistency, resulting in inconsistent supervision signals in the training data. This leads to unstable boundaries learned by the model and difficulty in generalization.

[0005] Therefore, existing technologies lack mechanisms for identifying and utilizing experts' implicit diagnostic intentions, making it impossible to address systemic disagreements in a targeted manner, thus affecting the rationality and stability of model training. Summary of the Invention

[0006] The purpose of this invention is to provide a training method for a medical data annotation model and a medical data annotation method to solve the problems mentioned in the background art.

[0007] This invention provides a training method for a medical data annotation model, applied to a medical data annotation platform, comprising the following steps: For each medical image in the training set, obtain the annotation data of multiple experts corresponding to that medical image and their annotation behavior characteristics. Based on the labeled behavioral features, behavioral feature representations of each expert are generated through a behavioral coding network, and the behavioral feature representations are input into the implicit intent inference model to obtain implicit diagnostic intent vectors representing the labeling strategies of each expert. Based on the similarity between the implicit diagnostic intent vectors of multiple experts, the expert-annotated data is divided into intent... Figure 1 Consistent annotation groups and annotation groups with conflicting intentions, regarding the meaning... Figure 1 Consistent annotation groups generate fused annotations, while annotation groups with divergent intents are retained as multiple intent sub-labels; The supervision weights are dynamically adjusted based on the implicit treatment intent vectors of each expert, and the annotation model is trained using the fused annotations and the intent sub-labels to obtain a medical data annotation model capable of simultaneously learning expert consensus and multiple treatment intent preferences; wherein, the annotation model includes methods for predicting intent... Figure 1 The main branch leads to the region labeling results, and multiple intent branches are used to learn the labeling preferences of different intent sub-labels.

[0008] As an example, the annotation behavior characteristics include behavioral information collected during the annotation process, including boundary drawing order, local adjustment actions, annotation cancellation behavior, and pause behavior.

[0009] As an example, the implicit intent inference model includes: The behavior attention extraction subnetwork is used to perform weighted modeling of boundary drawing order, local adjustment action, annotation cancellation behavior and pause behavior based on the importance difference of different behavior types in the labeled behavior features, so as to generate the first behavior embedding vector with behavior intention attention. The local conflict identification subnetwork is used to identify local conflict patterns presented by experts during the annotation process based on the first row of the embedding vector, including boundary expansion trend, boundary contraction trend and repeated confirmation behavior of local regions, and transforms the identification results into the second row of the embedding vector. The intent-behavior dual-stream fusion encoder independently encodes the first-behavior embedding vector and the second-behavior embedding vector, and aligns them across streams in the latent space to generate a joint behavior latent representation. The directional intent decoder, based on the joint behavioral latent representation, performs probabilistic modeling on the multidimensional clinical strategy directions related to expanding, shrinking, sensitive attention, and conservative screening of the annotation region, thereby outputting an implicit diagnostic and treatment intent vector representing the expert annotation strategy.

[0010] As an example, the implicit intent inference model also includes a behavior deviation calibration subnetwork, which is used to calibrate the behavior attention extraction subnetwork and the intent-behavior dual-stream fusion encoder in the implicit intent inference model. By reweighting the feature channels corresponding to abnormal behaviors, normalizing the features of offset behaviors, or suppressing noise in unstable behavior segments, the accuracy of generating implicit diagnostic intent vectors can be improved.

[0011] As an example, the behavior bias calibration subnetwork calibrates the behavior attention extraction subnetwork, including: Based on the degree of deviation of the labeled behavioral features, the attention weights of each behavioral feature channel in the behavioral attention extraction subnetwork are recalculated to suppress abnormal attention responses caused by excessively fast drawing, jittery trajectories, or abnormal pauses. The attention weights are also renormalized to improve the stability of the first behavioral embedding vector output by the behavioral attention extraction subnetwork.

[0012] As an example, the behavior deviation calibration subnetwork performs calibration processing on the intent-behavior dual-stream fusion encoder, including: The intermediate features generated by the behavioral flow and conflict flow are scale aligned or distribution corrected, and the cross-flow offsets that exist during the feature fusion process are compensated, so that the fused joint behavioral latent representation maintains a consistent feature scale and distribution in the latent space, thereby improving the accuracy of generating implicit diagnostic intent vectors.

[0013] As an example, the main branch of the annotation model shares an encoder structure with multiple intent branches, and outputs the prediction results of the main branch and the prediction results of multiple intent branches respectively during the decoding stage.

[0014] This invention also provides a medical data annotation method, applied to a medical data annotation platform, the method comprising: Receive medical images to be labeled; The medical image to be labeled is input into the medical data labeling model obtained according to any of the preceding training methods; Obtain the annotation results output by the medical data annotation model.

[0015] As an example, obtaining the annotation results includes at least one of the following methods: Obtain the consensus annotation results output by the main branch of the medical data annotation model; Receive the user's selected treatment intention preference, and obtain the annotation results that conform to the treatment intention preference output by the corresponding intention branch in the medical data annotation model; Simultaneously, the consensus annotation result output by the main branch and the specific intent annotation result output by one or more intent branches are obtained, and the consensus annotation result and the specific intent annotation result are merged and displayed.

[0016] As an example, the method also includes: Record user interaction behavior during the current annotation process to form new annotation behavior features. Input the new annotation behavior features into the implicit intent inference model associated with the medical data annotation model to update the implicit diagnosis and treatment intent vector. The output of the medical data annotation model is dynamically fine-tuned based on the updated implicit diagnostic intent vector.

[0017] This invention infers and quantifies the implicit diagnostic and treatment intent vectors from the annotation behavior of multiple experts, distinguishes between consensus and divergence regions based on intent similarity, and constructs a multi-branch model and dynamic supervision weight strategy accordingly. This enables the trained annotation model to robustly learn both expert consensus annotation results and various systematic diagnostic and treatment intent preferences, thereby effectively solving the technical problems of contradictory annotation signals and unstable model boundaries caused by differences in clinical strategies. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a training method for a medical data annotation model disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the implicit intent inference model disclosed in an embodiment of the present invention; Figure 3 This is another structural schematic diagram of the implicit intent inference model disclosed in an embodiment of the present invention; Figure 4 This is a flowchart illustrating a medical data annotation method disclosed in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of 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.

[0020] Please see Figure 1 This invention provides a training method for a medical data annotation model, applied to a medical data annotation platform, comprising the following steps: Step S10: For each medical image in the training set, obtain the annotation data of multiple experts corresponding to the medical image and their annotation behavior characteristics. In this step, the solution of the present invention is applied to a medical data processing platform, which obtains independent annotation results of the same medical image from multiple experts, including the annotation data finally submitted by the experts (such as the boundary of the lesion area, category label, etc.) and annotation behavior characteristics.

[0021] As an example, the annotation behavior characteristics include behavioral information collected during the annotation process, including boundary drawing order, local adjustment actions, annotation cancellation behavior, and pause behavior.

[0022] Among them, annotation behavior characteristics are behavioral information collected during the annotation process that reflects the annotation habits and thinking process of experts. These include, but are not limited to: the boundary drawing order of experts drawing lesion boundaries (e.g., from top to bottom or from inside to outside), local adjustment actions on drawn boundaries (e.g., repeatedly stretching or shrinking a certain section of boundary), annotation cancellation behavior (e.g., going back to a certain step and re-annotating), and pause behavior during the annotation process (e.g., staying in a specific area for a long time).

[0023] It's important to note that the order in which boundaries are drawn represents the expert's overall approach to observing and delineating lesions. For example, diffusion from the lesion's core outwards might suggest a primary-to-secondary confirmation strategy. Local adjustments reflect the expert's pursuit of boundary precision and their repeated deliberation on uncertain areas. The act of marking and withdrawing directly reflects the expert's decision-making process and is a key signal for capturing uncertainty or strategic hesitation. Pauses typically occur in areas of diagnostic difficulty; the duration and location of these pauses may be related to the complexity of the lesion or the expert's level of caution. This platform records and analyzes the timestamps, spatial locations, and action types of these behaviors to construct sequential features that reflect the expert's cognitive process. These behavioral features are recorded in time series or other structured forms for subsequent inference of the expert's deeper clinical intentions.

[0024] Step S20: Based on the labeled behavioral features, generate behavioral feature representations for each expert through a behavioral coding network, and input the behavioral feature representations into the implicit intent inference model to obtain implicit diagnostic intent vectors representing the labeling strategies of each expert; In this step, the platform utilizes a behavior encoding network to model the labeled behavioral feature sequence for each expert. This behavior encoding network can be a neural network capable of processing sequential data, such as a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), a Transformer network based on a self-attention mechanism, or a convolutional temporal network (such as 1D-CNN); this invention does not specifically limit it. The role of the behavior encoding network is to capture the dependencies and patterns of labeled behaviors in the temporal dimension, encoding the original, discrete behavioral sequence into a continuous, high-dimensional behavioral feature representation.

[0025] The aforementioned behavioral characteristics are input into a latent intent inference model. This model, once trained, is able to decode the expert's latent, unstated clinical strategy preferences from behavioral patterns and output a quantitative vector of implicit diagnostic and treatment intentions. This vector represents the expert's labeling strategy in the latent space, such as their preference for aggressive screening or robust follow-up, or their sensitivity to specific types of lesions.

[0026] Step S30: Based on the similarity between the implicit diagnostic intent vectors of multiple experts, the expert-annotated data is divided into intent... Figure 1 Consistent annotation groups and annotation groups with conflicting intentions, regarding the meaning... Figure 1 Consistent annotation groups generate fused annotations, while annotation groups with divergent intents are retained as multiple intent sub-labels; In this step, the platform calculates the similarity between all implicit diagnostic intent vectors of experts targeting the same image. Similarity can be determined using vector distance metrics, such as calculating at least one of Euclidean distance, cosine distance (measuring directional consistency), or Mahalanobis distance (considering the covariance of vector distributions). Based on a preset similarity threshold, experts are divided into several clusters.

[0027] Within the same cluster, Figure 1 Consistent annotation groups indicate that these experts employ highly similar clinical strategies, and differences in their annotation results are largely attributed to random errors or subtle individual variations. Therefore, methods such as pixel-level averaging, the STAPLE algorithm, or confidence-based fusion are used to generate a fused annotation, serving as a consensus supervision signal under this strategy.

[0028] For labeled groups belonging to different clusters with divergent intentions, the labeling differences are considered to stem from systemic disagreements in clinical strategies. Instead of simply mixing these labels, they are retained as multiple intention sub-labels, each representing a labeling result under a unique clinical intention preference.

[0029] Step S40: Dynamically adjust the supervision weights based on the implicit treatment intent vectors of each expert, and train the annotation model using the fused annotations and the intent sub-labels to obtain a medical data annotation model capable of simultaneously learning expert consensus and multiple treatment intent preferences; wherein, the annotation model includes methods for predicting intent... Figure 1 The main branch leads to the region labeling results, and multiple intent branches are used to learn the labeling preferences of different intent sub-labels.

[0030] In this step, the platform constructs a structured annotation model containing a main branch and multiple intent branches. The main branch and the multiple intent branches can share an encoder structure (such as a common CNN backbone network) for extracting general features from the input image. During the decoding phase, the main branch outputs the intent branches... Figure 1 The prediction results are directed to the region, while each intention branch outputs the prediction results under its corresponding treatment intention preference.

[0031] During training, the platform dynamically adjusts the supervision weights of the loss function based on the implicit treatment intent vector obtained in step S20 and the regions (consistent regions or divergent regions) divided in step S30. Specifically: For being judged as intentional Figure 1 In regions where there is consensus, the supervisory weight of the main branch in the prediction loss function is increased, forcing the model to learn in those regions in accordance with expert consensus. In regions where intentions are deemed divergent, the supervisory weight of the main branch in those regions is reduced to prevent it from learning contradictory signals; at the same time, the supervisory weight of each intention branch in the region covered by its corresponding intention sub-label is increased, so that each intention branch can focus on learning the annotation pattern under its specific clinical strategy.

[0032] The fused annotations and intent sub-labels processed using the aforementioned dynamic supervision strategy are used to train the annotation model end-to-end. The resulting medical data model can not only predict robust consensus boundaries but also output annotation results reflecting different clinical strategy preferences through different intent branches as needed, thereby simultaneously learning expert consensus and multiple diagnostic and treatment intent preferences.

[0033] This invention infers and quantifies the implicit diagnostic and treatment intent vectors from the annotation behavior of multiple experts, distinguishes between consensus and divergence regions based on intent similarity, and constructs a multi-branch model and dynamic supervision weight strategy accordingly. This enables the trained annotation model to robustly learn both expert consensus annotation results and various systematic diagnostic and treatment intent preferences, thereby effectively solving the technical problems of contradictory annotation signals and unstable model boundaries caused by differences in clinical strategies.

[0034] As an example, the implicit intent inference model includes: The behavior attention extraction subnetwork is used to perform weighted modeling of boundary drawing order, local adjustment action, annotation cancellation behavior and pause behavior based on the importance difference of different behavior types in the labeled behavior features, so as to generate the first behavior embedding vector with behavior intention attention. The local conflict identification subnetwork is used to identify local conflict patterns presented by experts during the annotation process based on the first row of the embedding vector, including boundary expansion trend, boundary contraction trend and repeated confirmation behavior of local regions, and transforms the identification results into the second row of the embedding vector. The intent-behavior dual-stream fusion encoder independently encodes the first-behavior embedding vector and the second-behavior embedding vector, and aligns them across streams in the latent space to generate a joint behavior latent representation. The directional intent decoder, based on the joint behavioral latent representation, performs probabilistic modeling on the multidimensional clinical strategy directions related to expanding, shrinking, sensitive attention, and conservative screening of the annotation region, thereby outputting an implicit diagnostic and treatment intent vector representing the expert annotation strategy.

[0035] In embodiments of the present invention, such as Figure 2As shown, the latent intent inference model includes a behavior attention extraction subnetwork, a local conflict recognition subnetwork, an intent-behavior dual-stream fusion encoder, and a directional intent decoder.

[0036] Specifically, the behavior attention extraction subnetwork receives the original sequence of behavior features. Based on a self-attention mechanism, this subnetwork learns the differences in importance of different behavior types within the labeled behavior features for inferring the final intent. For example, it assigns higher attention weights to repetitive local adjustment actions or critical undo actions, while giving standard weights to regular drawing sequences. Through weighted modeling, this subnetwork outputs a first behavior embedding vector with behavior intent attention, highlighting the most critical behavior patterns for intent inference.

[0037] The local conflict identification subnetwork focuses on mining local conflict patterns exhibited by experts during the annotation process that may indicate their strategic tendencies. This subnetwork analyzes behavioral sequences to identify behaviors such as boundary expansion trends (continuous outward fine-tuning), boundary contraction trends (continuous inward fine-tuning), or repeated confirmation behaviors of a local region. These patterns are transformed into a structured representation, outputting as a second-behavior embedding vector that encodes the expert's tendency in pixel-level decision-making.

[0038] The intent-behavior dual-stream fusion encoder receives the outputs from the two sub-networks mentioned above in parallel: the first line contains an embedding vector, and the second line contains an embedding vector. This encoder contains two independent encoding paths, each extracting deep features from one of the two embedding vectors. Subsequently, in a shared latent space, the information from the two paths is deeply fused using methods such as cross-stream attention or feature concatenation alignment to generate a comprehensive and unified joint behavioral latent representation that simultaneously implies global behavioral attention and local decision-making tendencies.

[0039] The directional intent decoder decodes based on a joint behavioral latent representation. Instead of outputting a single category, this decoder probabilistically models multiple pre-defined, clinically strategy-related multidimensional directions (e.g., directions related to strategies such as expanding or shrinking the labeled area, focusing on subtle features, or favoring conservative screening). By calculating the projection or score of this latent representation in each direction, a multidimensional, continuous-valued implicit diagnostic intent vector is finally output. Understandably, each dimension of this implicit diagnostic intent vector represents the strength of the expert's tendency in that specific clinical strategy direction, thus quantitatively characterizing their labeling strategy.

[0040] This invention employs an implicit intent inference model comprised of behavioral attention extraction, local conflict identification, dual-stream fusion, and directional decoding. This model can accurately separate and quantify the multidimensional clinical intent of experts in terms of global strategic focus and local decision-making tendencies from complex labeled behavioral sequences. It can provide high-fidelity and interpretable quantitative evidence for subsequent intent grouping and dynamic supervision.

[0041] As an example, the implicit intent inference model also includes a behavior deviation calibration subnetwork, which is used to calibrate the behavior attention extraction subnetwork and the intent-behavior dual-stream fusion encoder in the implicit intent inference model. By reweighting the feature channels corresponding to abnormal behaviors, normalizing the features of offset behaviors, or suppressing noise in unstable behavior segments, the accuracy of generating implicit diagnostic intent vectors can be improved.

[0042] In embodiments of the present invention, such as Figure 3 As shown, the latent intent inference model further integrates a behavior bias calibration subnetwork. This subnetwork is used to calibrate the behavior attention extraction subnetwork and the intent-behavior dual-stream fusion encoder in the latent intent inference model. The working basis of this subnetwork is to identify anomalous patterns in the expert's annotation behavior data. These anomalous patterns may originate from the expert's non-intent-based operations during the annotation process, such as drawing trajectory jitter caused by operating habits or external interference, rapid and repeated undoing and redoing that deviates from the normal logic, or abnormally long pauses in non-critical areas.

[0043] To eliminate the interference of such abnormal behavior on the intention inference process, the behavior deviation calibration subnetwork performs one or a combination of the following core calibration operations: (1) Feature channel reweighting: Identify feature channels that are strongly associated with abnormal behavior patterns and assign them lower weight coefficients to suppress the influence of these features in subsequent calculations; (2) Feature distribution normalization correction: Detect the intermediate feature distribution shift caused by systematic behavior deviation and apply a correction function to adjust the feature distribution to a preset normal range; (3) Noise suppression: For the identified, transient and meaningless abnormal behavior segments, smooth or attenuate the corresponding feature sequence segments to reduce random noise.

[0044] Through the above calibration mechanism, the behavioral bias calibration subnetwork can provide a pre-processing barrier for the implicit intention inference process, aiming to improve the robustness of the mapping from behavioral sequence to intention vector, so that the final generated implicit diagnostic intention vector is less affected by operational noise, thereby more stably and accurately reflecting the expert's true clinical strategy tendency.

[0045] As an example, the behavior bias calibration subnetwork calibrates the behavior attention extraction subnetwork, including: Based on the degree of deviation of the labeled behavioral features, the attention weights of each behavioral feature channel in the behavioral attention extraction subnetwork are recalculated to suppress abnormal attention responses caused by excessively fast drawing, jittery trajectories, or abnormal pauses. The attention weights are also renormalized to improve the stability of the first behavioral embedding vector output by the behavioral attention extraction subnetwork.

[0046] Specifically, when the behavior bias calibration subnetwork calibrates the behavior attention extraction subnetwork, the processing procedure is as follows: The behavior deviation calibration subnetwork receives the input or intermediate output of the behavior attention extraction subnetwork and evaluates the degree of abnormality or deviation of each segment in the labeled behavior feature sequence. The evaluation criteria include, but are not limited to, whether the duration of the behavior segment deviates significantly from the norm, whether the smoothness of the spatial trajectory is insufficient, or whether the logical coherence of the operation sequence is interrupted.

[0047] Based on the deviation information obtained from the evaluation, the behavioral deviation calibration subnetwork intervenes in the attention weight calculation process in the behavioral attention extraction subnetwork. Its main intervention method is to suppress excessive or unreasonable attention responses triggered by abnormal behaviors (e.g., instantaneous high-frequency signals generated by excessively fast drawing, cluttered spatial features introduced by jittery trajectories, or temporal gaps caused by abnormal pauses).

[0048] After suppression is complete, the behavioral bias calibration subnetwork triggers a re-normalization calculation of attention weights to ensure that the sum of all weights meets the constraints. The purpose of this process is to correct the distortion in attention allocation caused by the quality of the original behavioral data, and to guide the behavioral attention extraction subnetwork to allocate more attentional resources to behavioral patterns that are more stable and more likely to carry genuine clinical intentions.

[0049] After the above calibration, the first behavior embedding vector output by the behavior attention extraction subnetwork has improved representation quality and can more effectively serve the subsequent inference of implicit diagnostic intentions.

[0050] As an example, the behavior deviation calibration subnetwork performs calibration processing on the intent-behavior dual-stream fusion encoder, including: The intermediate features generated by the behavioral flow and conflict flow are scale aligned or distribution corrected, and the cross-flow offsets that exist during the feature fusion process are compensated, so that the fused joint behavioral latent representation maintains a consistent feature scale and distribution in the latent space, thereby improving the accuracy of generating implicit diagnostic intent vectors.

[0051] Specifically, when the behavior deviation calibration subnetwork calibrates the intent-behavior dual-stream fusion encoder, its processing focuses on feature preparation before fusion and optimization of the fusion process, including: Feature Scale and Distribution Alignment: This behavior bias calibration subnetwork analyzes the first behavior embedding vector (behavior flow) from the behavior attention extraction subnetwork and the second behavior embedding vector (conflict flow) from the local conflict recognition subnetwork. When a significant difference is detected between the two feature flows in numerical scale or statistical distribution (which may be due to the different sensitivities of the two subnetworks to anomalous behavior), the behavior bias calibration subnetwork performs scale scaling or distribution transformation operations to ensure that both have compatible feature scales and similar distribution characteristics before being input into the fusion encoder.

[0052] Cross-stream offset compensation: During the two-stream feature fusion process, the calibration sub-network monitors and identifies potential cross-stream offset phenomena. That is, because the same anomalous behavior is interpreted as different types of features in the two streams, feature points that should correspond semantically are misaligned in the latent space. To address this, the behavior bias calibration sub-network calculates a compensation vector or adjusts the fusion weights to correct the identified offset, ensuring that when the joint behavioral latent representation is generated, the information from the two streams can be effectively aligned and integrated at their positively correlated semantic level.

[0053] For example: The behavior bias calibration subnetwork receives the outputs from the upstream subnetworks, namely the first behavior embedding vector (from the behavior attention extraction subnetwork) and the second behavior embedding vector (from the local conflict recognition subnetwork). For ease of fusion, it is assumed that both have been preprocessed into feature vectors with the same dimension through scale alignment, or have been mapped to the same latent space.

[0054] The behavior bias calibration subnetwork first quantifies the local or global inconsistencies between and, for example, by calculating cross-flow feature difference maps or inconsistency scores. For example, by passing and respectively through lightweight projection layers (such as fully connected layers), two sets of key-value pairs for comparison are obtained: and , with dimensions and , where is the number of feature locations (such as time steps or spatial regions).

[0055] Next, the cross-current attention offset matrix is ​​calculated. : , where is the cosine distance or Euclidean distance. The diagonal elements of should theoretically be small, indicating that features at the same location in the two flows should be aligned. If the value at a location exceeds a preset threshold, it is determined that there is a significant cross-flow offset at that location.

[0056] For the detected offset position, the behavior bias calibration subnetwork generates a compensation vector. This generation process can be implemented using a small neural network (such as a multilayer perceptron MLP): the input to the MLP is the contextual information of the offset position, such as the concatenation of local feature windows. , where and are the value vectors of the position in the two streams, and and reflect the offset relationship between the position and all other positions.

[0057] The compensation vector is calculated using the following formula: , where are the learnable parameters of the MLP. This MLP is trained for prediction, determining which feature vectors need adjustment in which direction to eliminate offsets and make the semantics of and more consistent at their positions. Understandably, the generated compensation vector is applied to the features considered more susceptible to anomalous behavior or with relatively lower credibility.

[0058] For example, if the offset is determined to be mainly caused by abnormal conflict patterns in the data, then the corrected feature is: , where are learnable scalar gating coefficients or weights dynamically calculated based on the offset, used to control the compensation strength. The compensated features will replace the original features in subsequent fusion.

[0059] In addition to directly generating compensation vectors, the behavior bias calibration subnetwork can also choose to dynamically adjust the weights during the fusion process, as follows: In an intent-behavior dual-stream fusion encoder, the standard fusion approach can be weighted summation and / or attention-based fusion. For example, the joint representation can be initially computed as: , where is the original fusion weight.

[0060] The behavior bias calibration subnetwork outputs a weight correction factor based on the offset detection results. This factor can also be calculated by a small network based on the offset context. Specifically, if a significant offset is detected at a location, and the reliability of the determination is reduced due to the behavior bias, the behavior bias calibration subnetwork outputs a negative factor, thereby reducing its contribution to the fusion and relatively increasing the weights. The corrected fusion weights are: , where is the Sigmoid function, used to constrain the weights within the range (0,1). Then, is used for fusion calculation.

[0061] By performing the above calibration on the intent-behavior dual-stream fusion encoder, higher consistency can be ensured within the generated joint behavioral latent representation, reducing information fusion conflicts caused by noise in the input behavioral data. This provides clearer and more reliable input for the downstream directional intent decoder, ultimately improving the accuracy of implicit diagnostic intent vector generation.

[0062] As an example, the main branch of the annotation model shares an encoder structure with multiple intent branches, and outputs the prediction results of the main branch and the prediction results of multiple intent branches respectively during the decoding stage.

[0063] In this embodiment of the invention, the annotation model employs an efficient and parameter-sharing network architecture. In this architecture, the main branch and multiple intent branches are not completely independent networks, but rather share a common encoder structure. This encoder, for example, is a deep convolutional neural network, serving as the backbone of the annotation model and responsible for extracting hierarchical general visual features from the input raw medical images. Understandably, this design ensures consistency in the understanding of underlying features across different branches and significantly reduces the total number of parameters in the model, improving computational efficiency and training stability.

[0064] Following the shared encoder, the annotation model branches during the decoding phase. Specifically, the feature maps output by the shared encoder are fed into multiple independent decoder sub-networks. One decoder acts as the main branch, whose task is to generate a feature map tailored to specific needs based on the shared features. Figure 1 The final labeled prediction result for the region is the basic consensus output of the model. Meanwhile, multiple other decoders constitute different intent branches. Each intent branch's decoder receives the same shared features, but learns to map general features to its corresponding specific diagnostic intent preferences through its unique decoding layer parameters.

[0065] During training, each intent branch, under the dynamic weight supervision of step S40, focuses on fitting its associated intent sub-label. Therefore, in the inference phase, by selecting different intent branches, the model can output annotation results reflecting different clinical strategies such as aggressive screening and robust follow-up.

[0066] This invention, through the aforementioned shared encoding and split decoding architecture, achieves the ability to efficiently and flexibly integrate consensus learning and multi-intent learning in a single model, ensuring both the robustness of the model foundation and the flexibility of output to adapt to different clinical needs.

[0067] Please see Figure 4 This invention also provides a medical data annotation method, applied to a medical data annotation platform, the method comprising: Step S100: Receive the medical image to be labeled; Step S200: Input the medical image to be labeled into the medical data labeling model obtained by the training method according to any of the foregoing embodiments; Step S300: Obtain the annotation results output by the medical data annotation model.

[0068] This method is applied to a medical data annotation platform that integrates the aforementioned trained medical data annotation model. This method utilizes a pre-trained annotation model capable of understanding and separating different diagnostic and treatment intentions to provide intelligent, personalized, and interpretable annotation services for medical images.

[0069] First, the platform receives the medical images to be labeled. These images can be digital images such as CT, MRI, and X-rays from a medical image archiving and communication system. Next, the platform inputs the medical images to be labeled into a medical data labeling model obtained according to the training method described in any of the foregoing embodiments. This model has learned from multi-expert labeled data and their implicit diagnostic and treatment intentions, and is capable of generating consensus-based labels and labels with various intention preferences. Finally, the platform obtains the labeling results output by the medical data labeling model.

[0070] In addition, the solution offers a variety of flexible methods for obtaining annotation results based on different clinical needs or user preferences.

[0071] As an example, obtaining the annotation results includes at least one of the following methods: Obtain the consensus annotation results output by the main branch of the medical data annotation model. Specifically, the platform can directly call the output of the model's main branch. This result represents the most robust and common lesion boundaries or classifications predicted by the model after integrating expert consensus, and is suitable for standardized report generation scenarios that pursue high consistency and reproducibility.

[0072] The platform receives the user's selected treatment intention preferences and obtains the annotation results output by the corresponding intention branch in the medical data annotation model that conform to the treatment intention preferences. Specifically, the platform can receive the user's selected treatment intention preferences; for example, in the user interface, the user can choose an aggressive screening mode or a robust follow-up mode. Based on this selection, the platform activates the corresponding intention branch in the model and obtains its output annotation results. This approach allows the annotation results to proactively adapt to different clinical scenarios or different doctors' review habits, providing personalized auxiliary diagnostic opinions.

[0073] Simultaneously, the platform acquires the consensus annotation results output by the main branch and the specific intent annotation results output by one or more intent branches, and then merges and displays the consensus annotation results with the specific intent annotation results. Specifically, the platform can simultaneously acquire the consensus annotation results output by the main branch and the specific intent annotation results output by one or more intent branches, and then merge and display them. For example, in the visualization interface, the consensus boundary can be used as a baseline, and overlays of different colors or transparency can be used to display the boundary difference areas under different intent strategies. This approach can intuitively present the impact of different clinical strategies on lesion interpretation, helping doctors to fully understand the uncertainty range of lesion characteristics and support more prudent decision-making.

[0074] As an example, the method also includes: Record user interaction behavior during the current annotation process to form new annotation behavior features. Input the new annotation behavior features into the implicit intent inference model associated with the medical data annotation model to update the implicit diagnosis and treatment intent vector. The output of the medical data annotation model is dynamically fine-tuned based on the updated implicit diagnostic intent vector.

[0075] In this implementation, a closed-loop feedback and adaptive process is also set up, as detailed below: When a user (such as a doctor) uses the platform to annotate or review data, the platform records the user's interactive behavior during the annotation process (e.g., how the doctor adjusts the initial boundaries provided by the model, in which areas to focus on modifications, the frequency and pattern of modifications, etc.). These behaviors are encoded into new annotation behavior features in real time.

[0076] The platform inputs these new behavioral features into the implicit intent inference model associated with the medical data annotation model. This implicit intent inference model uses its learned behavior-intent mapping to update the user's implicit diagnostic intent vector based on the user's new behavior. This update process can be either real-time incremental learning or periodic batch optimization.

[0077] Furthermore, the platform dynamically fine-tunes the output of the medical data annotation model based on the updated implicit diagnostic intent vector. For example, it can adjust the weights of each intent branch in the final output to make the output more closely reflect the user's latest strategy preferences; or, it can directly use the intent branch that best matches the updated intent vector as the main output. This mechanism enables the intelligent annotation system to understand and adapt to the specific user's operating habits and clinical judgment tendencies, achieving an evolution from a general model to a personalized auxiliary tool, continuously improving annotation efficiency and user satisfaction.

[0078] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A training method for a medical data annotation model, applied to a medical data annotation platform, characterized in that, The methods and steps include the following: For each medical image in the training set, obtain the annotation data of multiple experts corresponding to that medical image and their annotation behavior characteristics. Based on the labeled behavioral features, behavioral feature representations of each expert are generated through a behavioral coding network, and the behavioral feature representations are input into the implicit intent inference model to obtain implicit diagnostic intent vectors representing the labeling strategies of each expert. Based on the similarity between the implicit diagnostic intent vectors of multiple experts, the expert annotation data is divided into an annotation group with consistent intent and an annotation group with divergent intent. A fusion annotation is generated for the annotation group with consistent intent, and the annotation group with divergent intent is retained as multiple intent sub-labels. The supervision weights are dynamically adjusted based on the implicit treatment intent vectors of each expert, and the annotation model is trained using the fused annotations and the intent sub-labels to obtain a medical data annotation model that can simultaneously learn expert consensus and multiple treatment intent preferences. The annotation model includes a main branch for predicting the annotation results of regions with consistent intent, and multiple intent branches for learning the annotation preferences of different intent sub-labels.

2. The training method for a medical data annotation model according to claim 1, characterized in that: The annotation behavior characteristics include behavioral information collected during the annotation process, including boundary drawing order, local adjustment actions, annotation cancellation behavior, and pause behavior.

3. The training method for a medical data annotation model according to claim 1, characterized in that: The implicit intent inference model includes: The behavior attention extraction subnetwork is used to perform weighted modeling of boundary drawing order, local adjustment action, annotation cancellation behavior and pause behavior based on the importance difference of different behavior types in the labeled behavior features, so as to generate the first behavior embedding vector with behavior intention attention. The local conflict identification subnetwork is used to identify local conflict patterns presented by experts during the annotation process based on the first row of the embedding vector, including boundary expansion trend, boundary contraction trend and repeated confirmation behavior of local regions, and transforms the identification results into the second row of the embedding vector. The intent-behavior dual-stream fusion encoder independently encodes the first-behavior embedding vector and the second-behavior embedding vector, and aligns them across streams in the latent space to generate a joint behavior latent representation. The directional intent decoder, based on the joint behavioral latent representation, performs probabilistic modeling on the multidimensional clinical strategy directions related to expanding, shrinking, sensitive attention, and conservative screening of the annotation region, thereby outputting an implicit diagnostic and treatment intent vector representing the expert annotation strategy.

4. The training method for a medical data annotation model according to claim 3, characterized in that: The implicit intent inference model also includes a behavior bias calibration subnetwork, which is used to calibrate the behavior attention extraction subnetwork and the intent-behavior dual-stream fusion encoder in the implicit intent inference model. By reweighting the feature channels corresponding to abnormal behaviors, normalizing the features of offset behaviors, or suppressing noise in unstable behavior segments, the accuracy of generating implicit diagnostic intent vectors can be improved.

5. The training method for a medical data annotation model according to claim 4, characterized in that: The behavior bias calibration subnetwork calibrates the behavior attention extraction subnetwork, including: Based on the degree of deviation of the labeled behavioral features, the attention weights of each behavioral feature channel in the behavioral attention extraction subnetwork are recalculated to suppress abnormal attention responses caused by excessively fast drawing, jittery trajectories, or abnormal pauses. The attention weights are also renormalized to improve the stability of the first behavioral embedding vector output by the behavioral attention extraction subnetwork.

6. The training method for a medical data annotation model according to claim 5, characterized in that: The behavior deviation calibration subnetwork performs calibration processing on the intent-behavior dual-stream fusion encoder, including: The intermediate features generated by the behavioral flow and conflict flow are scale aligned or distribution corrected, and the cross-flow offsets that exist during the feature fusion process are compensated, so that the fused joint behavioral latent representation maintains a consistent feature scale and distribution in the latent space, thereby improving the accuracy of generating implicit diagnostic intent vectors.

7. A training method for a medical data annotation model according to any one of claims 1-6, characterized in that: The main branch and multiple intent branches of the annotation model share the encoder structure, and the main branch prediction result and multiple intent branch prediction results are output respectively during the decoding stage.

8. A medical data annotation method, applied to a medical data annotation platform, characterized in that: The method includes: Receive medical images to be labeled; The medical image to be labeled is input into the medical data labeling model obtained by the training method according to any one of claims 1-7; Obtain the annotation results output by the medical data annotation model.

9. A medical data annotation method according to claim 8, characterized in that: The method of obtaining the annotation results includes at least one of the following: Obtain the consensus annotation results output by the main branch of the medical data annotation model; Receive the user's selected treatment intention preference, and obtain the annotation results that conform to the treatment intention preference output by the corresponding intention branch in the medical data annotation model; Simultaneously, the consensus annotation result output by the main branch and the specific intent annotation result output by one or more intent branches are obtained, and the consensus annotation result and the specific intent annotation result are merged and displayed.

10. A medical data annotation method according to claim 8, characterized in that: The method further includes: Record user interaction behavior during the current annotation process to form new annotation behavior features. Input the new annotation behavior features into the implicit intent inference model associated with the medical data annotation model to update the implicit diagnosis and treatment intent vector. The output of the medical data annotation model is dynamically fine-tuned based on the updated implicit diagnostic intent vector.

Images