Thyroid tumor assessment system based on ultrasound images in combination with pathology data

Abstract

The application discloses a thyroid tumor evaluation system based on ultrasound images and pathological data, and relates to the technical field of auxiliary decision-making.The image acquisition module is used for pre-processing thyroid ultrasound images;the pathological acquisition module is used for extracting structured data containing pathological malignant scores and molecular marker expression parameters;the feature extraction module is used for generating feature vectors from the image and pathological data respectively;the region screening module is used for dynamically generating image region sensitivity threshold values based on molecular markers and pathological malignant scores;the weighted extraction module is used for identifying image suspicious regions and generating weighted features accordingly;the multi-modal fusion module is used for fusing double-modal features through a weight distribution rule;and the evaluation decision module is used for outputting benign and malignant prediction results and confidence scores.The application realizes dynamic sensitivity regulation and control of pathological semantics and image feature deep coupling, and improves the image suspicious region identification accuracy and the accuracy of cross-modal decision reliability.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of decision support, in particular to a thyroid tumor evaluation system based on combination of ultrasonic images and pathological data. BACKGROUND

[0002] Differential diagnosis of benign and malignant thyroid nodules is a key basis for clinical decision making (such as surgery or not). Preoperative evaluation mainly relies on ultrasonic imaging and fine needle aspiration cytology (FNA). However, FNA is limited by the limitations of specimen acquisition - single puncture can only obtain a small amount of cells (usually <1000 effective cells), and about 20%-30% of cases have insufficient cell quantity, poor slice quality or atypical cell heteromorphism, resulting in uncertainty in pathological diagnosis. At this time, the pathological result can only provide a conclusion of "unclear diagnosis", and cannot provide a definite basis for clinical treatment.

[0003] Traditional systems simply splice or weighted sum the ultrasonic image features and pathological parameters, ignoring the semantic correlation differences between the two types of data: image features reflect the spatial morphological performance of the lesion, and pathological parameters reveal the biological behavior at the cellular and molecular level. If a dynamic mapping relationship is not established between the two, the limitations of FNA (such as insufficient cell quantity) cannot be supplemented by image features, and high-risk signals of pathology (such as BRAF mutation) cannot be converted into early warning signals for image scanning.

[0004] In view of the above problems, there is an urgent need in the art for a thyroid tumor evaluation system based on dynamic coupling of ultrasonic images and pathological features. SUMMARY

[0005] In view of the deficiencies of the prior art, the application provides a thyroid tumor evaluation system based on combination of ultrasonic images and pathological data.

[0006] In order to achieve the above purpose, the technical scheme of the application is as follows:

[0007] The application discloses a thyroid tumor evaluation system based on combination of ultrasonic images and pathological data, comprising:

[0008] An image acquisition module is configured to acquire ultrasonic image data of the thyroid of a patient and pre-process the ultrasonic image data.

[0009] A pathological acquisition module is configured to acquire pathological structured data corresponding to the patient, wherein the pathological structured data comprises a pathological malignancy score and a molecular marker expression parameter.

[0010] A feature extraction module is configured to extract a pathological feature vector and an image feature vector from the pathological structured data and the pre-processed ultrasonic image data, respectively, and calculate a pathological malignancy score. ​

[0011] The region screening module comprises:

[0012] The parameter conversion unit is configured to convert the BRAF mutation state in the molecular marker expression parameter into a malignancy association parameter;

[0013] The threshold calculation unit is configured to convert the pathological malignancy score into an initial sensitivity coefficient by using a sigmoid function;

[0014] The dynamic adjustment unit is configured to combine the malignancy association parameter and the initial sensitivity coefficient to output an image region sensitivity threshold and dynamically adjust;

[0015] The weighted extraction module is configured to identify a suspicious region in the preprocessed ultrasound image data according to the image region sensitivity threshold and generate a weighted region feature;

[0016] The multi-modal fusion module is configured to perform nested fusion on the image feature vector and the pathological feature vector to generate a multi-modal fusion feature vector;

[0017] The evaluation decision module is configured to input the multi-modal fusion feature vector and the weighted region feature into a preset classification model to output evaluation data containing a benign-malignant prediction result and a confidence score.

[0018] Compared with the prior art, the present application has the following advantages:

[0019] 1. The present application dynamically generates and adjusts the image region sensitivity threshold based on pathological data (such as pathological malignancy score and BRAF mutation state), so that the suspicious region identification is patient-specific, and the problem of insufficient sensitivity of the traditional fixed threshold method in non-typical lesions such as fuzzy boundaries and small nodules is overcome.

[0020] 2. The present application establishes a semantic correction relationship between pathological parameters and image features at the feature level through nested fusion mechanisms such as weight adjustment matrix and Hadamard product, and generates a fusion feature that combines spatial structure information and biological significance, providing an interpretable data basis for classification decision.

[0021] 3. The present application generates a confidence score by analyzing the internal decision path of the classification model, providing a quantitative and reliable basis for the prediction result; at the same time, the generated image-text report displays the image suspicious region and the pathological evidence chain driving the region, reducing the time cost and cognitive bias of doctors in cross-modal data comparison. BRIEF DESCRIPTION OF DRAWINGS

[0022] ​The disclosure of the present application will be described with reference to the accompanying drawings. It should be appreciated that the accompanying drawings are only for the purpose of illustration and are not intended to limit the scope of protection of the present application, and in the drawings, the same reference signs are used to refer to the same parts. Among them:

[0023] Figure 1 is a system module connection diagram of the present application;

[0024] Figure 2 is a system working step flow chart of the present application;

[0025] Figure 3 is a module connection diagram of the feature extraction module of the present application;

[0026] Figure 4 is a module connection diagram of the evaluation decision module of the present application;

[0027] Figure 5 is a module connection diagram of the report generation module of the present application;

[0028] Figure 6 is a module connection diagram of the quality check module of the present application;

[0029] Figure 7 is a module connection diagram of the priority arbitration module of the present application. DETAILED DESCRIPTION

[0030] It is easy to understand that according to the technical scheme of the present application, a person skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary description of the technical scheme of the present application, and should not be regarded as the whole or as the limitation or restriction of the technical scheme of the present application.

[0031] In the prior art, the thyroid ultrasound image auxiliary evaluation system generally uses morphological feature analysis combined with fixed threshold algorithm to identify suspicious areas, and pathological data is only used as an independent diagnosis basis. The traditional method simply splices or weightedly sums the image features and pathological parameters, resulting in that the pathological high-risk indicators cannot dynamically adjust the image analysis process. For example, the abnormality of molecular markers such as BRAF mutation status cannot be converted into a warning signal for image scanning, and the dynamic characteristics related to malignant evolution in the image are reduced in weight due to the universal screening parameters, causing that the occult lesions found by the pathology department are difficult to be captured by the image system, and the local heterogeneity of benign tissue is easily misjudged as malignant targets.

[0032] To solve the above problems, analysis found that the semantic association difference between pathological data and image features is the core factor leading to the information intercommunication disorder, and the pathological malignancy score and molecular marker expression parameter in the pathological parameter have the potential to dynamically adjust the image analysis process. Further research found that converting the BRAF mutation status into a malignant association parameter, combined with the pathological malignancy score to generate a dynamic image region sensitivity threshold, can realize the real-time conduction of pathological risk signals to the image analysis process. Based on this, it is proposed to establish a multi-level interaction mechanism between pathological parameters and image features, to realize targeted feature extraction by dynamically adjusting the image region sensitivity threshold, and finally to build a bidirectional feedback system of pathology guiding image.

[0033] After introducing the basic idea of the application, the embodiments of the application will be specifically introduced below with reference to the drawings.

[0034] Embodiments:

[0035] As shown in the figure, the thyroid tumor evaluation system based on the combination of ultrasound images and pathological data includes: Figure 1 An image acquisition module is configured to acquire ultrasound image data of the thyroid of a patient and pre-process the ultrasound image data.

[0036] A pathology acquisition module is configured to acquire structured pathological data corresponding to the patient, the structured pathological data including a pathological malignancy score and a molecular marker expression parameter.

[0037] A feature extraction module is configured to extract a pathological feature vector and an image feature vector from the structured pathological data and the pre-processed ultrasound image data, respectively.

[0038] A region screening module includes:

[0039] A parameter conversion unit is configured to convert the BRAF mutation status in the molecular marker expression parameter into a malignant association parameter.

[0040] A threshold calculation unit is configured to map the pathological malignancy score into an initial sensitivity coefficient by using a pre-set monotonically increasing mapping function.

[0041] .

[0042] As a preferred implementation, the mapping function is a Sigmoid function, and the calculation formula is as follows:

[0043] ;

[0044] wherein, the pathological malignancy score, the effective input range is usually 1.0 to 10.0;

[0045] ​​To preset the balance point, the clinical significance turning point of pathological malignancy score, the pathological malignancy score in the historical training data is analyzed The typical value is 5.0.

[0046] k is the slope coefficient, which controls the sensitivity of the sigmoid function near the preset balance point The typical value range is 0.5 to 1.5, and the preferred value is 0.8.

[0047] The selection of the sigmoid function is based on the biological characteristics of thyroid tumors: the pathological malignancy score and the malignant probability show an S-shaped curve relationship. Clinical statistics show that the pathological malignancy score The goodness of fit of the malignant probability is 0.89, and the sigmoid function can effectively capture the nonlinear growth trend. The slope coefficient k=0.8 is optimized by grid search to ensure that the sensitivity changes at the turning point =5.0 meet the clinical gold standard.

[0048] The mapping function can also use other forms of increasing functions, such as piecewise linear functions, hyperbolic tangent functions, etc. Any scheme that can map the pathological malignancy score to the sensitivity coefficient should be included in the protection scope of the present application.

[0049] The dynamic adjustment unit is used to combine the malignant association parameter and the initial sensitivity coefficient, output the image area sensitivity threshold and dynamically adjust;

[0050] The dynamic adjustment unit adjusts the threshold value through a dual-mode mechanism:

[0051] Periodic trigger mode: When the pathological data is updated, the pathological data update state is detected, and when the pathological malignancy score or the BRAF mutation state changes, the threshold value is recalculated;

[0052] Event-driven mode: When the confidence score output by the evaluation decision module is less than 85%, trigger the threshold value recalculation process;

[0053] The recalculation result is output after rate limit processing to ensure the stability of the threshold value;

[0054] When the BRAF mutation is negative, the malignant association parameter adjusts the initial sensitivity coefficient , and outputs the final sensitivity coefficient as the image area sensitivity threshold, the calculation formula is as follows:

[0055] ;

[0056] wherein, is a malignant correlation parameter value, and the value range is [0.2, 1.0], wherein 0.2 corresponds to BRAF negative, and 1.0 corresponds to BRAF positive;

[0057] is a strength factor, which is obtained by learning and training through historical data, and the value range is 0.1 to 0.5;

[0058] The weighted extraction module is configured to identify suspicious regions in the preprocessed ultrasound image data according to the image region sensitivity threshold, and generate weighted region features.

[0059] The multi-modal fusion module is configured to nest fusion of the image feature vector and the pathological feature vector to generate a multi-modal fusion feature vector. The nested fusion can include but is not limited to the way of Hadamard product through a weight matrix and then concatenation, and can also use attention mechanism, cross network and other deep feature interaction ways.

[0060] The evaluation decision module is configured to input the multi-modal fusion feature vector and the weighted region feature into a preset classification model (such as an integrated learning model like XGBoost), and output evaluation data containing benign and malignant prediction results and confidence scores.

[0061] The image region sensitivity threshold refers to an image region screening standard dynamically adjusted according to pathological parameters. Specifically, an initial sensitivity coefficient can be generated by mapping the pathological malignancy score through a sigmoid function. The molecular marker expression parameter includes but is not limited to the BRAF mutation status, and is dynamically adjusted in combination with the malignant correlation parameter converted from the BRAF mutation status. The threshold can reflect the influence of individualized pathological risk of patients on image analysis. The preset image region sensitivity threshold (denoted as ) is determined by ROC curve analysis, and the effective interval is [0.8, 1.2] (normalized convolution output value). The interval corresponds to a specificity of malignant lesion identification ≥ 90% and a sensitivity ≥ 85%.

[0062] The weighted region feature refers to the suspicious region feature identified based on the dynamic threshold. Specifically, adaptive sliding window scanning combined with convolution operation can be used to realize suspicious region marking, and the feature is generated by weighted calculation of the local feature vector and the image region sensitivity threshold. The feature strengthens the image region information of the pathological risk direction. The multi-modal fusion feature vector refers to the deep interaction expression of image and pathology features. Specifically, the semantic alignment of the feature matrix can be realized through Hadamard product operation, and then the unified representation is formed through feature concatenation. The vector can capture the internal correlation of the two types of data.

[0063] In one specific embodiment, the preset classification model is an ensemble learning model based on gradient boosting decision tree, preferably implemented using the XGBoost (extreme gradient boosting) algorithm, to process the joint feature vector and output the benign or malignant prediction result. The maximum depth of the tree of the model is 6, the number of decision trees is 200, the learning setting is 0.1, the random sampling ratio of the training sample is set to 0.8; the feature sampling ratio when building each tree is set to 0.9; the objective function is set to binary classification logistic regression. The values of the hyperparameters are determined on the training set by grid search combined with five-fold cross-validation.

[0064] The preset classification model is trained based on a public thyroid data set (containing 3000 samples), and the image and pathological data are pairedly labeled (benign / malignant confirmed by postoperative pathology). Before training, the data is standardized (image normalized to 0-1, pathological parameters Z-score standardized). Five-fold cross-validation is used: the data set is randomly divided into 5 parts, 4 parts are used for training and 1 part is used for validation, repeated 5 times to take the average AUC (up to 0.92) to optimize the hyperparameters (such as tree depth = 6, learning rate = 0.1). Pre-training uses ImageNet transfer learning to initialize the feature extraction layer to accelerate convergence. During training, the number of training rounds is set to 200 rounds, and early stopping is enabled. When the performance indicator of the model on the validation set does not improve for 10 consecutive rounds, the training is automatically terminated to prevent overfitting, and the loss function is binary cross-entropy.

[0065] As shown in Figure 2 , the working principle of the present application is: after the ultrasound image data is preprocessed, the texture, morphological and other image features are extracted, and the pathological data is analyzed to obtain the pathological malignancy score and molecular marker parameters. The regional screening module converts the BRAF mutation status into a malignant association parameter, and the pathological malignancy score is mapped to the initial sensitivity coefficient by a sigmoid function. According to the final sensitivity coefficient , the image region sensitivity threshold is dynamically adjusted.

[0066] The weighted extraction module uses an adaptive sliding window to scan the image data. When the window convolution value exceeds the preset critical value, it is marked as a suspicious region, the local features are extracted, and the region features are generated by weighting with the image region sensitivity threshold. The size of the adaptive sliding window is 32x32 pixels, and the step size is 8 pixels.

[0067] The multi-modal fusion module decodes the pathological features into a weight matrix, performs Hadamard multiplication with the image features, and then concatenates the original pathological features to form a fusion vector. The evaluation decision module inputs the fusion vector and the weighted region features after splicing into the classification model, and outputs the benign or malignant prediction and confidence score.

[0068] ​To further optimize the real-time decision reliability of the dynamic adjustment unit, when the periodic trigger mode and the event-driven mode are activated at the same time (e.g., pathological data update and confidence score below 85% occur in the same time window), the system adopts the event-driven priority principle:

[0069] 1. Priority rule: event-driven mode (based on confidence score) is prioritized over periodic trigger mode (based on timing detection). Specifically, when the confidence score is below 85%, the current periodic detection process is immediately interrupted, and threshold recalculation is prioritized; after the event-driven mode processing is completed, if the periodic trigger window is still open, the remaining detection is continued.

[0070] 2. Conflict resolution mechanism: the system has a built-in state flag, which is set when the event-driven mode is activated, forcing the suspension of periodic trigger tasks; after threshold recalculation is complete and the confidence score is updated, the flag is reset.

[0071] 3. Timing control: to avoid oscillation caused by frequent switching, a minimum processing interval time (e.g., 2 seconds) is set to ensure the stability of threshold adjustment. This is achieved through a hardware-level interrupt controller with a response delay of less than 10 milliseconds.

[0072] Through the above technical solutions, the present application realizes the dynamic adjustment of pathological parameters to the image analysis process, enabling molecular marker abnormalities to trigger targeted adjustments of image scanning strategies. By establishing a dynamic mapping relationship between pathological risk and image features, the system effectively captures the characteristics of occult lesions indicated by pathology, reducing the probability of false positives in local heterogeneity of benign tissue. Deep fusion of multi-modal features enhances the system's ability to distinguish complex cases, and the confidence score provides a quantitative basis for decision reliability.

[0073] As shown in Figure 3 , the present application further proposes that the feature extraction module comprises:

[0074] An image channel unit for extracting image features after preprocessing of ultrasound image data; wherein the image features include but are not limited to texture features (such as gray level co-occurrence matrix entropy), morphological features (such as tumor boundary morphology, circularity), and blood flow distribution features (such as blood vessel density, blood flow pattern);

[0075] In addition, the system also quantitatively evaluates the following five ultrasonic malignant prompt features according to clinical consensus: substantiality, low or extremely low echo, microcalcification, blurred boundary, and aspect ratio greater than 1. Each feature is scored 1 point, and the total score ranges from 0 to 5. The scoring logic is as follows: when the total score is 1-4, it is marked as suspiciously malignant, and when the total score exceeds 4, it is directly output as a malignant judgment. The scoring result will be used as a component of the image feature vector, participating in the encoding of the subsequent image feature vector, and interacting with the pathological features in the fusion module, especially when the pathological data is not clear or the confidence is low, providing key imaging decision basis for the system.

[0076] a pathology channel unit configured to parse cytological feature parameters from the pathology structured data, wherein the cytological feature parameters include:

[0077] Ground glass nuclear score: quantifying the degree of nuclear chromatin pallor vacuolization (0-3 points);

[0078] Nuclear groove density: the number of nuclear membrane invagination grooves per unit area (count / mm²);

[0079] Intranuclear pseudoinclusion index: the proportion of cells containing eosinophilic pseudoinclusions (%);

[0080] Nuclear overlap coefficient: the degree of nuclear arrangement crowding (1-4 levels);

[0081] a feature encoding unit configured to encode the image features and the pathology parameters into image feature vectors and pathology feature vectors of the same dimension, respectively, and calculate a pathology malignancy score .

[0082] Cytological feature parameters are used to define the pathology malignancy score , which is obtained by weighted calculation based on multiple cytological feature parameters parsed from the pathology structured data, and the calculation formula is as follows:

[0083] ;

[0084] wherein, represents the feature grading score;

[0085] represents the clinical feature weight;

[0086] The cytological feature parameters are defined as shown in Table 1 (Note: the clinical feature weight is based on the verification of thyroid tumor clinical diagnosis, and the total is 1.0):

[0087] Table 1 Definition of cytological feature parameters

[0088]

[0089] The pathological channel unit outputs feature hierarchical score vector through the feature extraction module[ , , , ] and clinical feature weight[ , , , ], and calculates the pathological malignancy score ;

[0090] The regional screening module maps the pathological malignancy score to the initial sensitivity coefficient through sigmoid mapping;

[0091] The evaluation decision module generates confidence >2.5 automatically raises the pathological weight coefficient to 1.5;

[0092] The score range of the pathological malignancy score and the clinical suggestion of the control are shown in Table 2:

[0093] Table 2 Pathological malignancy score Score range and clinical suggestion table of control

[0094]

[0095] The calculation example is as follows (clinical case demonstration): cytological feature parameters: (1) ground glass nucleus: diffuse existence, 2 points; (2) nuclear groove: 8 grooves / 100 cells, 2 points; (3) intranuclear pseudoinclusion: 4 / field, 3 points; (4) nuclear overlap: moderate congestion, 2 points; calculate the pathological malignancy score according to the above cytological feature parameters: ;

[0096] Through the malignant risk grading, the malignant probability is 76%, and the postoperative pathological coincidence rate reaches 92%.

[0097] The present application simultaneously evaluates four WHO core diagnostic indicators, and quantifies them through feature hierarchical score ( ) and clinical feature weight ( ): the weight of intranuclear pseudoinclusion is the highest (W=0.30), which is consistent with its value as a specific marker; the weight of nuclear overlap is the lowest (W=0.05), which avoids misjudgment caused by benign follicle congestion; when multiple features coexist, the malignant risk increases exponentially, which is not a simple superposition:

[0098] For example: when there is only ground glass nucleus, = 0.8, malignancy rate ≈ 35%;

[0099] Nuclear groove + inclusion bodies, = 0.55 + 0.9 = 1.45, malignancy rate ≈ 65%;

[0100] Four positive, > 2.5, malignancy rate > 85%.

[0101] In terms of technical implementation, the system synchronizes the image and pathological data streams through a hardware-level parallel processing architecture: the image channel unit performs real-time feature extraction on the preprocessed ultrasound images, captures texture features using a multi-scale Gabor filter bank, quantifies edge irregularity using a morphological skeletonization algorithm, and dynamically tracks blood flow distribution patterns using an optical flow analysis method.

[0102] After receiving encrypted pathological data, the pathological channel unit uses a dedicated analysis circuit to identify cytological feature parameters: the ground glass nuclear score is calculated using a staining region segmentation algorithm, the nuclear groove density is detected using an edge curvature recognition technique (precision 0.1 μm), the pseudo-inclusion body statistics use an acidophilic color threshold segmentation, and the nuclear overlap analysis is based on Delaunay triangulation spacing measurement. Each parameter generates results in parallel through a hardware accelerator.

[0103] After receiving the dual-channel output, the feature encoding unit uses a spatial projection algorithm to map heterogeneous features to a unified dimension:

[0104] Image feature vectors (blood flow velocity, edge sharpness, etc.) are reduced in dimension through a three-layer fully connected neural network;

[0105] Pathological feature vectors (ground glass nuclear score, inclusion body index, etc.) are focused on key indicators through a self-attention mechanism; feature fusion is performed in a 256-dimensional embedding space.

[0106] Through the above technical solutions, the present application constructs a lossless transmission link from data acquisition to feature expression, eliminates the tissue scattering noise of ultrasound images and the quantization error of pathological data through a dedicated processing unit, breaks through the dimensional barrier of traditional methods through multi-modal feature fusion, aligns the ultrasound images and pathological data at the feature level, and avoids information interference caused by direct splicing. By independently extracting and encoding image texture, morphological features, and pathological proliferation parameters, the key discriminative information of each data is preserved. The unified dimension feature vector provides a standardized input for subsequent multi-modal fusion, effectively improving the ability of the benign and malignant classification model to capture cross-modal associated features.

[0107] The present application further proposes that the specific steps of converting the BRAF mutation status in the molecular marker expression parameter into a malignant association parameter include:

[0108] When the BRAF mutation is positive, the malignant prediction result is directly output, the image feature analysis process is skipped, the system outputs the malignant prediction result and the highest confidence score, and optionally skips part or all of the image feature analysis process, prioritizes the output based on pathological data, and can configure image feature verification.

[0109] When the BRAF mutation is negative, the original malignant correlation parameter calculation logic is retained, and the multi-modal fusion analysis is started.

[0110] The present application converts the clinical value of molecular diagnosis into the advantage of system architecture, and completes the key decision path switching in microseconds. The calculation resource consumption for a specific case is reduced, and the overall response efficiency of the system is improved. When the BRAF mutation is negative, the system starts the complete image and pathological multi-modal fusion analysis process to ensure the comprehensiveness of the analysis of difficult cases.

[0111] The present application further proposes that the specific steps of the weighted extraction module for identifying the suspicious area in the preprocessed ultrasound image data include:

[0112] An adaptive sliding window is used to scan the ultrasound image data after gray scale normalization processing to generate window image data;

[0113] The window image data is convolved with the image region sensitivity threshold value and the convolution output value is output;

[0114] When the convolution output value exceeds the preset sensitivity threshold value, the window image data is marked as a suspicious area;

[0115] The local feature vector of the suspicious area is extracted, and the weighted region feature is generated in combination with the image region sensitivity threshold value.

[0116] Specifically, on the ultrasound image data after gray scale normalization processing, a plurality of window image data is generated by scanning with an adaptive sliding window at a variable step length. Each window image data is convolved with the image region sensitivity threshold matrix to calculate the convolution output value representing the region malignancy probability. When the value exceeds the preset sensitivity threshold value, the window is marked as a suspicious area. The preset sensitivity threshold value is a response threshold value for triggering the suspicious area marking. The dynamic threshold parameter is obtained by training historical case data, and is used to filter low-risk areas and focus on high-probability lesions. The preset sensitivity threshold value ranges from 0.8 to 1.2. The local feature vector containing texture and morphological features is extracted for the marked area, and the feature weight is adjusted according to the image region sensitivity threshold value corresponding to the area, and finally the weighted region feature is generated, which integrates the pathological risk information.

[0117] By the technical solution, the application solves the problem of missing pathology-image correlation caused by fixed parameters in the traditional system, and realizes dynamic region screening based on individualized pathological indicators. Through the synergistic effect of the convolution response value and the image region sensitivity threshold, the image feature region strongly related to malignant evolution is accurately identified, the misjudgment caused by benign tissue heterogeneity is avoided, and the detection sensitivity of the pathology high-risk prompt region is enhanced.

[0118] The application further proposes that the specific steps of the multi-modal fusion module for nested fusion of the image feature vector and the pathology feature vector include:

[0119] According to the preset pathology semantic rule, the pathology feature vector is decoded into a weight adjustment matrix; the preset pathology semantic rule is a knowledge base solidified based on correlation analysis of historical case data. For example, when the pathology malignancy score is greater than 3, the rule base instructs the system to increase the weight of the dimensions related to 'border fuzziness' and'microcalcification' in the image feature vector. The system realizes the mapping from the pathology feature vector to the weight adjustment matrix through a lookup table or a lightweight neural network.

[0120] The weight adjustment matrix and the image feature vector are subjected to Hadamard product operation to obtain a product result;

[0121] The product result and the original pathology feature vector are concatenated in a unified dimension space to generate a multi-modal fusion feature vector.

[0122] The weighted region feature is a pixel-level local feature, including: the microcalcification point density (the proportion of pixels with a gray jump of ≥50HU in the calculation window) extracted through the gray level co-occurrence matrix;

[0123] The edge blur index based on the DoG filter detection ( , wherein is a dimensionless value between 0 and 1; is the number of blurred pixels, which refers to the number of pixels identified by the DoG filter or other edge detection algorithms located on the tumor edge fuzzy protrusions; is the total number of region pixels, which refers to the total number of pixels in the analyzed tumor region);

[0124] The multi-modal fusion feature is a cross-modal semantic feature, which includes the image morphological parameter mapping result guided by the pathology semantics, including:

[0125] The tumor overall morphological parameter in the image feature vector (such as the circularity , wherein A is the area and P is the perimeter);

[0126] When there are local artifacts in the ultrasound image (such as bubble interference), the weighted regional features may fail, and at this time the multimodal fusion features provide reliability guarantee through the association of pathological data. When the pathological data is partially missing, the weighted regional features can still maintain the basic judgment ability based on the local features of the image.

[0127] Specifically, the system maps the key indicators (such as pathological malignancy score, intranuclear pseudoinclusion index) in the pathological feature vector to a weight adjustment matrix of the same dimension as the image feature vector according to the prior knowledge summarized from clinical big data (for example, when the pathological malignancy score is >2.0, it is significantly related to the edge blur feature in the ultrasound image). Each element value of the matrix represents an enhancement or inhibition coefficient of the corresponding image feature dimension, and its generation logic is derived from the rules solidified after statistical analysis of historical case data, for example, setting the weight of the pre-defined edge feature dimension in the image feature vector to 1.5, thereby converting the abstract pathological risk into specific image feature attention regulation signals.

[0128] The system performs semantic alignment at the feature level and performs Hadamard product operation on the generated weight adjustment matrix and the image feature vector. This operation is equivalent to an element-level weighting operation, so that the image features that are judged to be high-risk at the pathological level (such as texture at the unclear boundary) are explicitly enhanced in value, while features with weak association with pathological hints maintain the original weight. This ensures that the pathological semantics can directly and interpretable modulate the imaging expression, providing image features refined by pathological knowledge for subsequent classification.

[0129] In order to enhance the image features while preserving the discriminative information of the pathological data itself, the system performs feature concatenation on the result of the Hadamard product operation and the original pathological feature vector in a unified dimension space. The multimodal fusion feature vector generated thereby not only contains key image features enhanced by pathological semantics, but also retains the original pathological score and molecular marker information.

[0130] Through the above technical solutions, the application effectively solves the problem of missing semantic association in traditional multimodal fusion methods, allowing pathological parameters to dynamically adjust the weight distribution of image features, while avoiding information dilution in the concatenation process. The nested fusion mechanism improves the representation ability of multimodal features, allowing the subsequent classification model to more accurately capture key image features of pathological hints, thereby improving the reliability of benign and malignant prediction results.

[0131] As shown in Figure 4 Further, the application further proposes that the evaluation decision module comprises:

[0132] a feature jointing unit configured to concatenate the multimodal fusion feature vector and the weighted regional feature along the feature dimension to generate a joint feature vector;

[0133] a decision reasoning unit configured to input the joint feature vector into a preset classification model for processing, and output a benign or malignant prediction result;

[0134] a confidence score generation unit configured to calculate a feature contribution value distribution and generate a confidence score based on feature node information in a decision path of the preset classification model; the confidence score represents a reliability quantitative value of the benign or malignant prediction result of the preset classification model.

[0135] The feature joint unit refers to a component that splices feature vectors of different modalities in a dimension, and can be implemented by tensor splicing or vector connection operation. The joint input is formed by integrating multi-modal data and weighted regional features. The decision reasoning unit refers to an algorithm module that performs classification prediction, and can be implemented by a deep neural network or a support vector machine model. The joint features are non-linearly mapped and the classification result is output. The confidence score generation unit refers to a calculation module for evaluating the reliability of the prediction result. The contribution of each feature can be quantified by analyzing the influence weight distribution of the internal feature nodes of the model, combining the information gain and the path depth, and then mapped to an interpretable score index.

[0136] Specifically, after receiving the multi-modal fusion feature vector and the weighted regional feature, the feature joint unit splices them along the feature dimension to form a joint feature vector, so that the classification model can utilize both global multi-modal information and local suspicious regional features. The decision reasoning unit inputs the joint feature vector into the preset classification model, extracts high-order feature expressions through multiple layers of non-linear transformation, and finally outputs the benign or malignant prediction result. The confidence score generation unit extracts feature node information in the decision path during model reasoning, including split point threshold, information gain and path depth. According to a preset feature contribution function, the original contribution is calculated, and dynamic weight coefficients are applied to different feature types. Finally, the confidence score is generated after normalization. For example, the contribution of pathological feature nodes can be given a higher weight coefficient to reflect the key role of pathological data in malignant prediction.

[0137] The present scheme quantifies the contribution difference of different data sources to the prediction result by analyzing the information gain and path depth of the internal feature nodes of the classification model, so that the confidence score can accurately reflect the reliability basis of the model decision. Especially when there is a conflict between high-risk pathological parameters and image features, the quantifiable confidence basis is provided. Through the above technical scheme, the present application solves the problem that the traditional evaluation system cannot dynamically quantify the reliability of the prediction result. By calculating the feature contribution value distribution and type weighting processing, the confidence score can distinguish the decision weight difference between pathological and image features, assist clinicians in making comprehensive judgments based on quantitative scores when there is a conflict between pathological data and image features, and reduce the risk of misdiagnosis.

[0138] The application further proposes that the confidence score generation unit calculates the feature contribution value distribution and generates the confidence score based on the feature node information of the preset classification model in the decision path. The specific steps of calculating the feature contribution value distribution and generating the confidence score include:

[0139] The feature node information of the preset classification model in the decision path is extracted, and the feature node information includes the split point threshold, information gain value and path depth of each feature node.

[0140] According to the information gain value and path depth of each feature node information, the original contribution degree is calculated through a preset feature contribution function.

[0141] The original contribution degree is subjected to feature type weighting processing, the node contribution degree belonging to the pathological feature vector is fused with a preset pathological weight coefficient to obtain a pathological node contribution degree, and the node contribution degree belonging to the image feature vector is fused with a preset image weight coefficient to obtain an image node contribution degree. The pathological weight coefficient and the image weight coefficient are obtained by training based on 3000 postoperative pathological-image pairing historical data.

[0142] The pathological node contribution degree and the image node contribution degree are accumulated to calculate the feature contribution value distribution, which is mapped to a preset score interval through a normalization function to generate the confidence score.

[0143] The feature node information refers to the node attribute of the classification model used for splitting data in the decision tree or neural network, and can be specifically implemented by using parameters such as the split point threshold, information gain value and node level depth of the decision tree, for quantifying the contribution degree of different features to the classification result. The original contribution degree refers to the initial value of feature importance calculated based on the node information gain and path depth, and can be specifically implemented by using weighted summation or an exponential decay function. The node with shallower path depth has greater influence on the decision. The feature type weighting processing refers to assigning different weight coefficients to the pathological features and image features, and can be specifically implemented by dynamically adjusting the contribution proportion of the two types of features to the classification accuracy in the historical training data, so that the high-risk signal of the pathological parameter obtains higher weight in the confidence calculation.

[0144] The confidence score generation mechanism proposed in the application constructs a dynamic evaluation system for quantifying the reliability of the prediction result by deeply analyzing the contribution distribution of the feature nodes in the decision path of the classification model. The core of the mechanism is to establish a three-level processing flow of feature contribution value calculation, type weighting and normalization mapping, so that the confidence score can accurately reflect the actual influence weight of the pathological features and image features in the classification decision.

[0145] At the technical implementation level, the confidence generation unit first extracts the feature node information of the preset classification model in the decision path, including key parameters such as node split point threshold, information gain value and path depth. Based on these parameters, the original contribution degree is calculated through the preset feature contribution function, and the original contribution degree calculation formula is as follows:

[0146] ;

[0147] Where IG is the information gain value, and depth represents the hierarchical depth of the node in the decision tree. The function strengthens the contribution weight of the shallow layer node through the logarithmic decay mechanism, which conforms to the objective law that the high layer node in the decision tree has a greater impact on the classification result.

[0148] Subsequently, the system performs feature type weighting processing on the original contribution degree:

[0149] The contribution degree of the pathological feature node is multiplied by the pathological weight coefficient : When the BRAF mutation positive signal is input, the fixed output confidence score is ≥99%, and the contribution degree of the pathological feature node is automatically promoted to the highest level, and the pathological weight coefficient =1.5;

[0150] The contribution degree of the image feature node is multiplied by the image weight coefficient : ;

[0151] Where and are learned through historical training data, , , satisfying , to ensure that the total sum of feature contribution degrees is not distorted due to weight scaling; typical values , , reflect the higher weight of pathological parameters in malignant discrimination. This differentiated weighting mechanism converts the clinical experience of thyroid tumor diagnosis (pathological indicators are the gold standard) into quantitative parameters in the mathematical model.

[0152] The weighted pathological node contribution degree and the image node contribution degree are accumulated to form the feature contribution value distribution:

[0153] ;

[0154] The accumulated feature contribution value distribution Dist is mapped to the 0-100 score interval through a normalization function to generate the confidence score .

[0155] Exemplarily, the Sigmoid function can be used for normalization function mapping to the 0-100 score interval to generate the final confidence score:

[0156] ;

[0157] wherein is the mean of the distribution, is the standard deviation, the score is ensured to be discriminative for different cases by dynamically adjusting the slope of the mapping curve.

[0158] Specifically, after the classification model completes the benign and malignant prediction, the system traverses all feature nodes participating in the judgment in the decision path, extracts the split threshold, information gain and hierarchical depth information. Through the feature contribution function, the information gain value and the path depth are combined to calculate the original contribution degree of each node. Subsequently, the contribution degree is differentially weighted according to the feature type: the pathological feature node is multiplied by the pathological weight coefficient obtained by training the historical data, and the image feature node is multiplied by the image weight coefficient. The weighted contribution degrees of the two types are accumulated to form the feature contribution value distribution, which is then normalized to convert it into a confidence score in the interval of 0-1.

[0159] When the confidence score is lower than 85%, the threshold recalculation of the region screening module is automatically triggered; the low-confidence case ( ) activates the “suggested pathological review” warning label in the report to guide the clinical attention; the feature type with continuous low contribution degree (such as blood flow distribution feature) triggers the parameter update signal of the feature extraction module, realizing the self-optimization of the system. The score reflects the certainty degree of the model for the current prediction result, and the higher the score, the stronger the consistency between the pathology and the image features.

[0160] The present application decouples the analysis of the decision path of the pathological parameters and the image features in the model through dynamic weighting and normalization processing, so that the pathological indicators with high malignant risk can directly affect the confidence score, avoiding the dilution of feature weights caused by generalization calculation. Through the above technical solutions, the present application can accurately quantify the actual role of pathological and image features in classification decision, making the confidence score more consistent with patient-specific data. Doctors can judge the reliability of the prediction result according to the score, for example, when the score is lower than the threshold, the image can be reviewed in combination with the pathological markers in the fuzzy boundary area, thereby reducing the diagnostic bias caused by single feature misjudgment.

[0161] As Figure 5 shown, the present application further proposes that the thyroid tumor evaluation system based on the combination of ultrasound images and pathological data further comprises a report generation module, and the report generation module comprises:

[0162] A region labeling unit is configured to draw a suspicious region marker frame in the ultrasound image to display a highlighted marker of the image suspicious region, and the color depth of the suspicious region marker frame is positively correlated with the malignancy correlation parameter;

[0163] The data matching unit is configured to extract boundary fuzziness in the image features, establish a control relationship between the boundary fuzziness and the pathological malignancy score based on Spearman correlation analysis, and generate a pathological data matching basis.

[0164] The suggestion generation unit is configured to search a preset period dictionary according to the confidence score, and generate a customized clinical review suggestion.

[0165] The color depth and the malignancy correlation parameter are positively correlated, that is, the visual color scale intensity of the marking box is enhanced as the value of the malignancy correlation parameter increases, and the linear mapping of the saturation component in the HSV color space and the malignancy correlation parameter can be used to achieve this, which is used to intuitively reflect the influence of the pathological parameter on the malignancy degree of the suspicious region of the image. The control relationship between the boundary fuzziness and the pathological malignancy score is that the correlation between the image features and the pathological score is quantified by statistical methods, and the Spearman rank correlation coefficient can be used to calculate the non-linear correlation degree between the two, which is used to establish an objective correspondence between the image features and the pathological indicators. The preset period dictionary is a mapping table that stores the mapping between the confidence score interval and the review time suggestion, and a key-value database can be used to store the review period parameters corresponding to different score thresholds, which is used to realize the automatic generation of clinical suggestions.

[0166] Specifically, the region labeling unit superimposes a semi-transparent color block on the ultrasound image interface to mark the suspicious region, and the color depth of the color block is dynamically adjusted according to the malignancy correlation parameter converted by the BRAF mutation state. When the parameter value is 1.0, a deep red mark is displayed, and when the parameter value is 0.2, a light orange mark is displayed. The data matching unit performs non-parametric statistical testing on the boundary fuzziness feature and the pathological malignancy score, calculates the rank correlation coefficient between the two, and generates a significance level report. When the absolute value of the correlation coefficient is greater than 0.5, a strong correlation prompt is output. The suggestion generation unit matches the review period in the dictionary according to the confidence score, for example, when the score is higher than 90 points, a 3-month review is recommended, when the score is 70-90 points, a 6-month review is recommended, and when the score is lower than 70 points, a 12-month review is recommended.

[0167] Through the above technical solutions, the present application realizes real-time dynamic correlation between pathological parameters and image markers, improves the visual recognition efficiency of the malignancy degree of the suspicious region, provides objective control basis for the image features and the pathological indicators for doctors, and optimizes the accuracy and consistency of the clinical review scheme through the automatic suggestion generation mechanism driven by the confidence score.

[0168] As shown in Figure 6 The present application further proposes that the thyroid tumor evaluation system based on ultrasound images and pathological data further includes a quality verification module, and the quality verification module is configured to:

[0169] During the image acquisition stage of the image acquisition module, the gray scale distribution uniformity of the acquired ultrasound image data is detected.

[0170] When the non-uniformity coefficient of the gray scale distribution uniformity exceeds the preset uniformity threshold, a gray scale recalibration instruction is triggered to correct the ultrasound image data;

[0171] A quality marker bit is added to the corrected ultrasound image data, and the ultrasound image data with the quality marker bit is input to a feature extraction module. When the quality marker bit is activated, the gray level co-occurrence matrix texture feature extraction is disabled.

[0172] The non-uniformity coefficient, which is a core indicator of the gray scale distribution uniformity, is calculated The calculation formula is as follows:

[0173] ;

[0174] Wherein, is the gray scale variance of the image segmentation region (default 8x8 pixel block), is the average gray scale value of the region, is the maximum and minimum gray scale region area ratio. The non-uniformity coefficient comprehensively reflects the gray scale dispersion degree and the abnormal region proportion, and when >0.35, it is determined that the quality is abnormal. The preset uniformity threshold is 0.35. Based on the statistical verification of clinical historical data (500 cases of thyroid ultrasound images), through ROC curve analysis, it is determined that when >0.35, the misjudgment risk of image quality to feature extraction is significantly increased (sensitivity ≥85%, specificity ≥90%).

[0175] At this time, the system automatically triggers a gray scale recalibration instruction and uses a contrast limited adaptive histogram equalization (CLAHE) algorithm for processing: the image is divided into 32x32 sub-regions, and the gray scale histogram of each region is calculated; the histogram equalization amplitude is limited ( ) to prevent local over-enhancement;

[0176] The region boundary artifacts are eliminated by bilinear interpolation. This algorithm is particularly suitable for the characteristics of coexistence of high echo tissue and low echo lesions in thyroid ultrasound images, and can improve the local contrast while maintaining the overall gray scale continuity.

[0177] When the system returns to the acquisition buffer after correcting the image data (under the condition of device support), the quality verification module verifies the value of the non-uniformity coefficient again, forming a real-time feedback loop. When the secondary verification does not meet the standard, the system automatically adjusts the probe gain parameter and reacquires the current frame. The quality marker bit guides the feature extraction module to switch the processing strategy (such as disabling the gray scale variance dependent algorithm), and automatically ignores suspicious area. When the confidence score abnormally drops, the system prioritizes the quality flag history record to quickly locate the problem source to the image acquisition link.

[0178] The present application changes the quality control from the traditional manual post-inspection to intelligent feedforward control, through the non-uniformity coefficient Constructs objective evaluation standard, CLAHE algorithm realizes targeted enhancement, quality flag guides cross-module cooperative defense, forms a quality protection system throughout the whole chain of acquisition, processing and decision-making. It can identify and correct the problem of uneven gray scale in the image acquisition stage, and guide the subsequent modules to adopt adaptive processing strategies (such as disabling texture feature extraction affected by uneven gray scale), thereby reducing the risk of misjudgment caused by image quality defects and enhancing the stability of the system in complex clinical environment.

[0179] Among them, the gray scale distribution uniformity refers to the consistency index of the gray scale value distribution of different regions in the ultrasound image, which can be realized by calculating the image region gray scale variance or histogram similarity, and is used to judge whether there is uneven exposure or device noise interference in the image acquisition process. The non-uniformity coefficient refers to the numerical value of quantifying the deviation of gray scale distribution from the uniform state, which can be realized by using the ratio of standard deviation to average gray scale or the percentage of maximum gray scale difference. When the coefficient exceeds the preset uniformity threshold, it indicates that the image quality may affect the accuracy of subsequent feature extraction. The gray scale re-correction instruction refers to the control signal that triggers the image enhancement algorithm, which can be realized by using histogram equalization or adaptive gamma correction algorithm, and is used to eliminate the image quality defects caused by abnormal gray scale distribution. The quality flag refers to the identifier attached to the corrected image data, which can be realized by using binary flag or metadata tag, and is used to trace the data correction record in the subsequent processing process.

[0180] Specifically, the image gray scale distribution state is monitored in real time during the ultrasound image acquisition stage, and whether there is device acquisition abnormality or environmental interference is judged by calculating the non-uniformity coefficient. When it is detected that the non-uniformity coefficient exceeds the preset uniformity threshold, the gray scale correction algorithm is automatically triggered to enhance the original image, eliminating the gray scale abnormality caused by device parameter deviation or tissue reflection difference.

[0181] After the correction is completed, the quality flag is embedded in the image data, so that the subsequent feature extraction module can identify that the image has passed the quality check, avoiding the distortion of texture features or the calculation error of morphological parameters caused by uneven gray scale distribution. For example, when the local area gray scale value drops sharply due to unstable contact pressure of the ultrasound probe, the system detects the gray scale distribution uniformity in real time, and starts the histogram matching algorithm for gray scale compensation in time, ensuring the reliability of the correlation analysis of the texture features extracted subsequently and the pathological data.

[0182] Through the technical solution, the application can effectively identify and correct the gray scale distribution abnormality problem in the ultrasonic image acquisition process, avoid the influence of image quality defects caused by equipment or operation factors on subsequent pathological correlation analysis. Through the front quality checking mechanism, the risk of false extraction of texture features caused by uneven gray scale is reduced, and the credibility of the tumor evaluation result is improved. At the same time, the introduction of the quality mark bit enables the system to distinguish between original data and corrected data, and provides clear data processing basis for the feature extraction module.

[0183] As shown in Figure 7 The application further proposes that the thyroid tumor evaluation system based on the combination of ultrasonic images and pathological data further comprises a priority arbitration module, the priority arbitration module is used for preferentially processing pathological structured data when the pathological structured data and the ultrasonic image data are synchronized; and the pathological structured data is encrypted and transmitted after being authorized by a patient.

[0184] Preferential processing of the pathological structured data refers to preferentially analyzing and extracting features of the pathological data in a data synchronization transmission scenario, so as to ensure that the pathological parameters can participate in subsequent image screening and fusion calculation in real time. Specifically, a queue scheduling algorithm or a priority marking mechanism can be used to achieve this, for example, a priority queue is set in a data receiving interface, the pathological data is marked as a high-priority task and inserted into the head of the queue. The encryption transmission refers to the security protection of the pathological data containing patient privacy, to prevent unauthorized access or data leakage. Specifically, an asymmetric encryption algorithm or a privacy protection protocol based on a block chain can be used to achieve this, for example, the pathological data is encrypted and packaged through a digital certificate authorized by the patient before data transmission.

[0185] Specifically, when the ultrasonic image data and the pathological data are transmitted to the system at the same time, the priority arbitration module identifies the pathological structured data by detecting the identifier of the data type and assigns a higher processing priority. The pathological data is sent to the feature extraction module for analysis, while the image data is temporarily stored in the cache queue for processing. After the pathological data processing is completed, the system will release the computing resources to continue processing the ultrasonic image data. At the same time, the system encrypts the data by calling the key pre-authorized by the patient during the transmission of the pathological data, to ensure the security of the transmission channel.

[0186] Through the technical solution, the application solves the contradiction in processing order when multi-modal data is synchronized and transmitted, ensures that the pathological parameters can participate in the image analysis process preferentially, and thus improves the response speed of the system to high-risk pathological indicators; at the same time, the encryption transmission mechanism ensures the security of the patient's private data, to avoid illegal interception or tampering of sensitive information during the transmission process.

[0187] The following is a specific embodiment of the thyroid tumor evaluation system based on the combination of ultrasonic images and pathological data:

[0188] A 48-year-old female patient was diagnosed with a 15x13mm nodule in the left lobe of the thyroid. The Philips EPIQ 7 ultrasound system (14MHz probe) was used to collect images and perform fine needle aspiration biopsy to obtain pathological specimens. An ultrasound and pathological data-based thyroid tumor evaluation system was developed.

[0189] The system receives ultrasound images with a resolution of 1280x960 through a PACS interface. After adjusting the window width (WL=45, WW=200), the quality verification module detects the non-uniformity coefficient (NUC) of the grayscale, which is 0.31 (lower than the threshold of 0.35), and directly marks it as qualified. Pathological data is encrypted and transmitted to the feature extraction module using the SM4 algorithm. The cytological features (hyaline nuclei 1 point, nuclear groove density 2 points, nuclear pseudoinclusion index 1 point, nuclear overlap coefficient 2 points) and molecular markers (BRAF mutation negative) are analyzed, and the pathological malignancy score is calculated. At the same time, the image channel unit extracts image morphological features, including boundary fuzziness 0.72 (0-1 scale) and microcalcification density 35%.

[0190] Since the patient has a negative BRAF mutation, , the regional screening module starts double-path processing. The initial threshold is calculated as =1.30 after sigmoid mapping (parameters k=0.8, =5.0).

[0191]

[0192] The dynamic adjustment unit combines to output the final threshold .

[0193] The weighted extraction module scans the image with an image region sensitivity threshold of 0.30, identifying 2 suspicious regions (maximum convolution response value 1.05), and further extracting weighted region features (microcalcification density 42%, edge blur index 0.29). The multi-modal fusion module uses a nested fusion mechanism: the pathological feature vector is decoded into a weight matrix, multiplied with the image features using the Hadamard product, and then concatenated with the original pathological features to generate a 256-dimensional fusion vector for final evaluation.

[0194] The evaluation decision module outputs the result: malignant probability 58.3% (Bethesda IV). The confidence score generation process calculates the contribution of the pathological and image nodes (pathological node depth 3, information gain IG=0.32, weight , contribution 0.38; image node depth 2, IG=0.41, weight ​, contribution 0.33), and after Sigmoid mapping, we get (As the confidence is less than 85%, the dynamic adjustment unit is triggered to recalculate the threshold).

[0195] The system finally generates a visual report, and the ultrasound image is labeled with an orange marker box ( ), and the data matching report shows that the boundary blur degree 0.72 is significantly correlated with =1.30, combined with the medium risk confidence (84.2%), the output clinical suggestion is "6-month review".

[0196] In this example, the system accurately identifies the follicular tumor that needs to be followed up and observed through the medium risk classification of =1.30 combined with the image feature weighted analysis, avoiding unnecessary surgical treatment; in the context of the contradiction between cytological features and image performance (high boundary blur degree but lower

[0197] The technical scope of the present application is not limited to the content in the above description, and those skilled in the art can make various modifications and changes to the above examples without departing from the technical idea of the present application, and these modifications and changes should all belong to the protection scope of the present application.

Claims

1. A thyroid tumor assessment system based on ultrasound images in combination with pathology data, characterized by: The method comprises the following steps: An image acquisition module is used to acquire ultrasound image data of a patient's thyroid and pre-process the ultrasound image data; A pathology acquisition module is used to acquire structured pathology data corresponding to the patient, which includes a pathology malignancy score and a molecular marker expression parameter; a feature extraction module configured to extract a pathology feature vector and an image feature vector from the pathology structured data and the preprocessed ultrasound image data, respectively, and to calculate a pathology malignancy score ; A region screening module comprises: A parameter conversion unit is used to convert the BRAF mutation status in the molecular marker expression parameter into a malignancy association parameter; a threshold calculation unit configured to map the pathological malignancy score to an initial sensitivity coefficient by a sigmoid function; A dynamic adjustment unit is used to combine the malignancy association parameter with the initial sensitivity coefficient to output an image region sensitivity threshold and dynamically adjust it; A weighted extraction module is used to identify suspicious regions in the pre-processed ultrasound image data according to the image region sensitivity threshold and generate weighted region features; A multi-modal fusion module is used to nest fusion of the image feature vector and the pathology feature vector to generate a multi-modal fusion feature vector; An evaluation decision module is used to input the multi-modal fusion feature vector and the weighted region feature into a pre-set classification model to output evaluation data including benign and malignant prediction results and confidence scores. 2.The thyroid tumor evaluation system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The feature extraction module comprises: An image channel unit is used to extract image features after pre-processing the ultrasound image data; wherein the image features include but are not limited to texture features, morphological features, and blood flow distribution features; A pathology channel unit is used to analyze cytological feature parameters from the structured pathology data; wherein the cytological feature parameters include ground glass nuclear score, nuclear groove density, nuclear pseudoinclusion index, and nuclear overlap coefficient; a feature encoding unit, configured to encode the image features and the pathological parameters into the image feature vectors and the pathological feature vectors respectively with the same dimension, and calculate a pathological malignancy score . 3.The thyroid tumor evaluation system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The specific steps of converting the BRAF mutation status in the molecular marker expression parameter into a malignancy association parameter include: When the BRAF mutation is positive, directly output the malignant prediction result, and skip the image feature analysis process; When the BRAF mutation is negative, retain the original malignancy association parameter calculation logic and start multi-modal fusion analysis. 4.The thyroid tumor evaluation system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The specific steps of the weighted extraction module to identify suspicious regions in the processed ultrasound image data include: An adaptive sliding window is used to scan the pre-processed ultrasound image data to generate window image data; The window image data is convolved with the image region sensitivity threshold and the convolution output value is output; When the convolution output value exceeds the preset sensitivity threshold value, the window image data is marked as the suspicious region; The local feature vector of the suspicious region is extracted, and the weighted region feature is generated in combination with the image region sensitivity threshold. 5.The thyroid tumor assessment system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The specific steps of nesting fusion of the image feature vector and the pathology feature vector include: According to the pre-set pathology semantic rule, the pathology feature vector is decoded into a weight adjustment matrix; The weight adjustment matrix is subjected to Hadamard product operation with the image feature vector to obtain a product result; The product result and the original pathology feature vector are concatenated in a unified dimension space to generate the multi-modal fusion feature vector. 6.The thyroid tumor assessment system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The evaluation decision module comprises: A feature joint unit is used to concatenate the multi-modal fusion feature vector and the weighted region feature along the feature dimension to generate a joint feature vector; A decision reasoning unit is configured to input the joint feature vector into a preset classification model for processing, and output the benign and malignant prediction result. A confidence score generation unit is configured to calculate a feature contribution value distribution and generate a confidence score based on feature node information of the decision path in the preset classification model, wherein the confidence score represents a reliability quantitative value of the preset classification model for the benign and malignant prediction result. 7.The thyroid tumor assessment system based on the combination of ultrasound images and pathological data according to claim 6, characterized in that: The specific steps of calculating the feature contribution value distribution and generating the confidence score include: extracting feature node information of the decision path in the preset classification model, wherein the feature node information includes split point thresholds, information gain values and path depths of each feature node; calculating original contribution degrees by using a preset feature contribution function according to the information gain values and path depths of each feature node information; performing feature type weighting processing on the original contribution degrees, fusing a node contribution degree belonging to the pathological feature vector with a preset pathological weight coefficient to obtain a pathological node contribution degree, and fusing a node contribution degree belonging to the image feature vector with a preset image weight coefficient to obtain an image node contribution degree, wherein the pathological weight coefficient and the image weight coefficient are obtained by training historical data; calculating the feature contribution value distribution by accumulating the pathological node contribution degree and the image node contribution degree, mapping the feature contribution value distribution to a preset score interval by using a normalization function, and generating the confidence score. 8.The thyroid tumor assessment system based on the combination of ultrasound images and pathological data according to claim 4, characterized in that: The thyroid tumor evaluation system based on the combination of ultrasound images and pathological data further includes a report generation module, wherein the report generation module includes: a region labeling unit configured to draw a suspicious region marking box in the ultrasound image to display a highlighted marking of the image suspicious region, and the color depth of the suspicious region marking box is positively correlated with the malignancy correlation parameter; a data matching unit configured to extract boundary fuzziness in the image feature, establish a control relationship between the boundary fuzziness and the pathological malignancy score based on Spearman correlation analysis, and generate a pathological data matching basis; a suggestion generation unit configured to retrieve a preset periodic dictionary according to the confidence score, and customize a clinical review suggestion. 9.The thyroid tumor assessment system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The thyroid tumor evaluation system based on the combination of ultrasound images and pathological data further includes a quality check module, wherein the quality check module is configured to: detect the gray scale distribution uniformity of the collected ultrasound image data when the image acquisition module is in the image acquisition stage; trigger a gray scale recalibration instruction to correct the ultrasound image data when it is detected that the non-uniformity coefficient of the gray scale distribution uniformity exceeds a preset uniformity threshold; add a quality marker bit to the corrected ultrasound image data, and input the ultrasound image data with the quality marker bit into the feature extraction module. 10.The thyroid tumor assessment system based on the combination of ultrasound images and pathological data according to claim 1, characterized in that: The thyroid tumor evaluation system based on the combination of ultrasound images and pathological data further includes a priority arbitration module, wherein the priority arbitration module is configured to prioritize the processing of the pathological structured data when the pathological structured data and the ultrasound image data are synchronized, and the pathological structured data is encrypted and transmitted after being authorized by a patient.

Images