Medical image lesion analysis and clinical decision interpretation system based on concept activation vector

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a concept activation vector-based medical image lesion analysis and clinical decision interpretation system, which comprises a feature analysis module, a semantic mapping construction module, a concept activation extraction module, a boundary response fusion module and an interpretation path generation module. According to the method, the gradient direction and the gray abrupt change position of continuous elements in a focus area are detected, texture density and edge consistency are jointly judged to generate structured alignment mapping, microscopic image features are accurately captured, cross comparison is carried out on co-occurrence probability of lesion nouns and modifiers in phrase combinations and activation frequency of the same area, and therefore the accuracy of the lesion nouns and the modifiers in the phrase combinations is improved. Constructing a deep correlation map of pathological semantics and image features, screening an activation group according to continuous response of phrase pairs and image regions under case input change, generating a concept vector trigger track to dynamically track pathological feature evolution, and performing clinical consistency judgment; and an interpretable decision basis with strict logic support is output while subjective interference is eliminated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a medical image lesion analysis and clinical decision explanation system based on concept activation vectors. BACKGROUND

[0002] The field of artificial intelligence technology involves theoretical methods and application systems that simulate human intelligent behavior through computers. The core includes machine learning, computer vision, natural language processing, and expert systems. The field is committed to building intelligent agents that can perform perception, reasoning, learning, and decision-making tasks. Through pattern recognition and logical deduction of massive data, it realizes the automation of complex tasks and auxiliary decision support in medical health, industrial manufacturing, and autonomous driving scenarios. Among them, the traditional medical image lesion analysis and clinical decision explanation system refers to an auxiliary diagnosis platform that uses deep convolutional neural networks to extract features and classify CT or MRI image data. The class activation mapping (CAM) or gradient-weighted class activation mapping (GradCAM) algorithm is used to calculate the weight contribution of the convolution layer feature map to the predicted class, generate a heat map to highlight the pixel area that plays a key role in the classification result, and match the keywords in the electronic medical record. By manually comparing the overlap of the highlighted area of the heat map and the anatomical structure, the focus of the model is inferred, and the benign and malignant probability of the lesion is output according to statistical rules or pre-set medical guideline thresholds.

[0003] The existing technology relies too much on heat map display of pixel weight contribution, ignores the correlation between micro-texture density and edge gradient direction of the lesion area, resulting in lack of precision in image representation. Simply matching keywords with text is difficult to capture the specific semantic direction of disease nouns and modifiers, making it difficult to align the image activation area and pathological description. Manual comparison of the highlighted area and anatomical structure is easily affected by subjective experience, and relying solely on static statistical thresholds to determine benign and malignant cannot reflect the continuous response of features under varying input conditions, resulting in large semantic alignment deviation in complex lesion diagnosis, making it difficult to generate dynamic decision support with clinical consistency, and limiting the pathological reliability of auxiliary diagnosis results. SUMMARY

[0004] In order to solve the technical problems in the prior art that it is difficult to establish an effective connection between medical terms and model behaviors at the conceptual level, when facing the reasoning demand of high-level medical phrases in the clinical context, the response visualization of feature dimensions cannot support the discrimination and positioning of lesion concepts, and the continuity of model response fragments is poor when the input of cross-case data changes, which makes it difficult for doctors to grasp the continuity and stability of the model reasoning logic when evaluating the rationality of the diagnosis, affects the understanding depth and trust degree of the model judgment basis, and limits the practicability and generalizability in clinical auxiliary decision-making, the embodiments of the present application provide a medical image lesion analysis and clinical decision explanation system based on concept activation vector. The technical solution is as follows: In one aspect, a medical image lesion analysis and clinical decision explanation system based on concept activation vector is provided, which comprises: The feature analysis module obtains the two-dimensional element distribution in the medical sign, detects the gradient direction distribution and gray boundary mutation position of the continuous elements in the lesion candidate area, jointly judges the texture density sequence and edge direction consistency value in the same region, and generates a region feature structure alignment mapping result; The semantic mapping construction module calls the position index of the activated region in the region feature structure alignment mapping result, extracts the phrase combination in the standardized case record, cross-compare the co-occurrence probability of the lesion noun and the modifier in the phrase with the mapping region activation frequency, and generates a pathology semantic activation association graph; The concept activation extraction module detects the activation frequency distribution sequence of the phrase pair and the image region in multiple cases according to the pathology semantic activation association graph, screens the activated group with continuous response under the input variation of the case, and generates a concept vector activation trigger trajectory; The boundary response fusion module performs clinical consistency judgment on the fluctuation direction difference in the feature response interval of adjacent cases based on the activation frequency change sequence in the concept vector activation trigger trajectory, and generates an interpretable clinical decision basis.

[0005] As a further scheme of the present application, the region feature structure alignment mapping result includes texture density change, edge direction consistency value, and local gradient mutation marker, the pathology semantic activation association graph includes activation region position index, lesion phrase pair combination, and semantic structure corresponding label, the concept vector activation trigger trajectory includes continuous response fragment matching rate, activation frequency comparison trend value, and phrase image offset matching degree, and the interpretable clinical decision basis includes node frequency fluctuation record, direction consistency node combination, and feature response interval pairing relationship.

[0006] As a further scheme of the present application, the semantic mapping construction module comprises: The activation index extraction submodule calls the two-dimensional coordinate information corresponding to the structural unit in the structure feature alignment mapping result of the region, screens the coordinate points with an activation threshold value exceeding a positioning reference value in the gray level mutation position, regionally merges adjacent element points according to the activation frequency, and marks the position index number to obtain a mapping activated region position index group; The phrase co-occurrence calculation submodule obtains the phrase combination in the standardized case record by using the mapping activated region position index group, extracts the combination form of the lesion noun and the adjacent modifier in the phrase, and counts the joint frequency in the full text of the standardized case record, excludes the unstructured phrase combination, marks the position number in the record text for the legal combination, and generates a lesion phrase co-occurrence frequency vector group; The structural relationship determination submodule calls the lesion phrase co-occurrence frequency vector group, calculates the word activation position overlap probability of the phrase combination according to the corresponding distribution rule between the position number of the standardized case record and the image activation index, and performs structural matching marking on the phrase and the activated region with an overlap probability exceeding a comparison reference value to obtain a pathology semantic activation association graph.

[0007] As a further scheme of the application, the concept activation extraction module comprises: The activation frequency detection submodule calls the pathology semantic activation association graph, detects the activation frequency distribution of the combination of the phrase and the image region in the same structure group under multiple case sample inputs, sorts the case numbers in chronological order to obtain a multi-case activation frequency distribution sequence; The response continuity screening submodule extracts the phrase and image region combination with a non-zero activation frequency under the continuous case number based on the multi-case activation frequency distribution sequence, screens the activated paragraph that maintains continuous response when the case input changes, marks the paragraph according to the continuous length, and obtains a continuous response combination paragraph index set; The offset matching degree calculation submodule calls the continuous response combination paragraph index set, obtains the change direction sequence of each pair of phrase annotation values in the case sequence, calculates the amplitude change direction of the activation frequency of the corresponding image region, calculates the matching degree index according to the offset value of the change direction sequence of the phrase annotation value and the amplitude change direction of the activation frequency, and obtains a phrase region offset direction matching degree sequence; The trajectory triggering submodule extracts the phrase region combination with a matching degree index higher than an offset comparison reference value based on the phrase region offset direction matching degree sequence, respectively calls the activation frequency and the phrase appearance frequency in the corresponding continuous response combination paragraph index set, calculates the matching rate and the activation frequency trend value, integrates the trend value according to the combination order, and obtains a concept vector activation triggering trajectory.

[0008] As a further scheme of the present application, the calculation of the offset value is: the variation direction sequence of each pair of phrase label values in the case sequence is positionally matched with the amplitude change direction of the corresponding image region activation frequency, and the direction consistency at the position is calculated, and a matching value of "1" is given when the directions are consistent, a matching value of "-1" is given when the directions are opposite, and a matching value of "0" is given when the directions are not clear; The matching degree index is obtained by summing the matching values and normalizing them to the interval [-1, 1].

[0009] As a further scheme of the present application, the boundary response fusion module comprises: The concept projection extraction submodule calls the activation frequency change sequence in the concept vector activation trigger track, locates the concept vector identification number corresponding to the frequency fluctuation position, obtains the vector dimension of the corresponding node in the hidden space, and performs dimension projection on the activation response value under case input to obtain the hidden space node concept vector projection value sequence; The direction difference pairing judgment submodule extracts the activation response value change direction of the same node under adjacent case input based on the hidden space node concept vector projection value sequence, calculates the fluctuation direction difference value of the feature response interval, and then matches the difference value distribution after position numbering labeling with the corresponding positions of adjacent nodes to obtain a frequency fluctuation node direction pairing set; The segment coordination submodule calls the frequency fluctuation node direction pairing set, detects the synchronous fluctuation frequency of the node combination in the case sequence, calculates the direction consistency rate within the combination, and extracts the node combination interval whose synchronous rate and consistency rate both exceed the set standard value, and then performs continuous encoding processing on the combination nodes in the interval to obtain an interpretable clinical decision basis.

[0010] As a further scheme of the present application, the process of dimension projection comprises selecting a dimension subset whose response value fluctuation amplitude is greater than a set response change threshold value in the node dimension corresponding to each concept vector identification number; The response change threshold value is a percentile parameter calculated based on the standard deviation of the response value fluctuation of all node dimensions; When performing dimension projection, only the response values in the dimension subset are calculated, and the response value results are combined to form the hidden space node concept vector projection value sequence.

[0011] As a further scheme of the present application, the system further comprises an explanation path generation module: The explainable path generation module calls the explainable clinical decision basis, compares the element positions of the response ranking of nodes in consecutive training rounds, identifies the corresponding relationship between the nodes with response order mutation and the feature layer order offset nodes, constructs the order stability index of the response path in the node chain according to the corresponding relationship, and generates a structural explainability output path atlas; The structural explainability output path atlas includes node response order trajectory, order stability index and feature layer order offset chain.

[0012] As a further scheme of the application, the explainable path generation module comprises: The response ranking comparison submodule calls the node number in the explainable clinical decision basis, obtains the response value list of each node in consecutive training rounds, performs descending order arrangement in the round according to the response value and records the ranking serial number, calculates the ranking serial number difference value between adjacent rounds to obtain the node training round response ranking difference value sequence; The order offset identification submodule filters the node label whose ranking difference value exceeds the mutation reference value based on the node training round response ranking difference value sequence, extracts the response ranking structure of the feature nodes in the same layer in the corresponding training round, compares whether there is order mutation in the same position, matches the mutation point and the order offset point according to the node label corresponding relationship, and obtains the mapping relationship table of the response mutation node and the feature order offset node; The path atlas construction submodule calls the mapping relationship table of the response mutation node and the feature order offset node, constructs the response trajectory line segment according to the response value change curve of the node pair in consecutive training rounds, calculates the average difference value of the order fluctuation between the trajectory line segments as the order stability index, sorts the trajectory line segments according to the order stability index and connects the paths to obtain the structural explainability output path atlas.

[0013] The technical scheme provided by the embodiment of the application has at least the following beneficial effects: By detecting the gradient direction and the gray scale mutation position of the continuous elements in the lesion area, the texture density and the edge consistency are jointly judged to generate a structured alignment mapping, the microscopic image features are accurately captured, the cross comparison is performed on the co-occurrence probability of the lesion noun and the modifier in the phrase combination and the activation frequency of the same area, the deep association atlas of the pathological semantics and the image features is constructed, the activated group is screened according to the continuous response of the phrase pair and the image area under the case input change, the concept vector trigger trajectory is generated to dynamically track the pathological feature evolution, the clinical consistency is determined based on the fluctuation direction difference of the feature response interval, and the explainable decision basis with strict logical support is output while the subjective interference is excluded. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0015] Figure 1 is a flowchart provided by the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the present application will be described below with reference to the drawings.

[0017] In the embodiments of the present application, the words such as "example", "for example" are used to represent as an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0018] In the embodiments of the present application, "image" and "picture" can be used interchangeably at times. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent. "Of", "corresponding" and "corresponding" can be used interchangeably at times. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent.

[0019] In the embodiments of the present application, sometimes the subscript such as W1 can be written in the form of non-subscript such as W1. When the distinction is not emphasized, the meanings expressed are consistent.

[0020] In order to make the technical problems, technical solutions and advantages of the present application more clear, the following will be described in detail with reference to the drawings and specific embodiments.

[0021] The embodiments of the present application provide a medical image lesion analysis and clinical decision interpretation system based on concept activation vector, based on the flowchart as shown in Figure 1 The system comprises: The feature analysis module acquires the two-dimensional element distribution in the medical sign, detects the gradient direction distribution and gray boundary mutation position of the continuous elements in the lesion candidate region, jointly judges the texture density sequence and edge direction consistency value in the same region, and generates a region feature structure alignment mapping result; The semantic mapping construction module calls the position index of the activated region in the region feature structure alignment mapping result, extracts the phrase combination in the standardized case record, cross-compare the co-occurrence probability of the lesion noun and the modifier in the phrase with the position cross-comparison of the activated frequency of the mapping region, judge the structure corresponding relationship, and generate a pathological semantic activation correlation graph; The concept activation extraction module detects the activation frequency distribution sequence of the phrase pair and the image region in multiple cases according to the pathological semantic activation correlation graph, screens the activated group with continuous response under the input change of the case, compares each other through the matching degree of the trend of the phrase label change and the fluctuation range of the image activation, processes the matching rate of the phrase and the region combination in the continuous response segment and the trend value of the activation frequency comparison, and generates a concept vector activation trigger trajectory; The boundary response fusion module extracts the concept vector projection value of the node in the latent space based on the activation frequency change sequence in the concept vector activation trigger trajectory, judges the clinical consistency of the fluctuation direction difference in the feature response interval of adjacent cases, records the node combination with frequency change, judges the direction consistency between nodes, and generates an interpretable clinical decision basis; The explanation path generation module calls the interpretable clinical decision basis, compares the response ranking of the nodes in the continuous training round, identifies the corresponding relationship between the nodes with response order mutation and the feature layer sorting offset nodes, constructs the ordering stability index of the response path in the node chain according to the corresponding relationship, and generates a structure interpretable output path graph; The concept activation vector refers to distinguishing samples belonging to a certain semantic concept and samples not belonging to the concept in the hidden layer of the model (such as a certain layer of convolutional neural network), representing the direction of a specific concept (such as "lung nodule" or "hemorrhage" medical features); The latent space node concept vector refers to the activation direction of a certain concept of a node (such as a neuron or a channel) in the model (hidden layer); The region feature structure alignment mapping result includes texture density change, edge direction consistency value, and local gradient mutation marker. The pathological semantic activation correlation graph includes activated region position index, lesion phrase pair combination, and semantic structure corresponding label. The concept vector activation trigger trajectory includes continuous response segment matching rate, activation frequency comparison trend value, and phrase image offset matching degree. The interpretable clinical decision basis includes node frequency fluctuation record, direction consistency node combination, and feature response interval pairing relationship. The structure interpretable output path graph includes node response ordering trajectory, ordering stability index, and feature layer ordering offset chain.

[0022] Specifically, the feature analysis module includes: The gray matrix extraction submodule obtains the two-dimensional element distribution in the medical sign, extracts the gray value of the element unit to form a matrix structure, and performs row and column alignment and normalization processing on the gray value in the matrix to obtain a normalized two-dimensional gray matrix. According to the input medical image (such as a CT slice or an MRI image in DICOM format), the two-dimensional element array of the image is extracted using an image reading tool (such as the cv2.imread function in the OpenCV library of Python), and after being converted into a grayscale image, the gray value of each element point is extracted, which is an integer between 0 and 255, representing the brightness of the element point. Then the element gray value is organized in a matrix structure to generate an original two-dimensional gray matrix. The matrix is aligned in rows and columns, that is, the positions of the image elements in the horizontal and vertical directions are consistent with the corresponding positions of the actual image. The alignment method based on element index or coordinate remapping is used. In the case of rotation or misalignment in the image, geometric correction can be performed through an affine transformation matrix, and the physical positions of the element units in the matrix are adjusted by setting the reference point coordinates for coordinate mapping. The aligned matrix needs to be normalized, and the original gray value is normalized to the interval . The minimum-maximum normalization method is used, and the calculation formula is: ; wherein, represents the original gray value of the element at the th position, and are the minimum and maximum gray values of all elements in the image, respectively; After normalization, the gray value of each element point falls within the interval . If a lung CT image slice with a size of is selected, the minimum gray value of a certain region is 23, the maximum gray value is 189, and the original gray value of a certain element is , then the normalized gray value of the point is: ; The normalized elements collectively form a normalized two-dimensional gray matrix.

[0023] The gradient direction calculation submodule obtains the continuous element combination in the lesion candidate region of the normalized two-dimensional gray matrix, calculates the gray change rate value of each element unit along the horizontal and vertical axes, respectively, and extracts the corresponding gradient direction angle value in the neighborhood of the element point to obtain the element gradient direction angle sequence. Selecting each element point in the image as the center, the continuous element combination in the lesion candidate region (setting a window of ) is obtained, and the gray change rate values of the horizontal axis (x direction) and the vertical axis (y direction) are extracted, respectively. The Sobel operator can be used for approximate differential calculation. The convolution kernel for the x direction is: ; Perform convolution calculations between each element and its neighborhood to obtain the lateral gradient. With longitudinal gradient Then, the gradient direction angle value of that point is calculated. The calculation is performed on every element point in the entire image to obtain a complete element gradient direction angle sequence. The entire sequence is stored in matrix form to store the angle values ​​corresponding to the element points, thus obtaining the element gradient direction angle sequence.

[0024] The edge consistency determination submodule calls the element gradient direction angle sequence to obtain the position of abrupt gray-level change in the same region, calculates the direction difference between the texture density value and the gradient direction angle in the region, and judges the edge direction consistency in the sub-region according to the direction difference. The edge determination result is then processed by coordinate mapping to obtain the region feature structure alignment mapping result. Statistically determine the presence of abrupt changes in grayscale within the angles of each element's neighborhood, and set a threshold for the angle of such abrupt changes. When the difference between the direction angle of any element point in the neighborhood and the direction of the center element is... Exceed When a mutation point is reached, it is marked as a mutation point. Then, the texture density value is calculated for each sub-region, which can be calculated using the standard deviation of gray levels within the region. As a measure of texture density, it is set to Within the region, if the standard deviation of gray values ​​is If the density value is at a medium density level, the average directional difference is calculated based on the directional angle: ; in, This represents the number of element points within the sub-region. Set a reference value for the orientation of the central element. ,like If the direction is consistent, the region is determined to be a region with consistent direction and is set in a certain region. , ,but Therefore, since it is a consistent region, the determination result is backmapped to the original image coordinate system through affine transformation or image space coordinate mapping. The edge position is adjusted according to the image size ratio, and the region features are output. Structure alignment mapping results.

[0025] Specifically, the semantic mapping building module includes: The activation index extraction submodule calls the two-dimensional coordinate information corresponding to the structural units in the regional feature structure alignment mapping result, filters the coordinate points in the gray-scale abrupt change position where the activation threshold exceeds the positioning reference value, merges adjacent element points according to the activation frequency, and marks the position index number to obtain the mapped activation region position index group. The system evaluates each coordinate point individually, extracting those with significant abrupt changes in grayscale values. During this evaluation, the system iterates through the elements of the mapped image, calling the grayscale change trajectory of each element point during image changes, and determines whether the grayscale jump amplitude exceeds the activation threshold. If the grayscale abrupt change meets the set conditions, it is identified as a potential activation point. The activation points are then aggregated and filtered according to a preset positioning benchmark value, retaining only coordinate points with a change intensity higher than the benchmark. After filtering, the frequency of occurrence of the coordinate points is statistically analyzed. If certain points repeatedly exhibit abrupt changes in multiple detections or multiple frames of images, it indicates that the points have a high activation frequency. In this case, the system merges the high-frequency activation points with spatially adjacent activation points, determines whether they are located in the same region and form a relatively compact activation cluster, marks them as constituent units of the same region, and assigns them a unified number index, thus obtaining the mapped activation region location index group.

[0026] The phrase co-occurrence calculation submodule uses a mapping activation region location index group to obtain phrase combinations in standardized case records, extracts the combination forms of lesion nouns and adjacent modifiers in the phrases, and counts the joint frequency in the full text of the standardized case records. Unstructured phrase combinations are excluded, and the legal combinations are marked with position numbers in the record text to generate a lesion phrase co-occurrence frequency vector group. The text information in the case records is processed by calling standardized case documents and using natural language processing tools to divide the entire case content into several phrase fragments. The focus is on extracting nouns and modifiers with pathological features, identifying typical descriptive phrases such as "pulmonary nodules," "low-density areas in the brain," and "liver enhancement shadows." The position of each phrase is marked according to the original text structure, and the frequency of phrases appearing together in the whole text is counted to ensure that each group of lesion descriptions can be reflected in different paragraphs or sentences of the text. By judging the frequency of the combination, phrase combinations with incomplete semantic structure, illogical collocation, or loose structure are eliminated. For the remaining legitimate combinations, position indexes are marked in the original case text to ensure the precise location and retrieval of phrases. The legitimate phrase combinations and frequency data are summarized into a lesion phrase co-occurrence frequency vector group.

[0027] The structural relationship determination submodule calls the co-occurrence frequency vector group of pathological phrases, calculates the overlap probability of the word activation positions of phrase combinations according to the corresponding distribution pattern between the standardized case record location number and the image activation index, and performs structural matching and marking on phrases and activation regions whose overlap probability exceeds the comparison benchmark value to obtain the pathological semantic activation association map. The system compares the co-occurrence frequency vector of lesion phrases with the location index of the mapped activation region of the image item by item. The system determines whether a phrase can establish a stable correspondence with a certain image region by comparing the distance and positional relationship between the occurrence position of the phrase in the text and the spatial coordinate mapping of the activated region of the image. If the phrase describing "pulmonary nodule" in the text frequently appears in paragraphs mentioning the upper lobe of the right lung, and the high-frequency mutation points in the activated region of the image are also concentrated in the upper lobe of the right lung, then it can be determined that there is a correlation between the two and confirm whether they belong to the same lesion structure. The system evaluates the degree of overlap between space and semantics and determines whether it exceeds the preset comparison benchmark value. If it meets the benchmark, the phrase and the activated region are structurally labeled and paired, and the matching relationship and the resulting semantic structure label are recorded to obtain a pathological semantic activation association map.

[0028] Specifically, the concept activation extraction module includes: The activation frequency detection submodule calls the combination of phrases and image regions in the pathological semantic activation association map, and detects the activation frequency distribution of the corresponding image regions in the same structural group under multiple case sample inputs. The case numbers are sorted according to time order to obtain the activation frequency distribution sequence of multiple cases. With multiple different case samples input, the system sequentially calls the correspondence between phrases and image activation regions recorded in the structure group, and counts the frequency of image activation in different cases. In actual operation, the system assigns an independent number to each case, and during the loading of each case data, the image analysis module re-detects whether the activation position associated with the corresponding phrase in the image region exists, and counts whether the grayscale change reaches the previously set activation threshold standard. If the condition is met, it is recorded as a valid activation. The statistical operation is repeatedly performed on different case data to build a cross-case activation frequency record list for each phrase and image region combination, and arranges them according to the order of case numbers, thus forming an activation frequency distribution sequence under multiple cases. When processing 10 lung CT case samples in succession, if "lung nodules" and a certain activation region are frequently activated in cases 1, 2, 4, and 5, but there is no obvious response in the cases, the frequency sequence of the combination is non-uniformly distributed. The system organizes such sequences according to case numbers to obtain a multi-case activation frequency distribution sequence.

[0029] The response continuity filtering submodule extracts phrases and image regions with non-zero activation frequencies under consecutive case numbers based on the multi-case activation frequency distribution sequence, filters activation segments that maintain continuous response when case input changes, and numbers and marks the segments according to their continuous length to obtain a continuous response combination segment index set. The main objective is to determine whether a combination of a phrase and an image region maintains a stable response under continuously changing case numbers. The system searches for each combination, extracts sequence segments with non-zero activation frequencies and consecutive case numbers, filters these segments, and records the length of the continuous response. Only combination segments that do not interrupt the response during case changes are retained. For example, if a combination has response frequencies in case numbers 1 to 3 and 6 to 8, the system will identify two independent continuous response segments with corresponding lengths of 3 and 3, respectively. The system will assign independent index numbers to these continuous segments, setting them as Segment 1 and Segment 2, and establish a continuous response combination segment index set in the format of "combination number + segment number". In actual data processing, such as processing 20 case records, the system will analyze and mark the continuous activation performance of each phrase-image combination, using the continuity of activation as a filtering criterion to obtain the continuous response combination segment index set.

[0030] The offset matching degree calculation submodule calls the continuous response combined paragraph index set to obtain the change direction sequence of each pair of phrase annotation values ​​in the case sequence, and calculates the amplitude change direction of the activation frequency of the corresponding image region. Based on the offset value of the change direction sequence of phrase annotation values ​​and the amplitude change direction of activation frequency, the matching degree index is calculated to obtain the phrase region offset direction matching degree sequence. The system compares the response change trends of each phrase-image region combination in the case sample sequence. It extracts the direction of change in the phrase's position in the text; for example, if a phrase appears at the beginning of a paragraph in case 1 and in the middle of a paragraph in case 2, the direction of the annotation shifts backward. Simultaneously, the activation frequency records of the image region are processed in the same way, calculating the direction of frequency increase or decrease when the case number changes. For example, if the frequency increases from 3 times in case 1 to 5 times in case 2, the activation direction is considered to be an enhancement shift. The system compares the sequence of changes in the phrase annotation value with the direction of change in the magnitude of the activation frequency item by item to determine whether the change trend of each phrase-image combination at the corresponding position of each case number is consistent. It also calculates the percentage of items with consistent shifts and calculates the matching degree index between the two. For example, if a combination appears in 7 out of 10 cases with the phrase position change direction consistent with the image activation direction, the matching degree index is 0.7. This is then compared to the phrase region shift direction matching degree sequence.

[0031] The trajectory triggering submodule extracts phrase region combinations with matching degree indices higher than the offset comparison benchmark value based on the phrase region offset direction matching degree sequence. It then calls the activation frequency and phrase occurrence frequency in the corresponding continuous response combination paragraph index set, calculates the matching rate and activation frequency comparison trend value, and integrates the trend value according to the combination order to obtain the concept vector activation trigger trajectory. Combinations exhibiting outstanding consistency in offset direction are selected to establish highly reliable trend trajectories. The system traverses the phrase region offset direction matching degree sequence, extracts combination entries with matching degrees higher than the preset offset comparison benchmark value, and calls activation frequency data from the continuous response combination paragraph index set and phrase occurrence frequency data in the text for each combination. The two are then compared for trends. For example, in 5 consecutive cases, the occurrence frequency of a certain phrase increases from 1 to 5 times, and the corresponding image activation frequency also continues to increase. At this time, the system considers the combination to show a positive correlation trend. By constructing a combination sequence diagram, the corresponding matching rate and activation frequency trend change value are calculated for each set of data. The trend values ​​are then integrated according to the combination order to obtain the concept vector activation trigger trajectory.

[0032] Specifically, the boundary response fusion module includes: The concept projection extraction submodule calls the activation frequency change sequence in the concept vector activation trigger trajectory, locates the concept vector identifier corresponding to the frequency fluctuation position, obtains the vector dimension of the corresponding node in the latent space, and projects the dimension of the activation response value under the case input to obtain the concept vector projection value sequence of the latent space node. The system identifies and locates frequency fluctuations, scanning the entire activation frequency curve sequence to find points of significant change, including key nodes where the frequency suddenly increases or decreases. After locating the fluctuation point, the system extracts the corresponding concept vector identifier, determining which concept dimension caused the frequency fluctuation. Based on the concept vector identifier, the system enters the preset latent space vector structure model, calling the latent space node information corresponding to each identifier to obtain the original vector dimension of the node in the high-dimensional semantic space, such as 512, 1024, or higher dimensions. After obtaining the dimension information, the system collects the response values ​​of the nodes in the active state for each case input and maps the original high-dimensional vector to a latent space projection of a specified dimension through dimensionality reduction or linear projection. The mapping operation can use predefined matrix transformation rules or specific transformation benchmarks. In practice, for example, principal component analysis (PCA) can be used to project in a pre-selected direction to ensure that the changes in vector values ​​can truly reflect the change trajectory between concepts. The system will uniformly record the projection results of the nodes under the case, obtaining the latent space node concept vector projection value sequence.

[0033] The direction difference pairing judgment submodule extracts the direction of change of activation response value of the same node under adjacent case input based on the latent space node concept vector projection value sequence, calculates the fluctuation direction difference of the feature response interval, and then matches the difference distribution with the corresponding position of the adjacent node to obtain the frequency fluctuation node direction pairing set. The system performs differential analysis on the projected response values ​​of each node under different case inputs, extracts the direction of change of the activation response value of the node in adjacent case samples, that is, determines whether the projected value of the node is rising, falling or remaining unchanged when the case number changes. The system performs directional analysis on each node in turn, generating a fluctuation direction sequence of each node in the complete case sequence. Based on this, the system calculates the directional difference within each response interval, that is, calculates whether the node direction changes between two adjacent case samples. If the directions are consistent, the directional difference is zero; if there is a positive or negative change, the directional difference is one. The difference sequence is labeled with the corresponding case position number. The system compares the directional difference sequence of each node with adjacent nodes in the spatial location to analyze whether there is a pairing relationship of fluctuation direction, that is, whether they show consistent or similar directional change trends at the same position. If node A and node B both show a continuous upward trend between cases 4 and 7, they are considered to have directional consistency within the interval. The system combines nodes with synchronous change trends into pairing results, obtaining a frequency fluctuation node direction pairing set.

[0034] The segment collaboration submodule calls the frequency fluctuation node direction pairing set, detects the synchronization fluctuation frequency of node combinations in the case sequence, calculates the direction consistency rate within the combination, and extracts the node combination interval where the synchronization rate and consistency rate both exceed the set standard value. The combination nodes within the interval are continuously encoded to obtain interpretable clinical decision-making basis. The system detects the synchronicity of fluctuations in the case sequence of node combinations. Based on the combination structure of the node pairs, it checks whether the response fluctuations of the nodes at their respective positions occur simultaneously in the case number sequence. By comparing the directional change sequences of two nodes, if both exhibit fluctuations in a certain case segment, whether upward or downward, as long as there is a directional change within the same case pair, it is considered a synchronous fluctuation. After completing the detection of the entire case sequence, the system statistically analyzes the frequency of synchronous fluctuations between the combined nodes and calculates the proportion of combined nodes with consistent directions in all fluctuation events, i.e., the direction consistency rate. If a node combination not only frequently exhibits synchronous fluctuations in multiple consecutive cases but also shows a high degree of consistency in the direction of fluctuation, the system will identify it as a structural combination with a strong cooperative relationship. Based on this, the system sets a joint judgment standard value for the synchronization rate and consistency rate. If both are greater than 0.6 or 0.7, node combinations that meet the standard will be extracted. The system performs continuous encoding processing on the case sequence segment where the combined nodes are located, i.e., assigning a unique identifier to the combined nodes within the segment, thus obtaining interpretable clinical decision-making basis.

[0035] Specifically, the path generation module includes: The response ranking comparison submodule calls the node number in the interpretable clinical decision basis, obtains the response value list of each node in consecutive training rounds, sorts the response values ​​in descending order within the round and records the ranking number, calculates the difference between the ranking numbers of the nodes in adjacent rounds, and obtains the node training round response ranking difference sequence. The system retrieves the activation response values ​​of each node in consecutive training rounds, constructing a list of response values ​​for each node across different rounds. Each response value represents the activity level or activation intensity of the node within the model in each training round. For example, in five consecutive rounds, the response values ​​of node A are 0.85, 0.79, 0.93, 0.88, and 0.81, respectively. After collecting the response value sequence, the system sorts the node's response values ​​from highest to lowest within each training round, generating a descending list for each round. The system also records the specific ranking number of each node within the round. Furthermore, the system performs a difference calculation on the ranking numbers of nodes in adjacent training rounds, comparing the change in ranking of the same node between the current and previous rounds. The larger the difference, the more drastic the fluctuation in the node's response intensity during model training. The system records the node's ranking difference as a complete sequence, forming the node training round response ranking difference sequence.

[0036] The sorting offset recognition submodule is based on the sequence of ranking difference of the response of the node training rounds. It filters the node labels whose ranking difference exceeds the mutation benchmark value, extracts the response ranking structure of the feature nodes in the same layer under the corresponding training round, compares whether there is a sorting mutation at the same position, and matches the mutation point and sorting offset point according to the node label correspondence to obtain the mapping relationship table between response mutation nodes and feature sorting offset nodes. The system identifies and processes nodes whose response ranking changes abruptly during model training. A threshold value is set as the identification condition; a node whose ranking changes by more than 10 places in adjacent training rounds is considered a ranking mutation. The system sequentially filters the ranking differences of nodes in the training sequence, marking the node numbers whose ranking differences exceed the threshold value in any round. After identifying the ranking mutation node, the system extracts the complete response ranking structure of the remaining feature nodes at the same level within the corresponding round. This involves reloading the response rankings of all nodes in the round and sequentially determining whether any remaining nodes also show significant shifts or changes in their ranking positions. If multiple ranking mutations occur in the same or adjacent positions, the system establishes a correspondence between the node ranking mutation behaviors and performs clinical consistency based on the node labels. This identifies which ranking mutations and shifts occur in the same region or the same logical group, resulting in a mapping table between response mutation nodes and feature ranking shift nodes.

[0037] The path graph construction submodule calls the mapping table between response mutation nodes and feature sorting offset nodes. Based on the response value change curve of the node pair in consecutive training rounds, it constructs response trajectory segments. It calculates the average difference of sorting fluctuations between trajectory segments as the sorting stability index. The trajectory segments are sorted according to the sorting stability index and connected to the path to obtain the structural interpretability output path graph. The system begins constructing the graph structure. It curves the response value trends between each pair of nodes, plotting a curve representing the response value of the same node across consecutive training rounds. Each curve is then divided into multiple trajectory segments to refine the node's behavior within a specific round interval. The system calculates the average difference in ranking fluctuations between any two adjacent trajectory segments, serving as a measure of the node's stability in responding to training data. A smaller average difference indicates more consistent response behavior during training, while large fluctuations suggest frequent adjustments to the node utilization strategy during model learning. The system ranks the trajectory segments according to the ranking stability index, from most stable to least stable, and connects segments with similar rankings or stable trends to form a path graph structure containing multiple nodes and behavioral patterns across multiple training rounds. The system outputs the structural path graph and visualizes the node connections and response trends, providing a structurally interpretable output path graph to support the interpretability requirements of the model's internal mechanisms.

[0038] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A medical image lesion analysis and clinical decision explanation system based on concept activation vectors, characterized by, The system comprises: The feature analysis module obtains the two-dimensional element distribution in the medical sign, detects the gradient direction distribution and the gray boundary mutation position of the continuous elements in the lesion candidate region, jointly judges the texture density sequence and the edge direction consistency value in the same region, and generates the region feature structure alignment mapping result; The semantic mapping construction module calls the position index of the activated region in the region feature structure alignment mapping result, extracts the phrase combination in the standardized case record, cross-compare the co-occurrence probability of the lesion noun and the modifier in the phrase with the mapping region activation frequency, and generates the pathology semantic activation correlation graph; The concept activation extraction module detects the activation frequency distribution sequence of the phrase pair and the image region in multiple cases according to the pathology semantic activation correlation graph, screens the activated groups with continuous response under the change of case input, and generates the concept vector activation trigger trajectory; The boundary response fusion module judges the clinical consistency of the fluctuation direction difference in the feature response interval of adjacent cases based on the activation frequency change sequence in the concept vector activation trigger trajectory, and generates an interpretable clinical decision basis.

2. The concept activation vector based medical image lesion analysis and clinical decision explanation system according to claim 1, characterized in that: The region feature structure alignment mapping result includes texture density change, edge direction consistency value, and local gradient mutation marker, the pathology semantic activation correlation graph includes activated region position index, lesion phrase pair combination, and semantic structure corresponding label, the concept vector activation trigger trajectory includes continuous response segment matching rate, activation frequency comparison trend value, and phrase image offset matching degree, and the interpretable clinical decision basis includes node frequency fluctuation record, direction consistency node combination, and feature response interval pairing relationship. 3.The concept activation vector based medical image lesion analysis and clinical decision explanation system according to claim 1, wherein: The feature analysis module comprises: The gray matrix extraction submodule obtains the two-dimensional element distribution in the medical sign, extracts the gray value of the element unit to form a matrix structure, and performs row and column alignment and normalization processing on the gray value in the matrix to obtain a normalized two-dimensional gray matrix; The gradient direction calculation submodule obtains the continuous element combination of the lesion candidate region in the matrix based on the normalized two-dimensional gray matrix, calculates the gray change rate value along the horizontal and vertical axis directions for each group of element units, and extracts the corresponding gradient direction angle value in the element point neighborhood to obtain an element gradient direction angle sequence; The edge consistency determination submodule calls the element gradient direction angle sequence, obtains the gray change mutation position in the same region, calculates the direction difference value of the texture density value and the gradient direction angle in the region, judges the edge direction consistency in the sub-region according to the direction difference value, performs coordinate mapping processing on the edge determination result, and obtains the region feature structure alignment mapping result.

4. The concept activation vector based medical image lesion analysis and clinical decision explanation system of claim 3, wherein: The semantic mapping construction module comprises: The activation index extraction submodule calls the two-dimensional coordinate information of the structure unit corresponding to the region feature structure alignment mapping result, screens the coordinate points with an activation threshold value exceeding a positioning reference value in the gray mutation position, merges the adjacent element points according to the activation frequency, and marks the position index number to obtain a mapping activated region position index group; The phrase co-occurrence calculation submodule adopts the mapping activation region position index set to obtain phrase combinations in the standardized case record, extracts lesion noun and adjacent modifier combination forms in the phrases, and counts the joint frequency in the full text of the standardized case record, excludes non-structured phrase combinations, marks the position number of the legal combination in the record text, and generates a lesion phrase co-occurrence frequency vector set; The structural relationship determination submodule calls the lesion phrase co-occurrence frequency vector set, calculates the lexical activation position overlap probability of the phrase combination according to the corresponding distribution rule between the standardized case record position number and the image activation index, and performs structural matching marking on the phrases and the activation region whose overlap probability exceeds the comparison benchmark value, to obtain a pathology semantic activation correlation graph.

5. The concept activation vector based medical image lesion analysis and clinical decision explanation system of claim 4, wherein: The concept activation extraction module comprises: The activation frequency detection submodule calls the pathology semantic activation correlation graph, detects the activation frequency distribution of the combined image region in the same structure group under multiple case sample inputs, sorts the case numbers in chronological order, and obtains a multi-case activation frequency distribution sequence; The response continuity screening submodule extracts the phrase and image region combination with non-zero activation frequency under the continuous case number based on the multi-case activation frequency distribution sequence, screens the activation paragraph that maintains continuous response when the case input changes, marks the paragraph according to the continuous length, and obtains a continuous response combination paragraph index set; The offset matching degree calculation submodule calls the continuous response combination paragraph index set, obtains the change direction sequence of each pair of phrase annotation values in the case sequence, calculates the amplitude change direction of the corresponding image region activation frequency, calculates the matching degree index according to the offset value of the change direction sequence of the phrase annotation value and the amplitude change direction of the activation frequency, and obtains a phrase region offset direction matching degree sequence; The trajectory triggering submodule extracts the phrase region combination with a matching degree index higher than the offset comparison benchmark value based on the phrase region offset direction matching degree sequence, respectively calls the activation frequency and phrase appearance frequency in the corresponding continuous response combination paragraph index set, calculates the matching rate and activation frequency trend value, integrates the trend value according to the combination order, and obtains a concept vector activation trigger trajectory.

6. The concept activation vector based medical image lesion analysis and clinical decision explanation system of claim 5, wherein: The calculation of the offset value is: the change direction sequence of each pair of phrase annotation values in the case sequence and the amplitude change direction of the corresponding image region activation frequency are respectively position-corresponded, the direction consistency at the position is calculated, and when the directions are consistent, a matching value "1" is assigned, when the directions are opposite, a matching value "-1" is assigned, and when the directions are not clear, a matching value "0" is assigned; The matching degree index is obtained by summing the matching values and normalizing them to the [-1, 1] interval.

7. The concept activation vector based medical image lesion analysis and clinical decision explanation system of claim 5, wherein: The boundary response fusion module comprises: The concept projection extraction submodule calls the activation frequency change sequence in the concept vector activation trigger trajectory, locates the concept vector identification number corresponding to the frequency fluctuation position, obtains the vector dimension of the corresponding node in the latent space, and performs dimension projection on the activation response value under the case input, to obtain a latent space node concept vector projection value sequence; The direction difference pairing judgment submodule extracts the activation response value change direction of the same node under adjacent case inputs based on the hidden space node concept vector projection value sequence, calculates the fluctuation direction difference value of the feature response interval, and then matches the difference value distribution with the corresponding positions of adjacent nodes to obtain a frequency fluctuation node direction pairing set; The segment coordination submodule calls the frequency fluctuation node direction pairing set, detects the synchronous fluctuation frequency of the node combination in the case sequence, calculates the consistent rate of the direction within the combination, and extracts the node combination interval whose synchronous rate and consistent rate both exceed the set standard value, performs continuous coding processing on the combination nodes in the interval to obtain an interpretable clinical decision basis.

8. The concept activation vector based medical image lesion analysis and clinical decision explanation system of claim 7, wherein: The dimension projection process includes selecting a dimension subset with a response value variation amplitude greater than a set response change threshold in each node dimension corresponding to the concept vector identification number; The response change threshold is a percentile parameter calculated based on the standard deviation of the response value fluctuation of all node dimensions; When performing dimension projection, only the response values in the dimension subset are calculated, and the response value results are combined to form the hidden space node concept vector projection value sequence.

9. The concept activation vector based medical image lesion analysis and clinical decision explanation system according to claim 1, wherein: The system further includes an explanation path generation module: The explanation path generation module calls the interpretable clinical decision basis, compares the element positions of the response rankings of the nodes in the continuous training rounds, identifies the corresponding relationship between the nodes with response order mutations and the feature layer ordering offset nodes, constructs the ordering stability index of the response path within the node chain according to the corresponding relationship, and generates a structure explainability output path atlas; The structure explainability output path atlas includes node response ordering trajectories, ordering stability indexes, and feature layer ordering offset chains.

10. The concept activation vector based medical image lesion analysis and clinical decision explanation system of claim 9, wherein: The explanation path generation module includes: The response ranking comparison submodule calls the node numbers in the interpretable clinical decision basis, obtains the response value list of each node in the continuous training rounds, performs descending order arrangement within the rounds according to the response values and records the ranking serial numbers, calculates the ranking serial number difference value between adjacent rounds, and obtains the node training round response ranking difference value sequence; The ordering offset identification submodule filters the node numbers with ranking difference values exceeding the mutation benchmark value based on the node training round response ranking difference value sequence, extracts the response ranking structure of the feature nodes in the same layer in the corresponding training rounds, compares whether there is an ordering mutation in the same position, matches the mutation points and ordering offset points according to the node number corresponding relationship, and obtains a response mutation node and feature ordering offset node mapping relationship table; The path atlas construction submodule calls the response mutation node and feature ordering offset node mapping relationship table, constructs response trajectory line segments according to the response value change curves of the node pairs in the continuous training rounds, calculates the average difference value of the ordering fluctuations between the trajectory line segments as the ordering stability index, sorts the trajectory line segments according to the ordering stability index and connects the paths to obtain a structure explainability output path atlas.

Images