A web service-based cloud integration software running support method and system

Abstract

The application discloses a cloud integration software running support method and system based on a Web service, relates to the technical field of medical collaborative diagnosis, and comprises the following steps: constructing a double-channel event collection pipeline, capturing two types of interactive actions and converting the two types of interactive actions into structured atomic events; constructing a shared causal decoupling representation space, encoding two types of event sequences and decoupling the two types of event sequences into invariant attribute encoding and causal attribute encoding; introducing a structural causal model, constructing a directed acyclic graph containing operation nodes and entity nodes, and establishing a causal dependence direction; based on a counterfactual attention mechanism, constructing fact and counterfact context states, and quantifying the causal effect strength between operation events and modification events by comparison; generating a causal correlation link, screening strong causal pairs according to the causal strength to construct a traceability link, and solidifying the correlation data and the original flow hash timestamp; when there is a diagnostic conclusion difference in a collaborative session, the system can automatically backtrack and locate which specific operation is the root cause of the conclusion modification.

Description

Technical Field

[0001] This invention belongs to the field of medical collaborative diagnosis technology, and in particular relates to a cloud-based integrated software operation support method and system based on Web services. Background Technology

[0002] In cross-hospital collaborative diagnosis of medical images, users at each end of the conversation need to perform interactive operations such as adjusting window width and level, zooming and panning, and measuring and annotating the target image, while editing and modifying key medical entities in the diagnostic report. Existing technical solutions generally adopt timestamp alignment or event log sequential recording to collect and store the image interaction operation flow and the report text editing flow separately, and reproduce the operation sequence and the order of text modification according to the timeline when tracing back.

[0003] However, time-series-based recording methods have inherent flaws: when a user overturns an existing diagnosis and modifies the report entity through specific image operations, existing technologies can only record the temporal relationship between the operation and editing events, failing to automatically establish a causal link between them. The causal link between the operation flow and the text flow is broken. This break makes it easy for a large number of statistical co-occurrences between operation and editing events to be misjudged as causal relationships by time-window-based alignment, resulting in false associations. Furthermore, it leads to a lack of interpretable causal evidence chains between the report entity state before and after modification and the corresponding image operation sequence. The untraceability and non-repudiation of collaborative diagnostic records cannot be technically guaranteed. Therefore, we propose a cloud-integrated software operation support method and system based on Web services. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a cloud-based integrated software operation support method and system based on Web services, comprising the following steps: Step S1: Construct a dual-channel event stream acquisition and standardization pipeline. After the collaborative work session starts, simultaneously establish the first event acquisition channel and the second event acquisition channel to capture the first type of interactive action of image view interaction class and the second type of interactive action of report entity editing class, and convert them into structured atomic operation events. Step S2: Construct a shared causal decoupling intermediate representation space, encode the state of the first type of event sequence through the first encoder, and encode the semantic entities in the second type of event through the second encoder. Under the constraint of the orthogonal decoupling loss function, decouple the encoding result into two parts: invariant attribute encoding and causal attribute encoding. Step S3: Introduce a structural causal model to define a prior knowledge graph, construct a directed acyclic graph containing nodes of the first type of operation and nodes of the second type of semantic entity, clarify the direction of physical causal dependency between nodes, and prohibit reverse causal edges from semantic entity nodes to operation nodes. Step S4: Causal link discovery based on counterfactual attention mechanism. When a second type of modification event is captured, construct the fact global context state and the counterfactual global context state; apply do intervention to each first type of event to remove the corresponding causal path and generate the counterfactual global context state; quantify the causal effect strength between the first type of operation event and the second type of modification event by comparing the two types of context states. Step S5: Generate causal relationship links and solidify consensus versions. Select strong causal pairs based on the strength of causal effects, construct causal tracing links, and solidify the correlation description data and the original data stream into an immutable final version of the collaborative session after hash processing and adding a trusted timestamp.

[0005] Further, step S1 includes the following steps: Step S11: The first event acquisition channel acquires and parses the first type of interactive action, standardizes the first type of interactive action into a first atomic operation event, and defines a single first atomic operation event as a quadruple. In the formula For event timestamps, It is an element in the predefined set of first-class operation types. It is the dictionary of operation parameters corresponding to the first type of operation. It is the set of boundary parameters of the current viewport in the coordinate system of the target object, and also contains the unique identifier of the operating user; Step S12: The second event acquisition channel captures the editing action in the second type of interaction, and converts the editing action into a second atomic operation event. A single second atomic operation event is defined as a quadruple. In the formula For event timestamps, It is an element in the predefined set of second-class operation types. This refers to the affected semantic entity index, which is provided by a pre-built domain entity recognizer. It is an assertion state change encoding of semantic entities, and also contains the unique identifier of the editing user.

[0006] Furthermore, step S2 includes the following steps: Step S21: Design an operation status encoder The first type of event sequence within a continuous time period is input into the time series modeling network to obtain the hidden state. Furthermore, causal decoupling of the hidden state is achieved by forcibly decomposing the hidden state into two independent parts through an orthogonal constraint loss function: In the formula This represents the encoding of invariant physical attributes that are independent of the target semantics in the first type of operation. This represents the encoding of causal attributes that can alter perceptions of a target object; Step S22: Use an encoder Obtain the semantic state of key entities in the second type of event. And decouple semantic state to In the formula Encoding of invariant attributes in semantic concepts that are unrelated to the cognition of the target object. The encoding of the core causal features corresponding to the semantic concept.

[0007] Furthermore, step S3 includes the following steps: Step S31: Define the node set The node set contains all possible nodes of type I operation. Second type of semantic entity category node ; Step S32: Define the set of directed edges A directed edge set clearly defines the direction of physical causal dependencies between nodes, forming a path from... point to and the directed acyclic graph of intermediate state nodes Directed acyclic graphs explicitly prohibit reverse causal edges from semantic entity nodes to operation nodes, and restrict the propagation direction of edges to follow the hierarchical logic of operation-feature observation-diagnosis conclusion.

[0008] Furthermore, step S4 includes the following steps: Step S41: For time... Encode the causal attributes of all current first-type event sequences. And the causal attribute encoding of all current second-class entities Integrate into a single fact-based global context state ; Step S42: Based on the structural causal model defined in step S3 For any event of the first kind , implementation The operation propagates downstream along the topological order of the directed acyclic graph, blocking all causal paths affected by the event, and generates a counterfactual global context state through a pre-trained prediction network. ; Step S43: By comparing the factual global context state With counterfactual global context state Calculate the first type of event For the second type of entity causal effect score The causal effect score is defined as the ratio of the predicted match degree of the target entity state under factual conditions to the predicted match degree of the target entity state under counterfactual conditions, and the calculation formula is: In the formula, The cosine similarity function is used. For the second type of entity at time The causal attribute encoding, For a moment The factual global context state, For a moment The counterfactual global context state, molecule Measurement without removing the first type of event The predictive matching degree of the global context state to the target entity state, denominator Metrics in removing the first type of event The prediction matching degree of the post-global context state to the target entity state, and the causal effect score. The larger the value, the more likely it is a Type I event. With the second type of entity The stronger the causal relationship between state changes.

[0009] Furthermore, step S5 includes the following steps: Step S51: Based on the causal effect scores calculated in step S4 Values ​​that exceed a preset threshold are filtered out. Strong causal pairs are identified by taking the ultimately modified second-type entity as the root node, tracing back upstream, iteratively selecting the first-type event with the highest causal effect score as the parent node, and connecting them in chronological order to form a tree-like or chain-like causal origin graph, and labeling the operation user identity corresponding to each link. Step S52: Generate correlation description data. The correlation description data includes the state information of the entities before and after the modification, as well as the sequence information of the first type of event with the highest causal effect score associated with the entities before and after the modification. The correlation description data and the original data stream are processed by the SHA-256 hash algorithm and timestamped to solidify the final version of the collaborative session that is tamper-proof and traceable.

[0010] Furthermore, in step S42, the training objective of the pre-trained prediction network during the training phase is to minimize the prediction error between the fact global context state and the counterfact global context state on known causal edges. The training samples consist of operation sequences and entity modification records with known causal relationships. The prediction network adopts a two-layer fully connected structure, with the input being the set of causal attribute encodings after removing the target operation and the output being the counterfact global context state vector.

[0011] A cloud-based integrated software operation support system based on Web services includes an event acquisition and standardization module, a causal decoupling representation encoding module, a structural causal model management module, a counterfactual attention calculation module, a causal link generation and solidification module, a collaborative session management module, and a persistent storage module. The event acquisition and standardization module is used to independently acquire the first type of interactive action of image view interaction class and the second type of editing action of report entity editing class, and convert them into standardized atomic operation event quadruples, with timestamps, viewport parameters and user identity identifiers attached; The causal decoupling representation encoding module is used to perform temporal modeling on the first type of event sequence and generate hidden states; under the orthogonal decoupling loss constraint, causal decoupling is performed on the hidden states, forcibly decomposing them into invariant attribute encoding and causal attribute encoding; the semantic entities in the second type of events are also encoded and decoupled, providing operable vectorized inputs for subsequent causal calculations; The structural causal model management module is responsible for storing a directed acyclic graph composed of nodes of the first type of operation and nodes of the second type of semantic entity; defining and managing the directed edges between nodes, explicitly specifying the causal dependency direction from the operation node to the semantic entity node, prohibiting reverse causal edges, and providing topological constraints that conform to physical laws for counterfactual intervention. The counterfactual attention calculation module is used to integrate causal attribute encoding to construct a fact global context state; based on the constraints of the structural causal model, it applies do intervention to the target operation and propagates along the topology to execute the counterfactual intervention operation and generate a counterfactual global context state; by comparing the prediction matching degree of the two states to the target entity, it calculates the causal effect score between the first type of operation event and the second type of entity modification. The causal link generation and solidification module is used to filter out strong causal pairs that exceed the threshold based on the causal effect score; to construct a causal origin graph by tracing back upwards from the modified entity; to generate associated description data containing state information before and after modification and key operation sequences; and to perform SHA-256 hash operation on the above data and add a timestamp to complete the tamper-proof solidification. The collaborative session management module is responsible for creating, joining, leaving and destroying sessions; maintaining the global context state of the sessions; transmitting control signals between processing stages; triggering the start and termination of event acquisition and calculation processes; and coordinating the collaborative work of other modules. The persistent storage module is used to persistently save the original atomic event stream, intermediate representation state vectors, structural causal model diagram, calculated causal effect scores, the final generated causal link diagram, and solidified collaborative session version data for post-audit or model optimization.

[0012] The present invention has the following beneficial effects: 1. This invention constructs a shared causal decoupling intermediate representation space, mapping the first type of interactive actions and the second type of editing actions to causal attribute codes respectively. Under the constraints of a structural causal model, counterfactual intervention calculations are performed, establishing for the first time a quantifiable causal effect strength between the two types of events. When a collaborative session experiences a discrepancy in diagnostic conclusions, the system can automatically trace back and pinpoint the specific operation that caused the conclusion modification, rather than simply presenting the chronological order of the operations. This fundamentally solves the problem of the lack of causal explanatory power in existing technologies regarding correlation links.

[0013] 2. This invention introduces a structural causal model as a prior knowledge graph, explicitly defining the causal dependency direction from operation nodes to semantic entity nodes and prohibiting the existence of reverse causal edges. When calculating causal effect scores, counterfactual intervention is implemented to simulate the semantic state change after removing a certain operation, effectively blocking false associations introduced by confounding factors through backdoor paths. This ensures that the ultimately selected strong causal pairs have physical interpretability, rather than simply statistical correlation.

[0014] 3. After automatically discovering causal links, this invention generates correlation description data containing entity state information before and after modification, as well as key causal operation sequences. This correlation description data, along with the original data stream, is then hashed and timestamped. Since the tracing link and the original data together constitute a final version with cryptographic integrity, any subsequent tampering with the operation record or diagnostic conclusion will result in hash verification failure, thus ensuring the non-repudiation and reliability of the collaborative session records.

[0015] 4. This invention introduces a domain prior knowledge graph at the structural causal model level, requiring only a lightweight causal graph structure definition. During the training phase, the prediction network only needs to optimize the prediction error on known causal edges, and the difficulty of obtaining training samples is significantly lower than that of large-scale dense labeling. At the same time, the calculation of causal effect scores depends on the comparison between factual and counterfactual states. The core calculation process does not rely on the end-to-end fitting of black-box models, and has stronger generalization ability in scenarios with unseen operation types and entity combinations.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a cloud-based integrated software operation support method based on Web services according to the present invention. Figure 2 This is a schematic diagram of the architecture of a cloud-based integrated software operation support system based on Web services according to 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 1-2 As shown, the present invention is a cloud-based integrated software operation support method and system based on Web services, comprising the following steps: Step S1: Construct a dual-channel event stream acquisition and standardization pipeline. After the collaborative work session starts, simultaneously establish the first event acquisition channel and the second event acquisition channel to capture the first type of interactive action of image view interaction class and the second type of interactive action of report entity editing class, and convert them into structured atomic operation events. Step S2: Construct a shared causal decoupling intermediate representation space, encode the state of the first type of event sequence through the first encoder, and encode the semantic entities in the second type of event through the second encoder. Under the constraint of the orthogonal decoupling loss function, decouple the encoding result into two parts: invariant attribute encoding and causal attribute encoding. Step S3: Introduce a structural causal model to define a prior knowledge graph, construct a directed acyclic graph containing nodes of the first type of operation and nodes of the second type of semantic entity, clarify the direction of physical causal dependency between nodes, and prohibit reverse causal edges from semantic entity nodes to operation nodes. Step S4: Causal link discovery based on counterfactual attention mechanism. When a second type of modification event is captured, construct the fact global context state and the counterfactual global context state; apply do intervention to each first type of event to remove the corresponding causal path and generate the counterfactual global context state; quantify the causal effect strength between the first type of operation event and the second type of modification event by comparing the two types of context states. Step S5: Generate causal relationship links and solidify consensus versions. Select strong causal pairs based on the strength of causal effects, construct causal tracing links, and solidify the correlation description data and the original data stream into an immutable final version of the collaborative session after hash processing and adding a trusted timestamp.

[0021] Step S1 includes the following steps: Step S11: The first event acquisition channel acquires and parses the first type of interactive action, standardizes the first type of interactive action into the first atomic operation event, and defines a single first atomic operation event as a quadruple. In the formula For event timestamps, It is an element in the predefined set of first-class operation types. It is the dictionary of operation parameters corresponding to the first type of operation. It is the set of boundary parameters of the current viewport in the coordinate system of the target object, and also contains the unique identifier of the operating user; Step S12: The second event acquisition channel captures the editing action in the second type of interaction, and converts the editing action into a second atomic operation event. A single second atomic operation event is defined as a quadruple. In the formula For event timestamps, It is an element in the predefined set of second-class operation types. This refers to the affected semantic entity index, which is provided by a pre-built domain entity recognizer. It is an assertion state change encoding of semantic entities, and also contains the unique identifier of the editing user.

[0022] Step S2 includes the following steps: Step S21: Design an operation status encoder The first type of event sequence within a continuous time period is input into the time series modeling network to obtain the hidden state. Furthermore, causal decoupling of the hidden state is achieved by forcibly decomposing the hidden state into two independent parts through an orthogonal constraint loss function: In the formula This represents the encoding of invariant physical attributes that are independent of the target semantics in the first type of operation. This represents the encoding of causal attributes that can alter perceptions of a target object; Step S22: Use an encoder Obtain the semantic state of key entities in the second type of event. And decouple semantic state to In the formula Encoding of invariant attributes in semantic concepts that are unrelated to the cognition of the target object. The encoding of the core causal features corresponding to the semantic concept.

[0023] Step S3 includes the following steps: Step S31: Define the node set The node set contains all possible nodes of type I operation. Second type of semantic entity category node ; Step S32: Define the set of directed edges A directed edge set clearly defines the direction of physical causal dependencies between nodes, forming a path from... point to and the directed acyclic graph of intermediate state nodes Directed acyclic graphs explicitly prohibit reverse causal edges from semantic entity nodes to operation nodes, and restrict the propagation direction of edges to follow the hierarchical logic of operation-feature observation-diagnosis conclusion.

[0024] Step S4 includes the following steps: Step S41: For time... Encode the causal attributes of all current first-type event sequences. And the causal attribute encoding of all current second-class entities Integrate into a single fact-based global context state ; Step S42: Based on the structural causal model defined in step S3 For any event of the first kind , implementation The operation propagates downstream along the topological order of the directed acyclic graph, blocking all causal paths affected by the event, and generates a counterfactual global context state through a pre-trained prediction network. ; Step S43: By comparing the factual global context state With counterfactual global context state Calculate the first type of event For the second type of entity causal effect score The causal effect score is defined as the ratio of the predicted match degree of the target entity state under factual conditions to the predicted match degree of the target entity state under counterfactual conditions, and the calculation formula is: In the formula, The cosine similarity function is used. For the second type of entity at time The causal attribute encoding, For a moment The factual global context state, For a moment The counterfactual global context state, molecule Measurement without removing the first type of event The predictive matching degree of the global context state to the target entity state, denominator Metrics in removing the first type of event The prediction matching degree of the post-global context state to the target entity state, and the causal effect score. The larger the value, the more likely it is a Type I event. With the second type of entity The stronger the causal relationship between state changes.

[0025] Step S5 includes the following steps: Step S51: Based on the causal effect scores calculated in step S4 Values ​​that exceed a preset threshold are filtered out. Strong causal pairs are identified by taking the ultimately modified second-type entity as the root node, tracing back upstream, iteratively selecting the first-type event with the highest causal effect score as the parent node, and connecting them in chronological order to form a tree-like or chain-like causal origin graph, and labeling the operation user identity corresponding to each link. Step S52: Generate correlation description data. The correlation description data includes the state information of the entities before and after the modification, as well as the sequence information of the first type of event with the highest causal effect score associated with the entities before and after the modification. The correlation description data and the original data stream are processed by the SHA-256 hash algorithm and timestamped to solidify the final version of the collaborative session that is tamper-proof and traceable.

[0026] In step S42, the training objective of the pre-trained prediction network during the training phase is to minimize the prediction error between the fact global context state and the counterfact global context state on the known causal edge. The training samples consist of operation sequences and entity modification records with known causal relationships. The prediction network adopts a two-layer fully connected structure, with the input being the set of causal attribute encodings after removing the target operation and the output being the counterfact global context state vector.

[0027] A cloud-based integrated software operation support system based on Web services includes an event acquisition and standardization module, a causal decoupling representation encoding module, a structural causal model management module, a counterfactual attention calculation module, a causal link generation and solidification module, a collaborative session management module, and a persistent storage module. The event acquisition and standardization module is used to independently acquire the first type of interactive actions of the image view interaction class and the second type of editing actions of the report entity editing class, and convert them into standardized atomic operation event quadruples, with timestamps, viewport parameters and user identity identifiers attached; The causal decoupling representation encoding module is used to perform temporal modeling on the first type of event sequence and generate hidden states; under the orthogonal decoupling loss constraint, causal decoupling is performed on the hidden states, forcibly decomposing them into invariant attribute encoding and causal attribute encoding; the semantic entities in the second type of events are also encoded and decoupled, providing operable vectorized inputs for subsequent causal calculations; The structural causal model management module is responsible for storing the directed acyclic graph composed of nodes of the first type of operation and nodes of the second type of semantic entity; it defines and manages the directed edges between nodes, explicitly specifies the causal dependency direction from operation nodes to semantic entity nodes, prohibits reverse causal edges, and provides topological constraints that conform to physical laws for counterfactual intervention. The counterfactual attention calculation module is used to integrate causal attribute encoding and construct a factual global context state; based on the constraints of the structural causal model, it applies do intervention to the target operation and propagates along the topology to execute counterfactual intervention operations and generate a counterfactual global context state; by comparing the prediction matching degree of the two states to the target entity, it calculates the causal effect score between the first type of operation event and the second type of entity modification. The causal link generation and solidification module is used to filter out strong causal pairs that exceed the threshold based on the causal effect score; starting from the modified entity, it traces back upwards to build a causal origin graph; it generates associated description data containing state information before and after modification and key operation sequences; it performs SHA-256 hash operation on the above data and adds a timestamp to complete the tamper-proof solidification. The collaborative session management module is responsible for creating, joining, leaving and destroying sessions; maintaining the global context state of sessions; transmitting control signals between processing stages; triggering the start and termination of event acquisition and calculation processes; and coordinating the collaborative work of other modules. The persistent storage module is used to persistently save the original atomic event stream, intermediate representation state vectors, structural causal model diagram, calculated causal effect scores, the final generated causal link diagram, and solidified collaborative session version data for post-event auditing or model optimization.

[0028] One specific application of this embodiment is: This embodiment utilizes cloud-based integrated collaborative diagnostic software deployed on a web service architecture. The participants in the session include a radiologist's terminal from Hospital A and a radiology expert's terminal from Hospital B. Both parties synchronously access cloud-based image data and diagnostic reports via browsers, collaboratively completing lung CT scan interpretation and report writing. The session fully utilizes dual-channel event acquisition, causal decoupling representation, counterfactual causal calculation, and version persistence capabilities. The specific implementation steps are as follows: Step S1: Construct a dual-channel event stream acquisition and standardized pipeline After the collaborative work session is started, the cloud service simultaneously establishes the first event acquisition channel and the second event acquisition channel to capture the interactive actions of the image view and the editing actions of the diagnostic report, and converts them into structured atomic operation events.

[0029] Step S11: The first event acquisition channel performs full data tracking on mouse and keyboard interactions on the web-based image canvas, analyzing first-type interactive actions such as window width and level adjustment, zooming and panning, measurement annotation, and region of interest cropping, and standardizing them as first-atomic operation events. A single first-atomic operation event is defined as a quadruple: In the formula This is the event trigger timestamp (unit: seconds, with the session start time recorded as 0). For elements in the predefined first type of operation set, To operate on the corresponding parameter dictionary, This is the set of boundary parameters for the current viewport in the image coordinate system.

[0030] In this embodiment, a lung window adjustment operation is triggered at 12.3 seconds after the session starts, and the corresponding atomic event is: A nodule diameter measurement operation is triggered 14.1 seconds after the session starts; the corresponding atomic event is: Step S12: The second event acquisition channel captures text modification actions in the diagnostic report editor, locates the affected semantic entities using a pre-set lung medical entity recognizer, and converts the editing action into a second atomic operation event. A single second atomic operation event is defined as a quadruple: In the formula For event timestamps, For elements in the predefined set of second type of operation, A unique index for the affected semantic entities. Encodes the assertion state changes of semantic entities.

[0031] In this embodiment, 15.7 seconds after the session starts, the expert from Hospital B modifies the diagnosis of the lung nodule numbered 001 in the report, changing it from a benign nodule to a suspected malignant nodule. The corresponding atomic event is: Step S2: Construct a shared causal decoupling intermediate representation space By using dual encoders to learn representations of two types of event sequences, and forcibly decoupling the encoding results into two parts, invariant attribute encoding and causal attribute encoding, heterogeneous events are mapped to a unified intermediate representation space.

[0032] Step S21: Design the operation status encoder It employs a single-layer Transformer temporal modeling network, taking a sequence of events of the first class as input and outputting a 256-dimensional hidden state. By adding an orthogonal decoupling loss function as a constraint, the hidden state is forced to decompose into two independent 128-dimensional sub-vectors: In the formula Encodes the invariant physical attributes for the first type of operation, corresponding to attributes unrelated to diagnostic semantics, such as viewport position and mouse trajectory; It is a causal attribute encoding, corresponding to operational features that can change the results of image observation and thus affect the diagnostic judgment.

[0033] In this embodiment, the timing of the modified event is... The causal attribute encoding of all historical events of the first type It is a 128-dimensional vector with an L2 norm of 0.92.

[0034] Step S22: Employ a semantic entity encoder A lightweight BERT network finely tuned for the medical field is used to extract the semantic state of entities in the second type of event. The dimension is 256; similarly, it is split into two 128-dimensional sub-vectors through decoupling constraints: In the formula Encoding the invariant attributes of semantic concepts corresponds to fixed attributes such as the anatomical location and unique number of the entity; Encoding core causal features corresponds to the entity's diagnostic status, quantitative indicators, and other features that can be affected by image manipulation.

[0035] In this embodiment, the target entity at the modification time The causal attribute is encoded as .

[0036] Step S3: Introduce a structural causal model and define a prior knowledge graph. Based on domain prior knowledge of lung imaging diagnosis, a directed acyclic graph containing operation type nodes and semantic entity category nodes is constructed to clarify the physical causal dependency direction between nodes, providing topological constraints for counterfactual computation.

[0037] Step S31: Define the node set The first type of operation node set It includes four types of nodes: window width adjustment, scaling and translation, measurement annotation, and region of interest clipping; and a second type of semantic entity category node set. It includes four types of nodes: nodule size, nodule density, nodule margin, and diagnostic conclusion.

[0038] Step S32: Define the set of directed edges Based on the physical logic that image manipulation alters observed features and these features influence diagnostic conclusions, directed edges are established from operation nodes to entity nodes and from lower-level features to higher-level conclusions: This directed acyclic graph explicitly prohibits reverse causal edges from semantic entity nodes to operation nodes, structurally avoiding false associations with reversed causality.

[0039] Step S4: Causal link discovery based on counterfactual attention mechanism When captured The second type of modification event At that time, the system triggers the causal effect calculation process to quantify the causal driving strength of each first-type event in modifying the entity.

[0040] Step S41: Construct the global context state of the facts. Retrieve the time when the modification event occurred. The causal attribute codes of all first-type events are concatenated and fused with the causal attribute codes of all second-type entities to obtain the global context state of the facts: In this embodiment It is a 256-dimensional fusion vector.

[0041] Step S42: Generate counterfactual global context state. Based on structural causal model. Topological constraints, for the first type of events to be verified. (Window width adjustment operation), perform intervention operation. This involves shielding the causal path corresponding to the operation and propagating the intervention's impact along the topology of the directed acyclic graph; then, through a pre-trained prediction network, the counterfactual global context state after removing the operation is obtained. .

[0042] The pre-trained prediction network uses a 2-layer fully connected network. During the training phase, 12,000 image operation-report modification annotation samples with known causal relationships are used as the training set. The training objective is to minimize the mean squared prediction error of factual and counterfactual states on known causal edges. After training convergence, the validation set loss is 0.021.

[0043] Step S43: Calculate the causal effect score using the cosine similarity function. The matching degree of the factual and counterfactual contexts to the causal encoding of the target entity is calculated separately, and the ratio is taken after exponential normalization to obtain the first type of event. For the second type of entity Causal effect score: In the formula, the numerator measures the prediction matching degree of the global context to the target entity state when the first type of event is not removed, and the denominator measures the prediction matching degree of the global context to the target entity state after the first type of event is removed. The larger the value, the stronger the causal relationship between the two.

[0044] In this embodiment, for (Window width adjustment) and (Revised Diagnosis) Similarity under factual conditions: ; Similarity under counterfactual conditions: ; Substituting into the formula, we get: Similarly, the measurement annotation operation is calculated. Modified causal effect score for this entity The causal effect score for scaling and translation operations is approximately 1.08.

[0045] Step S5: Generate causal links and solidify consensus version Strong causal pairs are selected based on the strength of causal effects, a traceable causal tracing link is constructed, and data is solidified and protected against tampering through hash operations and timestamps.

[0046] Step S51: Construct a causal chain. Preset causal effect score thresholds. Filter out strong causal pairs that exceed the threshold: , Starting with the modified entity as the root node, the process traces upstream along the timeline, iteratively selecting the operation with the highest causal effect score as the parent node, and chaining them together in chronological order to form a chain-like causal origin graph. Step S52: Solidify the final version of the collaborative session. Generate correlation description data, including the target entity's pre-modification state (benign nodule), post-modification state (suspected malignant nodule), the operation sequence and timing information of causal relationships, and the causal effect score corresponding to each link. Merge the correlation description data with the original dual-channel event stream data, perform digest calculation using the SHA-256 hash algorithm, and add a trusted timestamp, finally solidifying it into an immutable collaborative session version; The solidified data is stored in a persistent storage module, supporting post-event auditing and integrity verification; any subsequent tampering with the operation records or report content will cause the hash verification to fail, ensuring the non-repudiation of collaborative records.

[0047] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0048] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A cloud-integrated software operation support method based on Web services, characterized in that, Includes the following steps: Step S1: Construct a dual-channel event stream acquisition and standardization pipeline. After the collaborative work session starts, simultaneously establish the first event acquisition channel and the second event acquisition channel to capture the first type of interactive action of image view interaction class and the second type of interactive action of report entity editing class, and convert them into structured atomic operation events. Step S2: Construct a shared causal decoupling intermediate representation space, encode the state of the first type of event sequence through the first encoder, and encode the semantic entities in the second type of event through the second encoder. Under the constraint of the orthogonal decoupling loss function, decouple the encoding result into two parts: invariant attribute encoding and causal attribute encoding. Step S3: Introduce a structural causal model to define a prior knowledge graph, construct a directed acyclic graph containing nodes of the first type of operation and nodes of the second type of semantic entity, clarify the direction of physical causal dependency between nodes, and prohibit reverse causal edges from semantic entity nodes to operation nodes. Step S4: Discover causal links based on the counterfactual attention mechanism. When a second type of modification event is captured, construct the fact global context state and the counterfactual global context state. For each Type I event, apply the do intervention to remove the corresponding causal path and generate a counterfactual global context state; By comparing two types of context states, the strength of the causal effect between the first type of operation event and the second type of modification event is quantified. Step S5: Generate causal relationship links and solidify consensus versions. Select strong causal pairs based on the strength of causal effects, construct causal tracing links, and solidify the correlation description data and the original data stream into an immutable final version of the collaborative session after hash processing and adding a trusted timestamp.

2. The cloud-based integrated software operation support method based on Web services according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: The first event acquisition channel acquires and parses the first type of interactive action, standardizes the first type of interactive action into a first atomic operation event, and defines a single first atomic operation event as a quadruple. In the formula For event timestamps, It is an element in the predefined set of first-class operation types. It is the dictionary of operation parameters corresponding to the first type of operation. It is the set of boundary parameters of the current viewport in the coordinate system of the target object, and also contains the unique identifier of the operating user; Step S12: The second event acquisition channel captures the editing action in the second type of interaction, and converts the editing action into a second atomic operation event. A single second atomic operation event is defined as a quadruple. In the formula For event timestamps, It is an element in the predefined set of second-class operation types. This refers to the affected semantic entity index, which is provided by a pre-built domain entity recognizer. It is an assertion state change encoding of semantic entities, and also contains the unique identifier of the editing user.

3. The cloud-based integrated software operation support method based on Web services according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Design an operation status encoder The first type of event sequence within a continuous time period is input into the time series modeling network to obtain the hidden state. Furthermore, causal decoupling of the hidden state is achieved by forcibly decomposing the hidden state into two independent parts through an orthogonal constraint loss function: In the formula This represents the encoding of invariant physical attributes that are independent of the target semantics in the first type of operation. This represents the encoding of causal attributes that can alter perceptions of a target object; Step S22: Use an encoder Obtain the semantic state of key entities in the second type of event. And decouple semantic state to In the formula Encoding of invariant attributes in semantic concepts that are unrelated to the cognition of the target object. The encoding of the core causal features corresponding to the semantic concept.

4. The cloud-based integrated software operation support method based on Web services according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Define the node set The node set contains all possible nodes of type I operation. Second type of semantic entity category node ; Step S32: Define the set of directed edges A directed edge set clearly defines the direction of physical causal dependencies between nodes, forming a path from... point to and the directed acyclic graph of intermediate state nodes Directed acyclic graphs explicitly prohibit reverse causal edges from semantic entity nodes to operation nodes, and restrict the propagation direction of edges to follow the hierarchical logic of operation-feature observation-diagnosis conclusion.

5. The cloud-based integrated software operation support method based on Web services according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: For time... Encode the causal attributes of all current first-type event sequences. And the causal attribute encoding of all current second-class entities Integrate into a single fact-based global context state ; Step S42: Based on the structural causal model defined in step S3 For any event of the first kind , implementation The operation propagates downstream along the topological order of the directed acyclic graph, blocking all causal paths affected by the event, and generates a counterfactual global context state through a pre-trained prediction network. ; Step S43: By comparing the factual global context state With counterfactual global context state Calculate the first type of event For the second type of entity causal effect score The causal effect score is defined as the ratio of the predicted match degree of the target entity state under factual conditions to the predicted match degree of the target entity state under counterfactual conditions, and the calculation formula is: In the formula, The cosine similarity function is used. For the second type of entity at time The causal attribute encoding, For a moment The factual global context state, For a moment The counterfactual global context state, molecule Measurement without removing the first type of event The predictive matching degree of the global context state to the target entity state, denominator Metrics in removing the first type of event The prediction matching degree of the post-global context state to the target entity state, and the causal effect score. The larger the value, the more likely it is a Type I event. With the second type of entity The stronger the causal relationship between state changes.

6. The cloud-based integrated software operation support method based on Web services according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Based on the causal effect scores calculated in step S4 Values ​​that exceed a preset threshold are filtered out. Strong causal pairs are identified by taking the ultimately modified second-type entity as the root node, tracing back upstream, iteratively selecting the first-type event with the highest causal effect score as the parent node, and connecting them in chronological order to form a tree-like or chain-like causal origin graph, and labeling the operation user identity corresponding to each link. Step S52: Generate correlation description data. The correlation description data includes the state information of the entities before and after the modification, as well as the sequence information of the first type of event with the highest causal effect score associated with the entities before and after the modification. The correlation description data and the original data stream are processed by the SHA-256 hash algorithm and timestamped to solidify the final version of the collaborative session that is tamper-proof and traceable.

7. The cloud-based integrated software operation support method based on Web services according to claim 5, characterized in that, In step S42, the training objective of the pre-trained prediction network during the training phase is to minimize the prediction error between the fact global context state and the counterfact global context state on known causal edges. The training samples consist of operation sequences and entity modification records with known causal relationships. The prediction network adopts a two-layer fully connected structure, with the input being the set of causal attribute codes after removing the target operation and the output being the counterfact global context state vector.

8. A cloud-integrated software runtime support system based on Web services, characterized in that, It includes an event acquisition and standardization module, a causal decoupling representation encoding module, a structural causal model management module, a counterfactual attention calculation module, a causal link generation and solidification module, a collaborative session management module, and a persistent storage module; The event acquisition and standardization module is used to independently acquire the first type of interactive action of image view interaction class and the second type of editing action of report entity editing class, and convert them into standardized atomic operation event quadruples, with timestamps, viewport parameters and user identity identifiers attached; The causal decoupling representation encoding module is used to perform temporal modeling on the first type of event sequence and generate hidden states; under the orthogonal decoupling loss constraint, causal decoupling is performed on the hidden states to forcibly decompose invariant attribute encoding and causal attribute encoding. The semantic entities in the second type of event are also encoded and decoupled to provide operable vectorized inputs for subsequent causal computation; The structural causal model management module is responsible for storing a directed acyclic graph composed of nodes of the first type of operation type and nodes of the second type of semantic entity type. Define and manage directed edges between nodes, explicitly specify the causal dependency direction from operation nodes to semantic entity nodes, prohibit reverse causal edges, and provide topological constraints that conform to physical laws for counterfactual intervention; The counterfactual attention calculation module is used to integrate causal attribute encoding to construct a factual global context state; based on the constraints of the structural causal model, it applies do intervention to the target operation and propagates along the topology to execute the counterfactual intervention operation and generate a counterfactual global context state. By comparing the predictive matching degree of the two states for the target entity, the causal effect score between the first type of operational event and the second type of entity modification is calculated. The causal link generation and solidification module is used to filter out strong causal pairs that exceed the threshold based on the causal effect score; to construct a causal origin graph by tracing back upwards from the modified entity; to generate associated description data containing state information before and after modification and key operation sequences; and to perform SHA-256 hash operation on the above data and add a timestamp to complete the tamper-proof solidification. The collaborative session management module is responsible for creating, joining, leaving and destroying sessions; maintaining the global context state of the sessions; transmitting control signals between processing stages; triggering the start and termination of event acquisition and calculation processes; and coordinating the collaborative work of other modules. The persistent storage module is used to persistently save the original atomic event stream, intermediate representation state vectors, structural causal model diagram, calculated causal effect scores, the final generated causal link diagram, and solidified collaborative session version data for post-audit or model optimization.

Images