A weakly supervised head and neck MRI report generation method based on anomaly detection

Abstract

The application provides a weakly supervised head and neck MRI report generation method based on anomaly detection, comprising: according to the enhancement intensity distribution and the necrosis area proportion, by comparing the prototype samples in the preset stage evolution library, the lesion evolution stage mark and the preliminary classification are obtained; the dominant degree of early solid enhancement performance in the preliminary classification is analyzed, and the matching degree of the lesion type and the early prototype in the prototype library is evaluated; the lesion evolution stage mark is fused through the matching degree correction mechanism to obtain the stage similarity score; the disease progression description field in the report generation template is determined according to the stage similarity score, the staging detailed content is obtained by embedding the necrosis area proportion and the lesion evolution stage mark; the late necrosis cystic change proportion is obtained by analyzing the staging detailed content, and the lesion real type recognition result is obtained by combining the enhancement intensity distribution and the stage similarity score.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for generating weakly supervised head and neck MRI reports based on anomaly detection. Background Technology

[0002] In the field of medical image analysis, particularly in report generation research of head and neck MRI images, the importance of automation technology is self-evident. This technology not only reduces the workload of physicians but also improves the efficiency and accuracy of diagnosis, providing irreplaceable value for timely patient treatment and disease management. With the continuous advancement of imaging technology, weakly supervised methods based on anomaly detection have gradually become a research hotspot, aiming to identify and describe lesion types using limited labeled data. However, existing methods often struggle to address the dynamic changes of lesions at different time points when faced with complex lesion characteristics. Many techniques focus more on the morphology or signal intensity of the lesion surface, neglecting the evolutionary patterns of the lesion's interior over time. This neglect leads to a reliance on comparisons with existing standards when determining lesion types, lacking a comprehensive consideration of the lesion's development process, thus affecting the comprehensiveness and reliability of the diagnosis. A deeper technical challenge lies in the fact that the differences in the manifestation of lesions at different stages of evolution can significantly interfere with type matching. The stage of lesion evolution is a key factor, as the same lesion may exhibit significant signal enhancement in the early stages, while in later stages it may show internal necrosis or structural changes. This stage-specific change makes comparisons with fixed standards unreliable. For example, a lesion may initially resemble the standard characteristics of one type, but in later stages, its characteristics may become more similar to another type. This fluctuation in characteristics due to different stages directly affects the accurate identification of lesion types and the complete description of disease progression. Therefore, how to fully consider the dynamic changes in the evolutionary stages of lesion type identification and adjust the judgment criteria according to the manifestations of different stages has become a key issue in improving the accuracy and completeness of report generation. This issue is not only related to the technical implementation, but also directly relates to whether doctors can obtain comprehensive information about disease progression when reading reports, thereby affecting the formulation of subsequent treatment plans. Summary of the Invention

[0003] This invention provides a weakly supervised head and neck MRI report generation method based on anomaly detection, comprising: Deep information of the lesion area is extracted from head and neck MRI by anomaly detection, and convolutional neural network is used to identify the internal texture structure of the lesion to obtain the enhancement intensity distribution and the proportion of necrotic area. Based on the enhanced intensity distribution and the proportion of necrotic area, the disease evolution stage markers and preliminary classifications are obtained by comparing with prototype samples in the preset stage evolution library. The degree of dominance of early solid enhancement in the preliminary classification was analyzed, and the matching degree between the obtained lesion types and early prototypes in the prototype library was assessed. The stage similarity score is obtained by fusing the disease evolution stage markers through a matching degree correction mechanism; Based on the stage similarity score, the disease progression description field in the report generation template is determined, and detailed stage information is obtained by embedding the proportion of necrotic area and the disease evolution stage marker. The proportion of late-stage necrosis and cystic degeneration was obtained by analyzing the detailed stage information. Combined with the enhancement intensity distribution and the stage similarity score, the true lesion type identification result was obtained. Based on the results of the identification of the true type of lesion, a generative adversarial network is used to integrate the deep information and the markers of the lesion evolution stage, evaluate the completeness of the progression description in the report, and form a final diagnostic report.

[0004] Furthermore, the step of extracting deep information of the lesion area from head and neck MRI through anomaly detection, and using a convolutional neural network to identify the internal texture structure of the lesion to obtain the enhancement intensity distribution and the proportion of necrotic area includes: Abnormal signal regions are extracted from T1-weighted enhanced sequences of head and neck MRI. The texture depth changes of the abnormal signals are analyzed by a three-dimensional convolutional neural network to obtain a signal gradient distribution map. The boundary of the solid enhancement region is determined based on the signal gradient distribution map, and the solid proportion is calculated. For regions where the solid percentage is lower than a preset threshold, a region growing algorithm is used to identify the necrosis morphology of the central low-signal region, determine the liquefaction degree index, mark the completely liquefied necrosis region, and obtain the accurate value of the necrosis region percentage. Based on the precise value of the proportion of the necrotic region and the boundary of the solid enhancement region, multi-scale feature vectors are extracted, the signal intensity decreasing gradient is determined, and the enhancement intensity distribution map is obtained. By performing frequency domain transformation on the enhanced intensity distribution map, the signal contributions of the solid enhancement component and the cystic necrosis component are separated, the complexity score is calculated, and a deep information set containing the enhanced intensity distribution and the proportion of the necrotic region is output.

[0005] Furthermore, based on the enhanced intensity distribution and the proportion of necrotic area, by comparing with prototype samples in a preset stage evolution library, lesion evolution stage markers and preliminary classifications are obtained, including: Based on the enhanced intensity distribution and the proportion of necrotic area, a lesion feature vector is constructed. A prototype feature set is extracted from a preset stage evolution library. The matching distance value between the lesion feature vector and the prototype feature is calculated. The three prototype samples with the smallest matching distance value are selected as the candidate set. For each prototype sample in the candidate set, stage templates and time markers are extracted, the width of the annular enhancement band, the edge gradient value and the liquefaction depth index are analyzed, the dominant component type is determined, the degree of difference between the necrosis ratio and the pathological indicators in the stage template is calculated, the similarity weight value is obtained, and the time node is determined according to the time markers. By matching the time node with the staging criteria in the preset stage evolution library, and judging the stage marker of the disease evolution based on the degree of conformity between the disease feature vector and the staging criteria, the corresponding classification result is obtained from the type label of the prototype sample.

[0006] Furthermore, the step of obtaining disease evolution stage markers and preliminary classification based on the enhanced intensity distribution and the proportion of necrotic area by comparing with prototype samples in a preset stage evolution library also includes: The width and signal peak of the annular enhancement band around the lesion are extracted by the enhanced intensity distribution. The continuity of the solid enhancement region is identified by morphological dilation and erosion operations. The difference between the edge gradient value of the solid enhancement region and the edge gradient value of the early solid dominant sample in the prototype library is calculated to determine the boundary clarity score and solid region integrity characteristics. The three-dimensional morphological parameters of the central necrotic cavity, including the major axis, minor axis, and volume ratio, are extracted from the necrotic area proportion data. The standard deviation of the signal inside the necrotic cavity is calculated to determine the liquefaction depth index. The liquefaction depth index is compared with the standard value of liquefaction depth of the late cystic degeneration dominant sample to obtain the cystic degeneration degree assessment value. Based on the boundary clarity score and the cystic degeneration degree assessment value, the component proportion coefficient is calculated to determine whether the solid component or the necrotic component is dominant, and the component dominance type is obtained. By querying a preset evolution stage reference table based on the component-dominant type, and determining the stage marker based on the combination of the solid region integrity characteristics, the liquefaction depth index, and the cystic degeneration degree assessment value, the corresponding classification index is extracted from the evolution stage reference table, and the lesion evolution stage marker and preliminary classification result are output.

[0007] Furthermore, the dominance of early solid enhancement in the preliminary classification was analyzed, and the matching degree between the obtained lesion types and early prototypes in the prototype library was assessed, including: The feature vectors of early solid enhancement are extracted from the preliminary classification. The feature vectors include three dimensions: the area ratio of the enhanced region, the mean signal intensity, and the boundary regularity. The difference between the value of each dimension and the preset dominant threshold is calculated to obtain the basic feature set of the lesion type. By comparing the basic feature set with the prototype features of early prototypes in the prototype library item by item, the similarity score of the feature vector is calculated. The similarity score is then multiplied by a preset category weight coefficient to obtain a weighted matching value sequence. Based on the maximum value in the weighted matching value sequence, an early prototype identifier is determined. By querying the pathological type and evolution stage information corresponding to the identifier in the prototype library, the maximum weighted matching value is taken as the final matching degree between the lesion type and the early prototype.

[0008] Furthermore, the step of fusing the lesion evolution stage markers through a matching degree correction mechanism to obtain a stage similarity score includes: Based on the matching degree value and the disease evolution stage marker, extract the temporal location information and disease progression degree value from the stage marker, calculate the absolute value of the difference between the matching degree and the preset benchmark value, and obtain the initial correction parameter by the reciprocal of the difference; The weighted average method is used to multiply the initial correction parameter by the confidence score of the evolution stage indicator, and then combine it with the preset category fusion weight to obtain the corrected stage matching factor. The stage similarity score is determined by multiplying the stage matching factor with the normalized temporal position information and mapping the product result to a preset scoring interval.

[0009] Furthermore, based on the stage similarity score, the disease progression description field in the report generation template is determined. By embedding the proportion of necrotic area and the disease evolution stage marker, detailed staging information is obtained, including: Based on the stage similarity score, query the preset report template library, select the report frame corresponding to the score range, the report frame includes a description of the disease stage, a rating of the rate of evolution and a prognostic trend judgment, and obtain the report frame to be filled; By embedding the percentage value of the necrotic area and the standardized description of the disease evolution stage markers through the preset data insertion points in the report framework, the numerical values ​​are connected with the text descriptions to generate mixed text paragraphs. Using the aforementioned mixed text paragraphs, the evolutionary stage markers are converted into standardized staging terms based on the clinical staging comparison table. The level of report detail is determined by the proportion of necrotic area, and the detailed staging content containing complete pathological staging information is output.

[0010] Furthermore, by analyzing the detailed staging information to obtain the proportion of late-stage necrosis and cystic degeneration, and combining the enhancement intensity distribution with the stage similarity score, the true lesion type identification result is obtained, including: Extract descriptions related to necrosis and cystic degeneration from the detailed staging information, calculate the ratio of the volume of the necrotic and cystic degeneration area to the total volume of the lesion, and obtain the percentage of late-stage necrosis and cystic degeneration and late-stage characteristic markers. Based on the late-stage necrosis and cystic degeneration proportion and the late-stage feature markers, combined with the annular enhancement width and the area of ​​the central non-enhanced region in the enhancement intensity distribution, the late-stage proportion and the annular feature value are multiplied by the corresponding preset weight coefficients and then added together to calculate the comprehensive feature value. The comprehensive feature value is then multiplied by the stage similarity score to obtain the type determination index. The pathological type is identified by querying a preset pathological type range table using the type determination index, determining the corresponding pathological type identifier based on the numerical range of the index, performing late-stage correction on the pathological type, and outputting the true type identification result of the lesion.

[0011] Furthermore, based on the results of the true lesion type identification, a generative adversarial network is used to integrate the deep information and the lesion evolution stage markers to assess the completeness of the progression description in the report, forming a final diagnostic report, including: Based on the identification results of the true type of lesion, an initial report framework is constructed. The enhancement intensity distribution value, necrosis area ratio and boundary features in the deep information are encoded into feature vectors and mapped to the text space through a fully connected layer to obtain a preliminary text sequence of pathological description. The preliminary text sequence is compared and matched with the disease progression stage markers, and the relevance score of each paragraph in the text sequence is evaluated by a discriminant network to obtain the completeness score of the progression description. Based on the completeness score, description fragments corresponding to the corresponding stage are extracted from a preset standard description library, and the fragments are embedded into the preliminary text sequence to obtain a structurally complete diagnostic text. Using the diagnostic text, combined with the type identifier and staging information in the lesion type identification results, a final diagnostic report containing complete pathological information is output.

[0012] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a weakly supervised head and neck MRI report generation method based on anomaly detection. Addressing the challenges of complex internal texture structures in lesion areas, difficulty in identifying evolutionary stages, and inaccurate classification and progression descriptions in clinical scenarios, this invention uses a convolutional neural network to identify the distribution of lesion enhancement intensity and the proportion of necrotic areas. Combined with comparison to a pre-set prototype sample library, it determines the markers of lesion evolution stages and provides a preliminary classification. Finally, a matching degree correction mechanism is used to fuse stage similarity scores, ultimately generating a diagnostic report containing a description of disease progression. This invention innovatively integrates generative adversarial networks to optimize deep information and stage markers, ensuring report completeness and accuracy, significantly improving the precision of lesion type identification and staging, and providing reliable support for clinical diagnosis. Attached Figure Description

[0013] Figure 1This is a flowchart of a weakly supervised head and neck MRI report generation method based on anomaly detection according to the present invention.

[0014] Figure 2 This is a schematic diagram of a weakly supervised head and neck MRI report generation method based on anomaly detection according to the present invention.

[0015] Figure 3 This is another schematic diagram of a weakly supervised head and neck MRI report generation method based on anomaly detection according to the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figures 1-3 This embodiment of a weakly supervised head and neck MRI report generation method based on anomaly detection may specifically include: S101. Extract deep information of the lesion area from head and neck MRI through anomaly detection, and use convolutional neural network to identify the internal texture structure of the lesion to obtain the enhancement intensity distribution and the proportion of necrotic area.

[0018] Abnormal signal regions are extracted from T1-weighted enhanced sequences of head and neck MRI. The texture depth changes of these abnormal signals are analyzed using a three-dimensional convolutional neural network to obtain a signal gradient distribution map from the periphery to the center of the lesion. Based on the boundary between high-signal and low-signal regions in the signal gradient distribution map, the boundary of the solid enhancement region is determined, and the solid proportion is calculated based on the ratio of the number of voxels in the high-signal region to the total number of voxels in the lesion. For regions where the solid proportion is below a preset threshold, a region growing algorithm is used to identify the necrosis morphology of the central low-signal area. By calculating the signal uniformity and boundary sharpness within the necrosis morphology, a liquefaction index is determined. If the liquefaction index exceeds a preset threshold, it is marked as a completely liquefied necrosis area, obtaining a precise value for the necrosis region proportion. Based on the precise value of the necrosis region proportion and the solid enhancement region boundary, multi-scale feature vectors of the internal structure of the lesion are extracted. The hierarchical distribution pattern of signal intensity in these multi-scale feature vectors is analyzed to determine the signal intensity decreasing gradient from the periphery to the center of the lesion, obtaining an enhancement intensity distribution map reflecting the evolution characteristics of the lesion. By performing frequency domain transformation on the enhanced intensity distribution spectrum, the signal contributions of the solid enhancement component and the cystic necrosis component are separated using fast Fourier transform. The complexity score is calculated based on the energy distribution of the two components in the frequency domain. The complexity score is determined by the ratio of the proportion of the solid component to the proportion of the necrosis component. The output is a set of deep information containing the enhanced intensity distribution and the proportion of the necrosis region.

[0019] In one implementation, when extracting abnormal signal regions from T1-weighted enhanced sequences of head and neck MRI, a three-dimensional convolutional neural network is used to process the image data. The three-dimensional convolutional neural network contains five convolutional layers, each with a kernel size of 3×3×3, and analyzes the texture depth of the lesion region by extracting spatial features layer by layer. After contrast agent injection, the lesion tissue exhibits varying degrees of enhancement due to increased vascular permeability, forming a signal gradient that decreases from the periphery to the center. By setting different window width and window level parameters, fine grayscale change information within the lesion is obtained, generating a signal gradient distribution map.

[0020] Specifically, the generation process of the signal gradient distribution map includes calculating the signal intensity difference between each voxel and its adjacent voxels, and obtaining gradient components in three directions using a gradient operator. In the lesion periphery, due to the abundance of solid components and rich blood supply, the signal gradient value is larger; while in the lesion center, if necrosis or liquefaction exists, the signal gradient value is significantly reduced. Based on the distribution characteristics of the gradient values, a threshold is set to divide the lesion area into high-signal and low-signal regions, and the boundary between the two is the boundary of the solid enhancement region.

[0021] It should be noted that the calculation of the solid component percentage is based on the ratio of the number of voxels in the high-signal region to the total number of voxels in the lesion. When head and neck tumors develop to a certain stage, ischemic necrosis occurs in the central region due to insufficient internal blood supply. This necrotic area appears as low signal or no enhancement on contrast-enhanced scans, forming a clear contrast with the surrounding solid components. By statistically analyzing the percentage of high-signal voxels relative to the total number of voxels in the lesion, the solid component percentage, reflecting the internal composition of the lesion, is obtained.

[0022] In one possible implementation, a region growing algorithm is used to identify necrosis morphology in lesion areas where the solid content is below a preset threshold. The region growing algorithm starts from a seed point at the center of the lesion and expands based on the signal intensity similarity of adjacent voxels. Signal uniformity is determined by calculating the standard deviation of the signal intensity of all voxels within the necrotic region; a smaller standard deviation indicates more uniform signal. Boundary sharpness is assessed by calculating the signal intensity gradient between the necrotic area and surrounding tissue; a larger gradient indicates a clearer boundary. A liquefaction index comprehensively considers both signal uniformity and boundary sharpness; when both reach a certain standard, it is determined to be complete liquefactive necrosis.

[0023] Preferably, the multi-scale feature vector extraction adopts a pyramid structure, analyzing lesion features at different resolution levels. The first level maintains the original resolution and extracts detailed texture features; the second level downsamples to half the original resolution to obtain medium-scale structural features; the third level further downsamples to capture overall morphological features. The features extracted at each level are integrated through a feature fusion mechanism to form a feature vector containing information at different scales. This multi-scale analysis method can simultaneously capture both local details and overall features of lesions, improving the ability to identify complex lesions.

[0024] For example, the process of determining the signal intensity decrease gradient involves performing concentric circle layering analysis on the lesion. Starting from the edge of the lesion, the process proceeds towards the center at fixed intervals, calculating the average signal intensity on each concentric circle layer. Under normal circumstances, the signal intensity of solid tumors gradually decreases from the periphery to the center; while cystic lesions show peripheral ring enhancement with extremely low signal intensity in the center. By fitting a curve of signal intensity changing with radial distance, a linear function S(r) = a*r + b is used, where S(r) is the signal intensity at a radial distance r, a is the slope, and b is the intercept; a is calculated as the decrease gradient parameter, which reflects the internal structural characteristics and blood supply distribution pattern of the lesion.

[0025] In one embodiment, the frequency domain transformation employs a Fast Fourier Transform (FFT) to convert the enhancement intensity distribution spectrum from the spatial domain to the frequency domain. In the frequency domain, the solid enhancement component mainly exhibits low-frequency components, while the cystic necrosis component, due to its inhomogeneous internal structure, generates more high-frequency components. By designing a low-pass Butterworth filter with a cutoff frequency of 0.2 times the total frequency and a high-pass Butterworth filter with a cutoff frequency of 0.8 times the total frequency, the energy spectra of the low-frequency and high-frequency components are extracted, respectively. The proportion of the solid component is obtained by calculating the ratio of low-frequency energy to total energy, while the proportion of the necrosis component corresponds to the relative proportion of high-frequency energy.

[0026] For example, in the early stages of a lesion, solid components dominate, with low-frequency energy accounting for over 70%. As the lesion progresses, the central necrotic area expands, and high-frequency energy gradually increases. In the late stages, after cystic degeneration, the spectral characteristics show a distinct bimodal distribution. The complexity score is calculated based on the entropy value of the spectral distribution; a higher entropy value indicates a more complex internal structure of the lesion. By comprehensively analyzing the frequency domain characteristics, a deep information set containing the enhancement intensity distribution and the proportion of necrotic areas is generated, providing a quantitative basis for subsequent lesion staging and type identification.

[0027] S102. Based on the distribution of enhanced intensity and the proportion of necrotic areas, obtain the disease evolution stage markers and preliminary classification by comparing with prototype samples in the preset stage evolution library.

[0028] A lesion feature vector is constructed based on the enhanced intensity distribution and the proportion of necrotic areas. A prototype feature set of prototype samples is extracted from a pre-defined stage evolution library. The sum of squared differences between each dimension of the lesion feature vector and the corresponding dimension of the prototype features is calculated to obtain a matching distance value. The three prototype samples with the smallest matching distance values ​​are selected as a candidate set. For each prototype sample in the candidate set, its corresponding stage template and time series marker are extracted. The dominant component type is determined by analyzing the width of the annular enhancement band, calculating the edge gradient value, and evaluating the liquefaction depth index. The degree of difference between the proportion of necrotic areas in the lesion feature vector and the pathological indicators in the stage template is calculated to obtain a similarity weight value. If the similarity weight value exceeds a pre-defined threshold, the time node in the lesion's evolution process is determined based on the time series marker. The time node is matched against the staging standards in the pre-defined stage evolution library. Based on the degree of conformity between the lesion feature vector and the staging standards, the lesion evolution stage marker is determined, and the classification result corresponding to the stage marker is obtained from the type labels of the prototype samples.

[0029] In one implementation, the lesion feature vector is constructed based on two core dimensions: enhancement intensity distribution and necrosis area proportion. The enhancement intensity distribution includes the average signal values ​​of three regions: the lesion periphery, the transition zone, and the center, while the necrosis area proportion reflects the degree of liquefactive necrosis within the lesion. A preset stage evolution library stores typical manifestations of different pathological types at various evolution stages, with each prototype sample corresponding to a specific lesion type and development stage.

[0030] Specifically, the matching distance value is calculated using the Euclidean distance metric. For each dimension of the lesion feature vector, the difference between it and the corresponding dimension of the prototype feature vector is calculated. The square root of the sum of the squares of all dimension differences is then taken to obtain the distance between the two feature vectors. The smaller the distance value, the higher the similarity between the current lesion and the prototype sample. By calculating and sorting the distances of all prototype samples, the three with the smallest distances are selected as the candidate set. These three candidate samples represent the pathological type and evolution stage most similar to the current lesion.

[0031] It should be noted that the stage template records typical pathological indicators of the lesion at different stages of its evolution, including parameters such as the proportion of solid components, the degree of necrosis, and the blood supply status. The time sequence marker indicates the relative time position of the prototype sample in the entire lesion evolution process, divided according to percentage from early to late stages.

[0032] Preferably, the calculation of the similarity weight value focuses on the key indicator of the necrosis ratio. When the difference between the necrosis ratio of the lesion and the pathological indicators in the stage template is less than a preset range, the two are considered to be highly consistent, and the similarity weight value will exceed the threshold. At this time, based on the temporal marker of the prototype sample, the specific time node in the evolution process of the current lesion can be determined, which reflects the relative time process experienced by the lesion from its occurrence to its current state.

[0033] In one embodiment, the staging criteria divide the lesion into three main stages: early, intermediate, and late, based on its temporal progression. Early lesions are predominantly solid with less than 20% necrosis; intermediate lesions show significant central necrosis, with necrosis covering 20% ​​to 60% of the area; and late lesions are characterized by extensive cystic degeneration, with necrosis exceeding 60%. By matching determined time points with the staging criteria, the stage of lesion evolution is identified, thus obtaining a preliminary classification result corresponding to that stage.

[0034] By enhancing the intensity distribution, the solid enhancement area and the surrounding ring enhancement pattern inside the lesion are determined. The morphology and liquefaction degree of the central necrotic cavity are collected from the proportion of necrotic area. The boundary clarity between the solid enhancement area and the early solid dominant sample in the prototype library is analyzed. The morphology of the central necrotic cavity and the liquefaction depth of the late cystic dominant sample are assessed. The dominance of solid and necrotic components inside the lesion is determined. The evolutionary stage markers of the lesion and the corresponding preliminary classification results are identified.

[0035] The width and signal peak of the annular enhancement band around the lesion are extracted by enhancing intensity distribution. Morphological dilation and erosion operations are used to identify the continuity of the solid enhancement region. The difference between the edge gradient value of the solid enhancement region and the edge gradient value of the early solid dominant sample in the prototype library is calculated. If the difference is less than a preset threshold, a boundary clarity score is obtained to determine the integrity characteristics of the solid region. The three-dimensional morphological parameters of the central necrotic cavity, including the major axis, minor axis, and volume ratio, are extracted from the necrosis region proportion data. The liquefaction depth index is determined by calculating the signal standard deviation inside the necrotic cavity. The liquefaction depth index is compared with the liquefaction depth standard value stored in the prototype library for late cystic dominant samples to obtain a cystic degree assessment value. Based on the boundary clarity score and the cystic degree assessment value, the component proportion coefficient is calculated by multiplying the boundary clarity score by a preset solid weight and adding the cystic degree assessment value by a preset necrosis weight. When the solid component proportion coefficient is greater than a preset dominance threshold, it is determined to be solid component dominant; otherwise, it is determined to be necrosis component dominant, thus obtaining the component dominance type inside the lesion. By querying a preset evolution stage reference table based on the component-dominant type, and determining the stage marker of the lesion in the evolution process according to the combination of solid region integrity characteristics, liquefaction depth index and cystic degeneration degree assessment value, the classification index corresponding to the stage marker is extracted from the evolution stage reference table, and the evolution stage marker and preliminary classification result of the lesion are output.

[0036] In one implementation, the analysis of the enhancement intensity distribution of the lesion begins with the identification of the annular enhancement band. Annular enhancement is a typical manifestation of head and neck malignancies, and its formation mechanism is closely related to the distribution of tumor neovascularization. Due to the rich blood supply to the periphery of the tumor, a large amount of contrast agent accumulates, forming a high-signal ring; while in the central region, due to ischemia and necrosis, the contrast agent cannot enter, resulting in a low signal. By measuring the width of the annular enhancement band, the invasiveness and growth rate of the tumor can be indirectly reflected. The signal peak value is extracted by setting multiple sampling points on the annular enhancement band and calculating the average signal intensity of each point.

[0037] Specifically, morphological dilation and erosion operations are used to identify the continuity of solid enhancement regions. Dilation uses a structuring element to convolve the image, expanding the boundaries of high-signal regions and filling in internal holes; erosion, conversely, shrinks high-signal regions, removing edge burrs and noise. By performing a closing operation of dilation followed by erosion, adjacent solid regions can be connected to form complete solid enhancement regions. Edge gradient values ​​are calculated using the Sobel operator in three orthogonal directions; larger gradient values ​​indicate clearer boundaries. The prototype library stores a large number of pathologically confirmed early solid-dominant samples, whose edge gradient values ​​serve as reference standards. When the difference between the edge gradient value of the lesion to be detected and the reference value is less than a preset threshold, it indicates that the lesion boundary is clear and the tissue structure is relatively intact.

[0038] It should be noted that the extraction of three-dimensional morphological parameters of the central necrosis cavity involves image segmentation and three-dimensional reconstruction. The major axis is the largest diameter of the necrosis cavity in the axial, coronal, and sagittal directions; the minor axis is the largest diameter perpendicular to the major axis; and the volume ratio is the ratio of the necrosis cavity volume to the total lesion volume. These parameters comprehensively reflect the extent and degree of necrosis. The calculation of the liquefaction depth index is based on the signal distribution characteristics within the necrosis cavity. In completely liquefied areas, the signal is uniform, and the standard deviation is close to zero; while in partially liquefied or solid necrosis areas, due to the presence of residual tissue structures, the signal is non-uniform, and the standard deviation is larger.

[0039] For example, the liquefaction depth standard value for the dominant late-stage cystic samples in the prototype library was established through extensive pathological control studies. When tumors progress to an advanced stage, the central necrotic area completely liquefies, forming a cystic cavity, which appears as a homogeneous watery signal on MRI. By comparing the liquefaction depth index of the lesion under test with the standard value, the degree of cystic degeneration of the lesion can be assessed.

[0040] It should be noted that the boundary clarity score and the cystic lesion assessment value are both normalized to the range of 0 to 1 to ensure the consistency of the dimensions in subsequent calculations. A value close to 1 indicates complete cystic lesion, and a value close to 0 indicates no obvious cystic lesion.

[0041] Preferably, the component proportion coefficient is calculated using a dynamic weighting mechanism. The weights for solid and necrosis are dynamically adjusted based on the overall size of the lesion. For small lesions, the weight of solid components is higher because early lesions are predominantly solid; for large lesions, the weight of necrosis components increases, reflecting the inevitable trend of central necrosis during tumor growth. Specifically, the boundary clarity score is multiplied by the solid weight, and the cystic degeneration assessment value is multiplied by the necrosis weight; the two are then added together to obtain the component proportion coefficient. A preset dominance threshold is typically set at 0.5; values ​​exceeding this threshold are considered to indicate solid component dominance, otherwise, necrosis component dominance. This method accurately reflects the main pathological features of the lesion.

[0042] In one possible implementation, the evolution stage lookup table is a pre-built knowledge base containing the complete evolution of the lesion from early to late stages. The table records characteristic combinations of different stages, including multiple dimensions such as the integrity of the solid region, the depth of liquefaction, and the degree of cystic degeneration. During a query, the system searches the lookup table for the best-matching stage based on the dominant component type of the current lesion.

[0043] For example, when the solid region exhibits clear boundaries and intact structure, with a liquefaction depth index below 0.3 and a cystic degeneration assessment value close to 0, the checklist matches it to the early solid stage; when the liquefaction depth index is between 0.3 and 0.7 and the cystic degeneration assessment value is moderate, it matches to the intermediate mixed stage; and when the liquefaction depth index exceeds 0.7 and the cystic degeneration assessment value is close to 1, it matches to the late cystic degeneration stage. Furthermore, the stage markers not only include period information but also an assessment of the lesion's malignancy. Early solid-dominated lesions typically have lower malignancy and relatively slow growth; late cystic-dominated lesions are often highly malignant and have a poor prognosis. The classification indicators, based on the stage markers and combined with auxiliary information such as the lesion's location and size, provide a preliminary judgment of the pathological type, such as squamous cell carcinoma, adenocarcinoma, or other types.

[0044] Understandably, this multi-dimensional, multi-stage comprehensive assessment method can more accurately identify the evolution of lesions, providing important reference for clinical diagnosis. This method overcomes the limitations of relying solely on morphological characteristics, fully considering the dynamic evolution of lesions, and improving the accuracy and reliability of classification.

[0045] S103. Analyze the dominance of early solid enhancement in the preliminary classification and assess the matching degree between the obtained lesion type and the early prototype in the prototype library.

[0046] Feature vectors representing early solid enhancement are extracted from the preliminary classification results. These feature vectors include three dimensions: the proportion of enhanced region area, the mean signal intensity, and the regularity of the boundary. The differences between the values ​​of each dimension and preset dominant thresholds are calculated. The summation of the inverses of the absolute values ​​of these differences is then normalized to obtain the basic feature set of the lesion type. This basic feature set is then compared item by item with the prototype features of early prototypes in a prototype library. A similarity algorithm is used to calculate the similarity score of the feature vectors. This similarity score is multiplied by a preset category weight coefficient to obtain a weighted matching value sequence. Based on the maximum value in the weighted matching value sequence, an early prototype identifier corresponding to the lesion type is determined. By querying the prototype library for the pathological type and evolution stage information corresponding to this identifier, the maximum weighted matching value is taken as the final matching degree between the lesion type and the early prototype.

[0047] In one implementation, the analysis of preliminary classification results focuses on the quantitative assessment of early solid enhancement. The area ratio of the enhanced region is obtained by calculating the ratio of the number of pixels in the enhanced region to the total number of pixels in the entire lesion area; this indicator directly reflects the spatial distribution characteristics of the solid component. The mean signal intensity is obtained by calculating the arithmetic mean of the signal values ​​of all pixels within the enhanced region. The mean signal intensity of normal solid tissue is usually within a specific range; deviations from this range may indicate a change in the nature of the lesion. Boundary regularity is assessed using a combination of two parameters: roundness and smoothness. A roundness closer to 1 indicates a more regular shape, while smoothness is measured by the standard deviation of the boundary curvature.

[0048] Specifically, the dominance coefficient is calculated using a normalization method. For each dimension, the difference between the actual measured value and the preset dominance threshold is first calculated. The dominance threshold is determined based on statistical analysis of a large number of cases. The reciprocal of the absolute value of the difference reflects the degree of closeness of that dimension to the ideal state. The dominance coefficient is obtained by summing the reciprocals of the three dimensions.

[0049] It should be noted that the similarity between feature vectors is evaluated by calculating the directional consistency between the vectors. Both the basic feature set and the prototype features are represented as multi-dimensional vectors. The similarity between two vectors is measured mathematically. The advantage of this method is that it is not affected by the size of the vectors, but only focuses on the consistency of the feature patterns. It is suitable for handling cases with different lesion sizes but similar feature patterns.

[0050] Preferably, the weighting coefficients for pathological types are set considering the clinical importance of different pathological types. The weighting coefficients for malignant tumors such as squamous cell carcinoma and adenocarcinoma are set to 0.8 to 1.0; the weighting coefficients for benign lesions such as inflammatory pseudotumors are set to 0.3 to 0.5; and the weighting coefficients for borderline lesions are set to 0.5 to 0.7. The weighted matching value obtained by multiplying the similarity score by the weighting coefficients better reflects the actual needs of clinical diagnosis.

[0051] In one embodiment, the prototype library stores typical case features confirmed by pathology. Each prototype is labeled with detailed pathological type, differentiation degree, and prognostic information. When the system identifies the prototype corresponding to the maximum weighted matching value, it directly extracts the complete information of that prototype as a diagnostic reference, and the final matching degree is the maximum value, providing a quantitative basis for subsequent treatment decisions.

[0052] S104. By integrating the disease evolution stage markers through the matching degree correction mechanism, the stage similarity score is obtained.

[0053] Based on the matching degree value and the disease evolution stage markers, the temporal location information and disease progression degree value are extracted from the stage markers. The absolute value of the difference between the matching degree and the preset benchmark value is calculated, and the reciprocal of this difference is used to obtain the initial correction parameter. A weighted average method is used to multiply the initial correction parameter by the reliability score of the evolution stage marker. The reliability score is preset based on the reliability of the stage marker source, and then combined with a preset category fusion weight to obtain the corrected stage matching factor. The stage matching factor is then multiplied by the normalized temporal location information. The normalization process uses a maximum-minimum standardization method. The product result is mapped to a preset scoring interval to determine the stage similarity score.

[0054] In one implementation, the matching degree correction mechanism achieves accurate stage determination through multi-dimensional information fusion. The core components of the evolution stage marker include temporal location information and lesion progression value. Temporal location information represents the relative time point of the lesion throughout the entire evolution process, normalized from 0 to 1, where 0 represents the early stage of the lesion and 1 represents the late, terminal stage. The lesion progression value reflects the rate of lesion deterioration, obtained by comparing the rate of change of lesion characteristics at adjacent time points.

[0055] Specifically, the initial correction parameter is calculated based on the deviation between the matching degree and a preset benchmark value. The preset benchmark value is determined based on statistical analysis of matching degrees from a large number of clinical cases. When the actual matching degree is higher than the benchmark value, it indicates that the lesion characteristics are clear and the correction magnitude is small; when the matching degree is lower than the benchmark value, it indicates that the lesion characteristics are ambiguous and a greater correction magnitude is needed. By calculating the reciprocal of the absolute value of the difference, a reverse adjustment mechanism is implemented, allowing the correction parameter to be dynamically adjusted between 0.5 and 2.

[0056] It should be noted that the reliability score is pre-set based on the reliability of the source of the evolution stage marker. If the stage marker comes from pathological diagnosis results, the reliability score is set to 0.9 to 1.0; if it comes from inference from imaging features, the score is 0.7 to 0.9; if it is based solely on clinical manifestations, the score is 0.5 to 0.7. This grading mechanism ensures a reasonable weight allocation of information from different sources during the fusion process.

[0057] Preferably, the weighted average method employs an adaptive weight allocation strategy. The initial correction parameter is multiplied by the confidence score to obtain the base fusion value. The category importance coefficient is set according to the clinical importance of the lesion type; for example, 0.8 for malignant tumors, 0.5 for benign lesions, and 0.3 for inflammatory lesions. Through the combined calculation of the base fusion value and the category importance coefficient, the resulting stage matching factor can comprehensively reflect the multidimensional characteristics of the lesion.

[0058] In one embodiment, the normalization process employs a maximum-minimum standardization method, mapping the temporal location information to a range of 0 to 1. The stage matching factor is multiplied by the normalized temporal location information, and the resulting product is mapped to a preset percentage scoring range to obtain a stage similarity score. This score intuitively reflects the degree of agreement between the current lesion and the standard evolution stage, providing a quantitative reference for clinical diagnosis. For example, when the product result is 0.8, mapping to a percentage scoring range yields a stage similarity score of 80, intuitively reflecting the degree of agreement between the lesion and the standard stage.

[0059] S105. Determine the disease progression description field in the report generation template based on the stage similarity score, and obtain detailed stage information by embedding the proportion of necrotic areas and the disease evolution stage markers.

[0060] The system queries a pre-defined report template library based on stage similarity scores, selects a report framework corresponding to the score range, and defines the report framework as including a description of the lesion stage, an evolution rate rating, and a prognostic trend assessment. A report framework to be filled is obtained through score matching. Using pre-defined data insertion points within the report framework, a percentage of necrotic area and standardized descriptions of lesion evolution stage markers are embedded. The numerical values ​​are connected to the text descriptions according to medical terminology combination principles, generating a mixed text paragraph containing quantitative data and qualitative descriptions. Using this mixed text paragraph, the evolution stage markers are converted into standardized staging terms based on a clinical staging reference table. The report detail level is determined by the necrotic percentage; when the necrotic percentage exceeds a pre-defined threshold, pathological descriptions are added, and the system outputs detailed staging content containing complete pathological staging information.

[0061] In one implementation, the report template library pre-stores multiple standardized disease progression description templates, each corresponding to a different stage similarity score range. When the score is between 80 and 100, the early-stage lesion template is selected; 60 to 79 corresponds to the mid-stage progression template; and below 60, the late-stage lesion template is used. Each template contains a fixed text frame and variable data insertion points. The lesion stage description illustrates the current evolutionary stage, the evolution rate rating reflects the speed of disease progression, and the prognostic trend assessment provides a possible future development direction.

[0062] Specifically, data insertion points are identified using specific markers, and the percentage of necrotic area directly replaces the corresponding numerical markers. Standardized descriptions of disease progression stages are converted from a pre-defined terminology database to ensure the standardization of reporting terminology. The principle of combining medical terminology requires that quantitative data be immediately followed by units, and that qualitative descriptions use standard medical terminology, avoiding colloquial expressions.

[0063] It should be noted that the clinical staging table establishes a mapping relationship between evolutionary stage markers and standard staging terminology. Early-stage solid predominant lesions correspond to T1 stage, intermediate-stage mixed lesions correspond to T2 or T3 stage, and late-stage cystic degeneration and necrosis correspond to T4 stage. This correlation is based on statistical analysis of a large number of pathological diagnostic results and has high accuracy.

[0064] Preferably, the level of detail in the report is dynamically adjusted based on the percentage of necrosis. When the percentage of necrosis is less than 20%, the report uses a concise format, focusing on describing the characteristics of the solid components; when the percentage of necrosis is between 20% and 60%, a standard format is used, providing a balanced description of both solid and necrotic components; when the percentage of necrosis exceeds 60%, a detailed format is used, adding descriptions of the necrosis morphology, degree of liquefaction, and cystic degeneration characteristics.

[0065] In one embodiment, the organization of the mixed text paragraphs follows a principle of moving from the whole to the part. First, the overall stage and type of the lesion are described, then specific quantitative indicators such as the percentage of necrotic area are given, and finally, morphological features are described. This hierarchical information organization method facilitates doctors' rapid grasp of key information.

[0066] For example, a typical staging detail might include: the lesion is in stage T3, moderately advanced, with necrosis accounting for 45%, liquefactive necrosis visible in the center, and obvious ring enhancement around the periphery, suggesting that the tumor has a relatively rich blood supply but the center has already shown ischemic changes.

[0067] S106. Analyze the detailed information of the stage to obtain the proportion of late-stage necrosis and cystic degeneration, and combine the enhancement intensity distribution and stage similarity score to obtain the results of identifying the true type of lesion.

[0068] The descriptions related to necrosis and cystic degeneration are extracted from the detailed stage information. The ratio of the volume of the necrotic and cystic degeneration area to the total lesion volume is calculated. If the ratio exceeds a preset late threshold, the proportion of late necrosis and cystic degeneration and late feature markers are obtained. Based on the proportion of late necrosis and cystic degeneration and late feature markers, combined with the annular enhancement width and the area of ​​the central non-enhanced region in the enhancement intensity distribution, a comprehensive feature value is calculated by multiplying the late proportion, annular feature value, and other auxiliary features by their respective preset weighting coefficients and then summing them. The comprehensive feature value is then multiplied by the stage similarity score to obtain the type determination index. The type determination index is used to query a preset pathological type interval table. The corresponding pathological type identifier is determined according to the numerical interval of the index. When a late feature marker is present, the pathological type is corrected for late stage, and the true lesion type identification result is output.

[0069] In one implementation, the detailed analysis of staging focuses on extracting late-stage lesion characteristics. A text matching algorithm identifies descriptive paragraphs containing keywords such as "necrosis," "cystic degeneration," and "liquefaction," from which relevant quantitative data is extracted. The volume of the necrotic and cystic area is obtained using 3D reconstruction technology, spatially superimposing the necrotic areas at various levels to calculate the total volume. The total lesion volume includes the sum of the solid portion and the necrotic and cystic portion. When the ratio of these two values ​​exceeds 0.6, the lesion is considered late-stage, and a late-stage characteristic marker is set.

[0070] Specifically, the intensity distribution provides detailed information about the internal structure of the lesion. The annular enhancement width is obtained by measuring the difference between the inner and outer diameters of the enhancement band, which is normally between 2 and 5 mm. The area of ​​the central non-enhanced region reflects the extent of necrosis and is calculated by counting the number of pixels in the low-signal area. These quantitative indicators provide the basis for subsequent comprehensive analysis.

[0071] It should be noted that the weighting coefficients were set based on clinical experience and statistical analysis. The weighting coefficient for the proportion of late-stage disease was set to 0.5, reflecting the important role of necrosis degree in lesion classification; the weighting coefficient for the ring feature value was 0.3, reflecting the diagnostic value of blood supply pattern; the remaining weight of 0.2 was allocated to other auxiliary features, such as density uniformity. The comprehensive feature value was obtained by weighted summation of these terms, i.e., CF = 0.5 * LO + 0.3 * RF + 0.2 * OF, where CF is the comprehensive feature value, LO is the proportion of late-stage disease, RF is the ring feature value, and OF is other auxiliary feature values. This value was then multiplied by the stage similarity score to obtain a type determination index distributed between 0 and 100.

[0072] Preferably, the pathological type interval table is divided into multiple intervals according to the index value. 0 to 30 corresponds to benign lesions, 30 to 50 corresponds to low-grade malignancy, 50 to 70 corresponds to intermediate-grade malignancy, and above 70 corresponds to high-grade malignancy. Each interval corresponds to a specific pathological type identifier, such as squamous cell carcinoma, adenocarcinoma, lymphoma, etc.

[0073] In one embodiment, the late correction mechanism adjusts the initial type determination based on late feature markers. When late feature markers are present, the system corrects the pathological type by one level towards a higher degree of malignancy, because late-stage lesions typically imply a worse prognosis.

[0074] For example, a lesion with a type determination index of 65 was initially identified as intermediate-grade adenocarcinoma. However, due to the presence of late-stage characteristic markers, the final identification result was revised to high-grade adenocarcinoma. This correction mechanism improves the accuracy of diagnosis and its clinical reference value.

[0075] S107. Based on the results of identifying the true type of lesion, a generative adversarial network is used to integrate deep information and markers of lesion evolution stages to assess the completeness of the progression description in the report and form a final diagnostic report.

[0076] An initial report framework is constructed based on the lesion type identification results. A deep learning network text generation module encodes the enhancement intensity distribution, necrosis area proportion, and boundary features from the deep information into feature vectors, which are then mapped to the text space through a fully connected layer to obtain a preliminary text sequence describing the pathology. This preliminary text sequence is compared and matched with lesion evolution stage markers. A discriminant network evaluates the consistency between each paragraph and the stage marker, calculating a relevance score for each paragraph. If the score is below a preset threshold, it is marked as a missing content area, resulting in a completeness score for the progression description. Based on the completeness score and missing content areas, description fragments corresponding to the stage are extracted from a preset standard description library and embedded into the missing positions of the preliminary text sequence. Medical terminology standardization is used to adjust the logical relationships between sentences, resulting in a structurally complete diagnostic text. Using this diagnostic text, combined with the type identifier and staging information from the lesion type identification results, a diagnostic conclusion and clinical recommendations are added at the end of the report. Formatting is used to standardize the paragraph structure, outputting a final diagnostic report containing complete pathological information.

[0077] In one implementation, the application of generative adversarial networks (GANs) in medical report generation is based on the text generation principles of deep learning. The generator network receives the true lesion type identification result as conditional input, which includes the specific type identifier and malignancy rating of the lesion. The integration of deep information unfolds from multiple dimensions: the enhanced intensity distribution numerical value reflects the blood supply characteristics inside the lesion, and parameters such as ring enhancement width and signal peak value are converted into standardized numerical vectors; the proportion of necrotic area directly reflects the degree of necrosis of the lesion, encoded as a percentage; boundary features are quantified by edge gradient values ​​and regularity scores. These heterogeneous medical features are processed through feature engineering and uniformly mapped to the same vector space.

[0078] Specifically, the mapping from feature vectors to the text space is implemented using a multi-layer fully connected neural network. The input layer receives the encoded feature vectors, typically 256-dimensional, and performs a non-linear transformation through three hidden layers, each containing 512 neurons, using rectified linear units as the activation function. The output layer generates a sequence of word embeddings, each corresponding to a word in a medical terminology dictionary. A sequence decoder then converts the word embedding sequence into a pathological description text in natural language form. The initial text sequence contains basic feature descriptions of the lesions, but may contain logical inconsistencies or missing information.

[0079] It should be noted that the discriminant network's role is to evaluate the quality and completeness of the generated text. The discriminator employs a convolutional neural network architecture to extract features from the text sequence, capturing semantic information at different granularities through multiple convolutional layers. Disease evolution stage markers serve as reference standards, encompassing typical descriptive elements expected at each stage. The comparison and matching process is achieved by calculating the semantic similarity between text paragraphs and the standard descriptions, using cosine similarity as the metric. A preset threshold was determined through cross-validation experiments, and the optimal decision point was obtained by optimization on the test dataset. When the relevance score of a paragraph falls below this threshold, it indicates that the paragraph lacks key information and needs to be supplemented. This threshold of 0.6 was determined through cross-validation experiments; on the test dataset, when the similarity is below this value, the accuracy of information loss detection reaches 95%.

[0080] Preferably, the standard description library is organized according to lesion type and evolution stage, with each category containing standardized description templates reviewed by experts. The description fragments in the library cover multiple aspects, including morphological features, signal characteristics, enhancement patterns, and internal structure. When a missing content area is detected, the system extracts a suitable description fragment from the corresponding category based on the specific type and stage of the lesion. The extracted fragment is not simply copied and pasted, but rather parameterized according to the specific characteristics of the current lesion to ensure accuracy of the description.

[0081] In one embodiment, the standardization of medical terminology involves optimization at multiple levels. At the lexical level, colloquial expressions are replaced with standard medical terms, such as standardizing "mass" to "space-occupying lesion." At the grammatical level, sentence structures are adjusted to conform to the writing standards of medical reports, using passive voice and objective descriptions. At the logical level, consistency in descriptions is ensured, avoiding contradictory or ambiguous expressions. Through this multi-layered standardization process, the professionalism and readability of diagnostic texts are guaranteed.

[0082] For example, the formation of a diagnostic conclusion comprehensively considers multiple factors. The true type identification of the lesion provides the core diagnostic basis; the type identifier clarifies the pathological nature, such as squamous cell carcinoma or adenocarcinoma. Staging information reflects the progression of the lesion, with T1 to T4 representing the evolution from early to late stages. Clinical recommendations are formulated based on the diagnostic conclusion and staging information; surgical treatment is recommended for early-stage lesions, while comprehensive treatment plans are recommended for intermediate-to-late-stage lesions. Furthermore, formatting ensures that the report meets the standard requirements of medical institutions. The paragraph structure is organized in the order of "chief complaint - present illness - imaging findings - diagnostic opinion - recommendations," with each section having a clear title and content requirements. The imaging findings section describes in detail the location, size, shape, signal characteristics, and enhancement pattern of the lesion; the diagnostic opinion section provides a clear pathological type judgment and staging assessment.

[0083] For example, a complete diagnostic report might include: the patient complains of head and neck discomfort; MRI shows an irregular mass in the parapharyngeal space, approximately 3.5 × 2.8 × 3.2 cm in size, showing isointense signal on T1WI and mixed signal on T2WI, with ring enhancement after contrast administration, and a central non-enhancing necrotic area, accounting for approximately 40%. Based on the imaging findings, a squamous cell carcinoma is diagnosed, and a pathological biopsy is recommended for further confirmation.

[0084] Understandably, through training and optimization of generative adversarial networks, the system can learn the writing patterns and standards of medical reports. The generated reports not only contain complete pathological information but also meet the requirements of clinical practice, providing valuable references for doctors' diagnostic decisions.

[0085] Based on the embodiments of the present invention described above, and through the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A weakly supervised head and neck MRI report generation method based on anomaly detection, characterized in that, include: Deep information of the lesion area is extracted from head and neck MRI by anomaly detection, and convolutional neural network is used to identify the internal texture structure of the lesion to obtain the enhancement intensity distribution and the proportion of necrotic area. Based on the enhanced intensity distribution and the proportion of necrotic area, the disease evolution stage markers and preliminary classifications are obtained by comparing with prototype samples in the preset stage evolution library. The degree of dominance of early solid enhancement in the preliminary classification was analyzed, and the matching degree between the obtained lesion types and early prototypes in the prototype library was assessed. The stage similarity score is obtained by fusing the disease evolution stage markers through a matching degree correction mechanism; Based on the stage similarity score, the disease progression description field in the report generation template is determined, and detailed stage information is obtained by embedding the proportion of necrotic area and the disease evolution stage marker. The proportion of late-stage necrosis and cystic degeneration was obtained by analyzing the detailed stage information. Combined with the enhancement intensity distribution and the stage similarity score, the true lesion type identification result was obtained. Based on the results of the identification of the true type of lesion, a generative adversarial network is used to integrate the deep information and the markers of the lesion evolution stage, evaluate the completeness of the progression description in the report, and form a final diagnostic report.

2. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, The method involves extracting deep information of the lesion area from head and neck MRI through anomaly detection, and using a convolutional neural network to identify the internal texture structure of the lesion to obtain the enhancement intensity distribution and the proportion of necrotic area, including: Abnormal signal regions are extracted from T1-weighted enhanced sequences of head and neck MRI. The texture depth changes of the abnormal signals are analyzed by a three-dimensional convolutional neural network to obtain a signal gradient distribution map. The boundary of the solid enhancement region is determined based on the signal gradient distribution map, and the solid proportion is calculated. For regions where the solid percentage is lower than a preset threshold, a region growing algorithm is used to identify the necrosis morphology of the central low-signal region, determine the liquefaction degree index, mark the completely liquefied necrosis region, and obtain the accurate value of the necrosis region percentage. Based on the precise value of the proportion of the necrotic region and the boundary of the solid enhancement region, multi-scale feature vectors are extracted, the signal intensity decreasing gradient is determined, and the enhancement intensity distribution map is obtained. By performing frequency domain transformation on the enhanced intensity distribution map, the signal contributions of the solid enhancement component and the cystic necrosis component are separated, the complexity score is calculated, and a deep information set containing the enhanced intensity distribution and the proportion of the necrotic region is output.

3. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, Based on the enhanced intensity distribution and the proportion of necrotic area, by comparing with prototype samples in a preset stage evolution library, disease evolution stage markers and preliminary classifications are obtained, including: Based on the enhanced intensity distribution and the proportion of necrotic area, a lesion feature vector is constructed. A prototype feature set is extracted from a preset stage evolution library. The matching distance value between the lesion feature vector and the prototype feature is calculated. The three prototype samples with the smallest matching distance value are selected as the candidate set. For each prototype sample in the candidate set, stage templates and time markers are extracted, the width of the annular enhancement band, the edge gradient value and the liquefaction depth index are analyzed, the dominant component type is determined, the degree of difference between the necrosis ratio and the pathological indicators in the stage template is calculated, the similarity weight value is obtained, and the time node is determined according to the time markers. By matching the time node with the staging criteria in the preset stage evolution library, and judging the stage marker of the disease evolution based on the degree of conformity between the disease feature vector and the staging criteria, the corresponding classification result is obtained from the type label of the prototype sample.

4. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 3, characterized in that, The step of obtaining disease evolution stage markers and preliminary classification based on the enhanced intensity distribution and the proportion of necrotic area by comparing prototype samples in a preset stage evolution library also includes: The width and signal peak of the annular enhancement band around the lesion are extracted by the enhanced intensity distribution. The continuity of the solid enhancement region is identified by morphological dilation and erosion operations. The difference between the edge gradient value of the solid enhancement region and the edge gradient value of the early solid dominant sample in the prototype library is calculated to determine the boundary clarity score and solid region integrity characteristics. The three-dimensional morphological parameters of the central necrotic cavity, including the major axis, minor axis, and volume ratio, are extracted from the necrotic area proportion data. The standard deviation of the signal inside the necrotic cavity is calculated to determine the liquefaction depth index. The liquefaction depth index is compared with the standard value of liquefaction depth of the late cystic degeneration dominant sample to obtain the cystic degeneration degree assessment value. Based on the boundary clarity score and the cystic degeneration degree assessment value, the component proportion coefficient is calculated to determine whether the solid component or the necrotic component is dominant, and the component dominance type is obtained. By querying a preset evolution stage reference table based on the component-dominant type, and determining the stage marker based on the combination of the solid region integrity characteristics, the liquefaction depth index, and the cystic degeneration degree assessment value, the corresponding classification index is extracted from the evolution stage reference table, and the lesion evolution stage marker and preliminary classification result are output.

5. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, The degree of dominance of early solid enhancement in the preliminary classification was analyzed, and the matching degree between the obtained lesion types and early prototypes in the prototype library was assessed, including: The feature vectors of early solid enhancement are extracted from the preliminary classification. The feature vectors include three dimensions: the area ratio of the enhanced region, the mean signal intensity, and the boundary regularity. The difference between the value of each dimension and the preset dominant threshold is calculated to obtain the basic feature set of the lesion type. By comparing the basic feature set with the prototype features of early prototypes in the prototype library item by item, the similarity score of the feature vector is calculated. The similarity score is then multiplied by a preset category weight coefficient to obtain a weighted matching value sequence. Based on the maximum value in the weighted matching value sequence, an early prototype identifier is determined. By querying the pathological type and evolution stage information corresponding to the identifier in the prototype library, the maximum weighted matching value is taken as the final matching degree between the lesion type and the early prototype.

6. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, The process of fusing the disease evolution stage markers through a matching degree correction mechanism to obtain a stage similarity score includes: Based on the matching degree value and the disease evolution stage marker, extract the temporal location information and disease progression degree value from the stage marker, calculate the absolute value of the difference between the matching degree and the preset benchmark value, and obtain the initial correction parameter by the reciprocal of the difference; The weighted average method is used to multiply the initial correction parameter by the confidence score of the evolution stage indicator, and then combine it with the preset category fusion weight to obtain the corrected stage matching factor. The stage similarity score is determined by multiplying the stage matching factor with the normalized temporal position information and mapping the product result to a preset scoring interval.

7. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, Based on the stage similarity score, the disease progression description field in the report generation template is determined. By embedding the proportion of necrotic area and the disease evolution stage markers, detailed staging information is obtained, including: Based on the stage similarity score, query the preset report template library, select the report frame corresponding to the score range, the report frame includes a description of the disease stage, a rating of the rate of evolution and a prognostic trend judgment, and obtain the report frame to be filled; By embedding the percentage value of the necrotic area and the standardized description of the disease evolution stage markers through the preset data insertion points in the report framework, the numerical values ​​are connected with the text descriptions to generate mixed text paragraphs. Using the aforementioned mixed text paragraphs, the evolutionary stage markers are converted into standardized staging terms based on the clinical staging comparison table. The level of report detail is determined by the proportion of necrotic area, and the detailed staging content containing complete pathological staging information is output.

8. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, Analysis of the detailed staging information yields the proportion of late-stage necrosis and cystic degeneration. Combined with the enhancement intensity distribution and the stage similarity score, the true lesion type is identified, including: Extract descriptions related to necrosis and cystic degeneration from the detailed staging information, calculate the ratio of the volume of the necrotic and cystic degeneration area to the total volume of the lesion, and obtain the percentage of late-stage necrosis and cystic degeneration and late-stage characteristic markers. Based on the late-stage necrosis and cystic degeneration proportion and the late-stage feature markers, combined with the annular enhancement width and the area of ​​the central non-enhanced region in the enhancement intensity distribution, the late-stage proportion and the annular feature value are multiplied by the corresponding preset weight coefficients and then added together to calculate the comprehensive feature value. The comprehensive feature value is then multiplied by the stage similarity score to obtain the type determination index. The pathological type is identified by querying a preset pathological type range table using the type determination index, determining the corresponding pathological type identifier based on the numerical range of the index, performing late-stage correction on the pathological type, and outputting the true type identification result of the lesion.

9. The method for generating weakly supervised head and neck MRI reports based on anomaly detection according to claim 1, characterized in that, Based on the results of the true lesion type identification, a generative adversarial network is used to integrate the deep information and the lesion evolution stage markers to evaluate the completeness of the progression description in the report, forming a final diagnostic report, including: Based on the identification results of the true type of lesion, an initial report framework is constructed. The enhancement intensity distribution value, necrosis area ratio and boundary features in the deep information are encoded into feature vectors and mapped to the text space through a fully connected layer to obtain a preliminary text sequence of pathological description. The preliminary text sequence is compared and matched with the disease progression stage markers, and the relevance score of each paragraph in the text sequence is evaluated by a discriminant network to obtain the completeness score of the progression description. Based on the completeness score, description fragments corresponding to the corresponding stage are extracted from a preset standard description library, and the fragments are embedded into the preliminary text sequence to obtain a structurally complete diagnostic text. Using the diagnostic text, combined with the type identifier and staging information in the lesion type identification results, a final diagnostic report containing complete pathological information is output.

Images