Medical image report generation informatization planning system

Abstract

The invention relates to the technical field of medical image processing and artificial intelligence auxiliary diagnosis, in particular to a medical image report generation informatization planning system, which comprises a multi-modal time sequence difference feature analysis module used for acquiring current medical image data and historical image data of a patient and generating a time sequence difference residual vector; the semantic entropy flow conservation and negentropy injection control module is used for calculating a system net entropy target; the dynamic planning generation module based on residual driving is used for generating a diagnosis report text sequence; the diagnostic specificity and normativity collaborative optimization module is used for constructing a composite loss function containing a cross entropy loss term and a specificity penalty term and carrying out iterative updating on system parameters based on the composite loss function, and the specificity penalty term is used for constraining the description accuracy of the generated text on the dynamic change of the image; the problem that the dynamic change of the disease course cannot be captured only by relying on static image analysis in the prior art is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and artificial intelligence assisted diagnosis, in particular to a medical image report generation information planning system. BACKGROUND

[0002] With the continuous improvement of medical image informatization level, the accumulation of medical data has increased significantly; this trend makes it inevitable to use artificial intelligence to assist in generating diagnosis reports, especially in terms of reducing the burden of doctors and improving efficiency; At present, the generation of medical image reports generally adopts an end-to-end image description model, or is manually written by doctors referring to static images of a single examination; the existing automation technology usually only focuses on the image features at the current time, lacking effective utilization and comparative analysis of the patient's historical image data; however, the traditional static analysis method cannot capture the dynamic evolution information of the lesion over time, and it is difficult to identify complex situations such as the same disease with different images; in addition, the existing model often faces the problem of modal collapse when generating text, and tends to output general and vague descriptions, resulting in generated reports that are grammatically correct but lack specific descriptions of key dynamic changes such as lesion enlargement, reduction or new onset, and cannot meet the accuracy requirements of clinical diagnosis; therefore, how to effectively utilize time-series image data to capture dynamic changes of lesions and generate high-quality diagnosis reports with both standardization and specificity has become a problem that needs to be solved in the field. SUMMARY

[0003] To solve the above technical problems, the present application provides a medical image report generation information planning system, specifically, the technical scheme of the present application comprises: A multi-modal time-series differential feature analysis module is used to obtain the current medical image data and historical image data of a patient, unify them to the same spatial coordinate system using a rigid registration algorithm, and call a convolutional neural network to extract features to generate a time-series differential residual vector; A semantic entropy flow conservation and negative entropy injection control module is used to calculate the mutual information between the time-series differential residual vector and the text sequence to be generated, and combine a pre-set negative entropy injection factor to calculate the system net entropy target; A dynamic programming generation module based on residual driving is used to project the time-series differential residual vector to the vocabulary space, and combine the hidden state of the decoder at the current time and the residual gate weight to generate a diagnosis report text sequence; A diagnosis specificity and standardization collaborative optimization module is used to construct a composite loss function comprising a cross-entropy loss term and a specificity penalty term, and iteratively update the system parameters based on the composite loss function, wherein the specificity penalty term is used to constrain the accuracy of the generated text in describing the dynamic changes of the image.

[0004] Preferably, the multi-modal time-difference feature analysis module generates a time-difference residual vector, specifically including the following steps: A preset convolutional neural network feature extraction function is called to process the current medical image data and the historical image data respectively, and output a current feature tensor and a historical feature tensor; A preset baseline decay coefficient is obtained, which is a scalar value used to adjust the weight of the historical data; The product of the historical feature tensor and the baseline decay coefficient is calculated to obtain a weighted historical feature tensor; An element-wise subtraction operation is performed between the current feature tensor and the weighted historical feature tensor, and the difference value obtained by the operation is defined as a time-difference residual vector.

[0005] Preferably, the semantic entropy flow conservation and negative entropy injection control module determines a system net entropy target, specifically including the following steps: The conditional entropy value of generating a diagnosis report text sequence under the condition of the given current medical image data is calculated; The mutual information value between the diagnosis report text sequence and the time-difference residual vector is calculated; A negative entropy injection factor is obtained, which is a preset scalar used to control the dynamic information focus; The product of the mutual information value and the negative entropy injection factor is calculated; The difference between the conditional entropy value and the product is calculated, and the difference is determined as the system net entropy target.

[0006] Preferably, the residual-driven dynamic programming generation module generates a diagnosis report text sequence, specifically including the following steps: The hidden state vector of the decoder at the current time is obtained and transformed using an output layer linear transformation matrix to obtain a first logical vector; A learnable residual mapping matrix is called to project the time-difference residual vector to the vocabulary space to obtain a second logical vector; A residual gating weight is obtained, which is a scalar with a value range of 0 to 1, and the second logical vector is weighted using the weight; The first logical vector and the weighted second logical vector are added and summed, and the sum result is applied to the Softmax activation function to obtain the word probability distribution and generate the word accordingly.

[0007] Preferably, the diagnosis specificity and normativity collaborative optimization module constructs a compound loss function, specifically including the following steps: The probability distribution difference between the diagnosis report text sequence and the true report is calculated to generate a standard cross-entropy loss value; Based on the hinge loss principle, a specificity penalty term for dynamic change feature description is calculated; obtaining a specificity penalty coefficient corresponding to the current training round, the coefficient being a scalar dynamically changing with the training process; calculating the product of the specificity penalty term and the specificity penalty coefficient, and adding the product to the standard cross-entropy loss value to obtain a composite loss function.

[0008] Preferably, the specificity penalty term is calculated, specifically including the following steps: searching in the diagnostic report text sequence to identify a key phrase matching a preset keyword set, the preset keyword set containing pre-stored words for describing the dynamic changes of the lesion morphology; obtaining a hidden layer state vector corresponding to the key phrase in the decoder as a text semantic feature representation; calling a cross-modal alignment projection matrix to linearly map the time series difference residual vector to a text semantic space to obtain a mapped residual vector; calculating a cosine similarity value between the text semantic feature representation and the mapped residual vector.

[0009] Preferably, the calculation of the specificity penalty term further includes the following logical judgment steps: obtaining a preset semantic matching threshold, the threshold being a preset scalar value; judging whether the cosine similarity value is lower than the semantic matching threshold; If yes, calculating the difference between the semantic matching threshold and the cosine similarity value, and adding the difference to the specificity penalty term; If no, setting the penalty value for the key phrase to zero.

[0010] Preferably, the diagnostic specificity and normative collaborative optimization module is further configured to perform the following parameter adjustment strategy: in the initial stage of system training, the specificity penalty coefficient is set to a first preset value; with the increase of the training round, the specificity penalty coefficient is gradually increased according to a preset annealing strategy until it reaches a second preset value greater than the first preset value, so as to strengthen the description accuracy of the model to the time series difference residual vector in the later training stage.

[0011] Compared with the prior art, the present application has the following beneficial effects: 1. The application effectively solves the problem that the prior art cannot capture the dynamic changes of the disease course by relying only on static image analysis through the multi-modal time sequence difference feature analysis module; the system uses a rigid registration algorithm to eliminate body position differences, and introduces a baseline decay coefficient to perform weighted difference operation on historical and current image features; this processing method can effectively filter out non-pathological differences caused by device noise or non-rigid deformation, accurately quantify the biological evolution characteristics of the lesion over time, and provide a solid physical basis for subsequent generation of diagnostic reports with longitudinal contrast value; 2. The application overcomes the modal collapse problem in the process of mapping visual information to natural language by using the semantic entropy flow conservation and negative entropy injection control module; the mutual information between the time sequence difference residual vector and the text sequence to be generated is calculated, and the system net entropy target is established combined with the negative entropy injection factor, which forces the model to pay attention to the high-frequency detail changes in the image at the information theory level; this effectively prevents the model from generating hollow and universal descriptions, ensuring that the generated report content is substantial and has substantive diagnostic information; 3. The application adopts a dynamic programming generation module based on residual driving, which realizes explicit causal driving from image change characteristics to text vocabulary selection; the system projects the time sequence difference residual vector into the vocabulary space, and uses adaptive residual gating weights to adjust the intervention intensity of image features on vocabulary generation; this dual-drive mechanism ensures that when the image changes significantly, the generation probability of related descriptive words is significantly improved, ensuring that each diagnostic keyword generated is directly controlled by the dynamic changes of the image, accurately reflecting the enlargement, reduction or new state of the lesion; 4. The application solves the contradiction between report syntax smoothness and diagnostic accuracy through the diagnostic specificity and normative collaborative optimization module; the system constructs a composite loss function including cross-entropy loss and specificity penalty term, aligns the text semantics and image features using a cross-modal projection matrix; combined with a dynamic annealing strategy to adjust the penalty coefficient, simulates the gradual learning process from learning to speak to learning to diagnose, ensuring that the generated report meets the medical grammar specifications while accurately describing the dynamic evolution of key lesions. BRIEF DESCRIPTION OF DRAWINGS

[0012] The application will be further explained in conjunction with the drawings and examples: Figure 1 is a structural diagram of the system of the application. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in conjunction with specific examples.

[0014] Example 1 Please refer to Figure 1A medical image report generation information planning system comprises: A multi-modal time series difference feature analysis module is configured to obtain current medical image data and historical image data of a patient, unify the two to the same spatial coordinate system by using a rigid registration algorithm, and call a convolutional neural network to extract features respectively to generate a time series difference residual error vector; A semantic entropy flow conservation and negative entropy injection control module is configured to calculate mutual information between the time series difference residual error vector and a text sequence to be generated, and calculate a system net entropy target in combination with a preset negative entropy injection factor, thereby forcing the model to focus on high-frequency detail changes in the image at the information theory level; A residual error driven dynamic planning generation module is configured to project the time series difference residual error vector to a vocabulary space, and generate a diagnostic report text sequence in combination with a hidden state of the decoder at the current time and a residual error gating weight; A diagnosis specificity and normativity collaborative optimization module is configured to construct a composite loss function comprising a cross-entropy loss term and a specificity penalty term, and iteratively update system parameters based on the composite loss function, wherein the specificity penalty term is used to constrain the accuracy of the generated text in describing the dynamic changes of the image.

[0015] The embodiment provides a medical image report generation information planning system; in view of the problems of difficult identification of the same disease and different images and lack of description of dynamic evolution of a lesion in a generated report in the prior art, the system realizes full-process control from image difference perception to high-precision text generation through collaborative work of four core modules; The multi-modal time series difference feature analysis module aims to solve the problem that static image analysis cannot capture dynamic changes of a disease course; in the embodiment, the module is configured to obtain current medical image data of a patient and historical image data , unify the two to the same spatial coordinate system by using a rigid registration algorithm or a non-rigid registration algorithm such as a B-spline free-form deformation algorithm to eliminate body position difference and organ soft tissue deformation, call a convolutional neural network to extract features respectively, and finally generate a time series difference residual error vector capable of representing dynamic evolution information of a lesion; The semantic entropy flow conservation and negative entropy injection control module aims to solve the modal collapse problem in the process of mapping high-dimensional visual information to low-dimensional natural language; in the embodiment, the module is configured to calculate mutual information between the time series difference residual error vector and a text sequence to be generated, and calculate a system net entropy target in combination with a preset negative entropy injection factor, thereby forcing the model to focus on high-frequency detail changes in the image at the information theory level; a dynamic planning generation module based on residual driving, which aims to endow the generation model with causal reasoning ability; in this embodiment, the module is used to project the time-series difference residual vector into the vocabulary space, and generate a diagnostic report text sequence in combination with the hidden state of the decoder at the current time and the residual gate weight, ensuring that each word generated is explicitly driven by the dynamic changes in the image; a diagnostic specificity and normative collaborative optimization module, which aims to solve the contradiction between report fluency and accuracy; in this embodiment, the module is used to construct a composite loss function containing a cross-entropy loss term and a specificity penalty term, and iteratively update the system parameters based on the function; wherein the specificity penalty term is specifically used to constrain the accuracy of the generated text in describing the dynamic changes of the image, realizing the gradual optimization from learning to speak to learning to diagnose; Through the synergistic effect of the above four modules, the system breaks through the limitation of traditional end-to-end models that only focus on static features; the differential features extracted by module one provide the core physical basis for subsequent generation; modules two and three ensure that the differential features are effectively encoded into the text report from the dimensions of information flow control and generation path planning; module four ensures that the generated report can accurately reflect the dynamic characteristics such as the increase, decrease or new occurrence of the lesion while being grammatically fluent, significantly improving the clinical value of auxiliary diagnosis.

[0016] Embodiment 2 The multi-modal time-series difference feature analysis module generates a time-series difference residual vector, specifically including the following steps: A preset convolutional neural network feature extraction function is called to process the current medical image data and the historical image data respectively, and output the current feature tensor and the historical feature tensor; A preset baseline decay coefficient is obtained, which is a scalar value used to adjust the weight of the historical data; The product of the historical feature tensor and the baseline decay coefficient is calculated to obtain a weighted historical feature tensor; An element-wise subtraction operation is performed between the current feature tensor and the weighted historical feature tensor, and the difference value obtained by the operation is defined as the time-series difference residual vector.

[0017] This embodiment details the specific process of the multi-modal time-series difference feature analysis module generating a time-series difference residual vector; in order to quantify the evolution of the lesion over time, this embodiment introduces a difference calculation model based on baseline decay, and the specific calculation logic is shown in the following formula: ; Wherein, , represent the input current image matrix and historical image matrix respectively, and the dimension is , wherein Image height, Image width, The number of image channels is derived from the raw data acquisition of medical imaging equipment. This represents a predefined convolutional neural network (CNN) feature extraction function. This network removes deep global pooling layers to preserve spatial structure information, or uses dilated convolutional structures to extract multi-scale morphological and geometric features. Its function is to map high-dimensional pixel data into feature tensors, with output dimensions of... ,in The dimension representing the feature vector; The baseline attenuation coefficient is a dimensionless scalar. In this embodiment, it is derived from a preset value set based on clinical experience, with an initial value of 1.0, used to adjust the reference weight of historical images for the current diagnosis. Specifically, this coefficient is related to the time interval. A negative correlation exists, for example, following The decay function, where The attenuation rate is used to characterize the physical decay characteristics of the reference value of historical images over time. This represents the output time-series difference residual vector, with dimensions of... , is the core input variable for subsequent modules to perform entropy flow control and dynamic programming; This indicates that tensors are multiplied element-wise; This indicates that the tensor is subtracted element-wise, and its operational logic is consistent with the element-wise subtraction operation described in the steps, ensuring the time-series difference residual vector. Calculation; By introducing a baseline attenuation coefficient This system can not only calculate simple pixel differences, but also flexibly adjust the reference weights of historical data according to clinical needs. This difference calculation based on depth feature space rather than pixel space effectively filters out non-pathological differences caused by equipment noise or slight positional differences, resulting in a higher quality residual vector. It purely reflects the biological changes of the lesion.

[0018] Example 3 The semantic entropy flow conservation and negative entropy injection control module determines the system's net entropy target, specifically including the following steps: Calculate the conditional entropy value for generating a sequence of diagnostic report texts given current medical image data; Calculate the mutual information value between the diagnostic report text sequence and the time-series difference residual vector; Obtain the negative entropy injection factor, which is a preset scalar used to control the attention given to dynamic information; Calculate the product of the mutual information value and the negative entropy injection factor; The difference between the calculated conditional entropy value and the product is determined as the system net entropy target.

[0019] Specifically, the embodiment constructs a discriminator neural network to estimate the lower bound of mutual information, based on the MINE algorithm idea, and the calculation formula is as follows: ; Wherein, is the batch size; is the sample joint distribution sample pair, that is, the matched text and image difference; is the edge distribution sample pair generated by randomly shuffling the order, that is, the unmatched text and image difference; the discriminator is optimized by maximizing the above formula, thereby outputting the approximate mutual information value; The embodiment details the process of determining the system net entropy target by the semantic entropy flow conservation and negative entropy injection control module; in order to prevent the generated text from losing key image details, that is, modal collapse, the embodiment constructs an entropy flow model based on Shannon information theory, and the specific mathematical model is as follows: ; Wherein, represents the diagnostic report text sequence to be generated; represents the conditional entropy of generating text under the condition of the given current image , the unit is bit or nat, and the value is derived from the probability distribution calculation of the generation model, representing the uncertainty of generation; represents the mutual information between the generated text and the time series difference residual vector , the unit is bit or nat; this value represents the amount of effective information about image dynamic changes contained in the text; represents the negative entropy injection factor, which is a dimensionless scalar; it is derived from the model hyperparameter optimization process and is used to control the degree of attention of the system to the changing details; represents the system net entropy target, which is part of the loss function of system optimization; The formula defines a clear optimization direction: by maximizing the mutual information , that is, reducing , the system forces the model to inject the extracted difference features into the generated text; this mechanism prevents the model from generating hollow and general descriptions such as unseen obvious abnormalities from the bottom logic of information theory, ensuring the substantive content of the report.

[0020] Embodiment 4 The residual-driven dynamic programming generation module generates a sequence of diagnostic report texts, specifically including the following steps: Obtain the hidden state vector of the decoder at the current time step, and transform it using the linear transformation matrix of the output layer to obtain the first logical vector; The learnable residual mapping matrix is ​​invoked to project the temporal difference residual vector onto the vocabulary space to obtain the second logical vector; Obtain the residual gating weight, which is a scalar with a value between 0 and 1, and use this weight to perform weighted processing on the second logic vector; The first logical vector is summed with the weighted second logical vector, and the Softmax activation function is applied to the sum to obtain the word probability distribution and generate vocabulary accordingly.

[0021] This embodiment describes the specific implementation of the residual-driven dynamic programming generation module for generating diagnostic report text sequences. To directly convert physical layer image changes into lexical selection at the language layer, this embodiment constructs a probability distribution model based on residual gating. The specific probability generation formula is as follows: ; in, Indicates the decoder is in The hidden state vector at time t, with dimensions of This comes from real-time state updates from the RNN or Transformer decoder; This represents the linear transformation matrix of the output layer, with dimensions of This is used to map hidden states to a vocabulary space; This represents the temporal difference residual vector from the module, with dimensions of... ; Denotes the residual mapping matrix, with dimensions of , is a learnable parameter; its function is to directly project the residual features of the image space onto the vocabulary space; · represents matrix multiplication; in The middle represents scalar multiplication; This represents the residual gating weight, which is a value within a certain range. The dimensionless scalar between them comes from the calculation of the adaptive gating network, and the specific calculation formula is as follows: ; in, This represents the hidden state vector of the decoder at the current time step. With time-series difference residual vector Perform the splicing operation; Let the learnable weight matrix of the gated network have the following dimensions: ; here Represents the weight matrix concatenate the transpose of the vector with a matrix multiplication, and concatenate the result as a row vector, resulting in a 1x1 scalar; is a gating bias term; is a Sigmoid activation function, which is used to constrain the output value in the interval (0, 1), so as to realize the adaptive adjustment of the intensity of the dynamic feature injection of the image; denotes a Softmax activation function, which is used to output the final dimensionless probability distribution; By introducing the residual mapping matrix and the gating weight , the system not only refers to the context semantics when generating each word , but also directly superimposes the change characteristics of the image ; This dual driving mechanism ensures that when the image changes significantly, the value is large, and the generation probability of the related descriptive words will be explicitly improved, thereby ensuring the accuracy of the report generation.

[0022] Embodiment 5 The diagnostic specificity and normativity collaborative optimization module constructs a composite loss function, specifically including the following steps: Calculate the probability distribution difference between the diagnostic report text sequence and the real report, and generate a standard cross-entropy loss value; Based on the hinge loss principle, calculate the specificity penalty term for dynamic change feature description; Obtain the specificity penalty coefficient corresponding to the current training round, which is a scalar that changes dynamically with the training process; Calculate the product of the specificity penalty term and the specificity penalty coefficient, and add the product to the standard cross-entropy loss value to obtain the composite loss function.

[0023] The specificity penalty term is calculated, specifically including the following steps: Search in the diagnostic report text sequence to identify key phrases matching the preset keyword set, which contains pre-stored words describing the dynamic changes of the lesion morphology; Obtain the hidden layer state vector corresponding to the key phrase in the decoder as the text semantic feature representation; Call the cross-modal alignment projection matrix to linearly map the time difference residual vector to the text semantic space to obtain the mapped residual vector; Calculate the cosine similarity value between the text semantic feature representation and the mapped residual vector.

[0024] The calculation of the specificity penalty term also includes the following logical judgment steps: obtaining a preset semantic matching threshold, the threshold being a preset scalar value; determining whether the cosine similarity value is lower than the semantic matching threshold; if yes, calculating the difference between the semantic matching threshold and the cosine similarity value, and adding the difference to the specificity penalty term; if no, setting the penalty value for the key phrase to zero.

[0025] The embodiment details the construction of the composite loss function in the diagnostic specificity and normative collaborative optimization module and its calculation logic; in order to solve the trade-off contradiction between text fluency and diagnostic accuracy, the system designs a composite objective function containing a specificity penalty term; the overall definition of the composite loss function is as follows: ; wherein, represents the standard cross-entropy loss value, which is used to constrain the probability distribution difference between the generated sequence and the true report , and ensure the grammatical normativity of the language; represents the corresponding true labeled report text sequence; represents the specificity penalty coefficient, which is a dimensionless scalar that changes dynamically with the training process; represents the total number of key phrases retrieved in the current training sample; represents the index of the th key phrase; represents the specificity penalty term for the th key phrase; It should be noted that during the model training phase, in order to ensure the differentiability of the computational graph and the continuity of the gradient backpropagation, the teacher forcing strategy is adopted; the retrieval in the diagnostic report text sequence specifically refers to retrieving the position index of the key phrase in the true labeled report sequence, and extracting the hidden layer state vector generated by the decoder when generating the word corresponding to the position index for subsequent calculation, rather than directly using the predicted output text sequence; when the key phrase consists of multiple word units, the hidden layer state vectors of all time steps covered by the phrase are averaged or the state vector of the last time step is taken, so that they are aggregated into a single vector with a dimension of The specific calculation logic of the specificity penalty term is as follows: ; wherein, key phrase retrieval: the system retrieves the key phrase matching the preset key word set such as increase, decrease, and new in the generated text sequence; denotes the corresponding hidden layer state vector of the key phrase in the decoder, with dimensionality , i.e., the text semantic feature dimension , and as its text semantic feature representation; denotes the time-difference residual vector from the multi-modal time-difference feature analysis module, with dimensionality ; denotes the cross-modal alignment projection matrix, which is a learnable parameter, with dimensionality ; its role is to linearly map the image space residual vector to the text semantic space, solving the dimension inconsistency problem; Here, · denotes matrix multiplication, which maps the image residual vector from dimension to dimension; denotes the cosine similarity function, which is used to calculate the consistency between the text vector and the mapped image residual ; denotes the semantic matching threshold, which is a preset dimensionless scalar; in this embodiment, the value is set to 0.7, and this threshold is determined by hyperparameter search on the validation set; specifically, the cosine similarity boundary when the F1-Score reaches the peak is selected as the value of to ensure that the cross-modal alignment projection matrix can effectively distinguish between synonymous descriptions and irrelevant descriptions after mapping the feature space distribution; denotes the hinge loss logic; only when the similarity is lower than the threshold , a positive gradient penalty is generated; This composite loss function achieves mathematical alignment between the text semantic space and the image feature space through the projection matrix ; the hinge loss mechanism creates a tolerance interval, as long as the generated description is sufficiently similar to the image feature, greater than , no penalty is imposed; otherwise, the model is forced to correct; this mechanism ensures that the report grammar is correct, controlled by , while forcing the key diagnostic words in the report to be highly consistent with the actual changes in the image, achieving mathematical synergy of specificity and standardization.

[0026] Embodiment 6 The diagnostic specificity and standardization synergy optimization module is further configured to perform the following parameter adjustment strategy: In the initial stage of system training, the specificity penalty coefficient is set to a first preset value; As the number of training rounds increases, the specificity penalty coefficient is gradually increased according to the preset annealing strategy until it reaches a second preset value that is greater than the first preset value, so as to enhance the accuracy of the model in describing the temporal difference residual vector in the later stage of training.

[0027] This embodiment describes the parameter adjustment strategy in the diagnostic specificity and standardization co-optimization module; to simulate the learning curve of human doctors learning language first and then diagnosis, this system adjusts the specificity penalty coefficient. A dynamic annealing strategy was adopted; Initial phase settings: In the initial phase of system training, such as the first 10 epochs, the specificity penalty coefficient is set... Set it to the first preset value, for example, 0.01; at this time, the loss function is mainly composed of cross-entropy loss. The model primarily learns the grammatical structure and common terminology of medical reports. Dynamic scaling: As the training rounds increase, the scale gradually increases according to a preset annealing strategy, such as linear or exponential growth. ; Later enhancements: until Once the second preset value is reached, for example, 1.0, the weight of the specificity penalty term in the total loss increases significantly, and the model will be forced to focus on image features. Consistency with the text description, thereby correcting the earlier all-purpose description; This dynamic adjustment strategy avoids the problem of model convergence difficulties or incoherent sentences generated due to overemphasis on feature matching in the early stages of training. By progressively introducing specific constraints, this system achieves a smooth transition in model capabilities, ultimately obtaining a report generation model with both high language quality and high diagnostic accuracy.

[0028] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A medical image report generation information planning system, characterized in that, include: The multimodal temporal difference feature parsing module is used to acquire the patient's current medical image data and historical image data, use a rigid registration algorithm to unify the two to the same spatial coordinate system, and call a convolutional neural network to extract features to generate temporal difference residual vectors. The semantic entropy flow conservation and negative entropy injection control module is used to calculate the mutual information between the temporal difference residual vector and the text sequence to be generated, and to calculate the system net entropy target in combination with the preset negative entropy injection factor. The residual-driven dynamic programming generation module projects the temporal difference residual vectors onto the vocabulary space and combines the hidden state of the decoder at the current time with the residual gating weights to generate a diagnostic report text sequence. The diagnostic specificity and standardization co-optimization module is used to construct a composite loss function that includes a cross-entropy loss term and a specificity penalty term, and to iteratively update the system parameters based on the composite loss function. The specificity penalty term is used to constrain the accuracy of the generated text in describing the dynamic changes of the image.

2. The medical image report generation information planning system according to claim 1, characterized in that, The multimodal temporal difference feature parsing module generates temporal difference residual vectors, specifically including the following steps: Call the preset convolutional neural network feature extraction function to process the current medical image data and historical image data respectively, and output the current feature tensor and the historical feature tensor. Obtain the preset baseline attenuation coefficient, which is a scalar value used to adjust the weight of historical data; The weighted historical feature tensor is obtained by multiplying the historical feature tensor by the baseline attenuation coefficient. Perform an element-wise subtraction operation between the current feature tensor and the weighted historical feature tensor, and define the difference obtained by the operation as the temporal difference residual vector.

3. The medical image report generation information planning system according to claim 1, characterized in that, The semantic entropy flow conservation and negative entropy injection control module determines the system's net entropy target, specifically including the following steps: Calculate the conditional entropy value for generating a sequence of diagnostic report texts given current medical image data; Calculate the mutual information value between the diagnostic report text sequence and the time-series difference residual vector; Obtain the negative entropy injection factor, which is a preset scalar used to control the attention given to dynamic information; Calculate the product of the mutual information value and the negative entropy injection factor; The difference between the conditional entropy value and the product is calculated, and this difference is determined as the target net entropy of the system.

4. The medical image report generation information planning system according to claim 1, characterized in that, The residual-driven dynamic programming generation module generates a sequence of diagnostic report texts, specifically including the following steps: Obtain the hidden state vector of the decoder at the current time step, and transform it using the linear transformation matrix of the output layer to obtain the first logical vector; The learnable residual mapping matrix is ​​invoked to project the temporal difference residual vector onto the vocabulary space to obtain the second logical vector; Obtain the residual gating weight, which is a scalar with a value between 0 and 1, and use this weight to perform weighted processing on the second logic vector; The first logical vector is summed with the weighted second logical vector, and the Softmax activation function is applied to the sum to obtain the word probability distribution and generate vocabulary accordingly.

5. The medical image report generation information planning system according to claim 1, characterized in that, The diagnostic specificity and standardization synergistic optimization module constructs a composite loss function, specifically including the following steps: Calculate the probability distribution difference between the diagnostic report text sequence and the actual report, and generate a standard cross-entropy loss value; Based on the hinge loss principle, a specific penalty term for describing dynamic changes is calculated. Obtain the specificity penalty coefficient corresponding to the current training round, which is a scalar that changes dynamically with the training process; Calculate the product of the specificity penalty term and the specificity penalty coefficient, and add this product to the standard cross-entropy loss value to obtain the composite loss function.

6. The medical image report generation information planning system according to claim 5, characterized in that, The calculation of the specific penalty term includes the following steps: The search is performed on the text sequence of the diagnostic report to identify key phrases that match a preset keyword set, which contains pre-stored words used to describe the dynamic changes in the morphology of lesions. Obtain the hidden layer state vectors corresponding to key phrases in the decoder as text semantic feature representations; The cross-modal alignment projection matrix is ​​invoked to linearly map the temporal difference residual vector to the text semantic space, resulting in the mapped residual vector. Calculate the cosine similarity between the semantic feature representation of the text and the mapped residual vector.

7. The medical image report generation information planning system according to claim 6, characterized in that, Calculating the specificity penalty term also includes the following logical judgment steps: Obtain a preset semantic matching threshold, which is a preset scalar value; Determine whether the cosine similarity score is lower than the semantic matching threshold; If so, calculate the difference between the semantic matching threshold and the cosine similarity value, and include this difference in the specificity penalty term; If not, the penalty value for that key phrase will be set to zero.

8. The medical image report generation information planning system according to claim 1, characterized in that, The diagnostic specificity and standardization co-optimization module is also configured to implement the following parameter tuning strategy: In the initial stage of system training, the specificity penalty coefficient is set to the first preset value; As the number of training rounds increases, the specificity penalty coefficient is gradually increased according to the preset annealing strategy until it reaches a second preset value that is greater than the first preset value, so as to enhance the accuracy of the model in describing the temporal difference residual vector in the later stage of training.

Images