A method and system for ultrasound-assisted diagnosis of breast and thyroid glands that integrates expert systems and deep learning
By combining multimodal data preprocessing with expert system logic verification, the generalization ability and interpretability issues of deep learning models in breast and thyroid ultrasound diagnosis were resolved, achieving efficient and low-cost diagnostic support and improving diagnostic accuracy and physician trust.
Patent Information
- Application Number
- CN202611382837.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-08
- Publication Date
- 2026-11-10
AI Technical Summary
Existing deep learning models have weak generalization ability, poor interpretability, high hardware requirements, high misjudgment rate and serious latency in the diagnosis of breast and thyroid ultrasound, which cannot meet the actual clinical needs.
By employing multimodal data preprocessing, deep learning feature perception, expert system logic verification based on medical knowledge graphs, and dynamic weighted fusion, combined with lightweight forward inference and Bayesian inference, an interpretable structured report is generated.
It significantly improves the accuracy and interpretability of cross-device diagnosis, reduces hardware costs and latency, greatly reduces the false positive rate, and increases doctors' acceptance of AI conclusions.
Smart Images

Figure CN123072684A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent medical image analysis and computer-aided diagnosis technology, specifically to a method, system, and electronic device for ultrasound-assisted diagnosis of breast and thyroid glands that combines a deep learning perception model with logical reasoning from a medical knowledge graph. The innovation of this patent lies not in simply superimposing the perception and reasoning modules, but in exploring a deeper synergistic mechanism. This "coupling" approach is expected to fundamentally improve the intelligence level and clinical applicability of the auxiliary diagnostic system. Background Technology
[0002] Ultrasound examination is the preferred imaging method for screening and diagnosing breast and thyroid nodules. Currently, computer-aided diagnostic systems are mainly divided into two categories: traditional CAD systems based on manually designed features, and deep learning models based on convolutional neural networks. Although deep learning models perform excellently in feature extraction, they still have significant limitations in practical clinical application.
[0003] Data and algorithm bottlenecks: Deep learning models have weak generalization ability and their performance degrades significantly on images across devices and operators; the model decision-making process is a "black box" with poor interpretability and low doctor trust; and the misjudgment rate is high for rare cases (such as postoperative scars and concentrated colloids).
[0004] Computational resource constraints: Complex models suffer from high inference latency and large memory consumption, rely on expensive GPU hardware, and are difficult to popularize in primary healthcare institutions.
[0005] Disconnected clinical workflow: False positive markers generated by the system interfere with diagnosis, structured report generation is inefficient and lacks logical connection with medical history.
[0006] In summary, existing technologies cannot simultaneously meet the practical clinical needs of "high precision and high interpretability", "low latency and low cost", and "strong generalization and strong logic". Summary of the Invention
[0007] The present invention aims to provide an ultrasound-assisted diagnostic method and system that integrates expert systems and deep learning, in order to solve the technical problems of poor interpretability of deep learning models, weak cross-device generalization ability, high misjudgment rate of rare cases, and demanding hardware computing resource requirements in the prior art.
[0008] (II) Technical Solution
[0009] To achieve the above objectives, this invention provides a method for ultrasound-assisted diagnosis of breast and thyroid glands that integrates expert systems and deep learning, comprising the following steps:
[0010] Step S1: Multimodal data preprocessing and standardization
[0011] Step S1: Multimodal data preprocessing and standardization
[0012] It receives raw images or video streams output by ultrasound equipment, parses DICOM header files to extract device fingerprints, and uses a style transfer model built with adversarial generative networks to map images from different equipment sources to a unified standard grayscale space, eliminating device texture-specific interference.
[0013] Step S2: Deep Learning Feature Perception and Preliminary Classification
[0014] Step S2: Deep Learning Feature Perception and Preliminary Classification
[0015] The standardized image is input into a multi-task deep learning model consisting of an object detection network (YOLO variant), a semantic segmentation network (nnUNet), and a feature classification network (Vision Transformer), which outputs lesion bounding boxes, pixel-level segmentation masks, morphological quantization parameters, and preliminary benign and malignant probability vectors in parallel.
[0016] Step S3: Logical verification of the expert system based on medical knowledge graph
[0017] Step S3: Logical verification of the expert system based on medical knowledge graph
[0018] A rule engine is constructed, incorporating BI-RADS / TI-RADS dictionary encoding, clinical guidelines, and expert consensus. The morphological parameters output from step S2 are read and correlated with structured fields from electronic medical records. If a specific logical conflict pattern is identified (e.g., a history of breast surgery coexisting with an irregular hypoechoic area, or punctate hyperechoic areas within an anechoic thyroid region), the expert system's intervention weights are triggered to correct the initial probabilities of the deep learning output and generate a natural language logical explanation text.
[0019] Step S4: Dynamic weighted fusion of divergence and confidence calibration
[0020] Calculate the discrepancy between the deep learning prediction probability and the expert system prior probability. If the discrepancy is below a threshold, directly adopt the deep learning result to maintain a low-latency response; if the discrepancy is above the threshold, start the Bayesian inference engine to calibrate the posterior probability, and output the final benign / malignant classification and confidence level.
[0021] Step S5: Visualizing the chain of evidence and generating a structured report
[0022] A Grad-CAM heatmap is generated to display the region of interest for deep learning, while highlighting the specific medical terminology codes that trigger the logic verification. The above chain of evidence is then automatically filled into the structured ultrasound report template.
[0023] (III) Beneficial Effects
[0024] Compared with the prior art, the present invention has the following significant advantages:
[0025] Significantly improves generalization ability:
[0026] By using a front-end image normalization network, the cross-device performance degradation rate is reduced from 8%-15% to less than 3%.
[0027] Significantly reduce misdiagnosis of rare cases:
[0028] By using expert rules to logically intercept visual illusions in deep learning, the misjudgment rate of postoperative breast scars was reduced from 25% to below 5%, and the misjudgment rate of thyroid colloids was reduced from 20% to below 2%.
[0029] Significantly reduce hardware requirements and latency:
[0030] For regular images, only lightweight forward inference is required, and complex calculations are triggered only when there is high divergence. It supports real-time operation in a pure CPU environment, reduces the average response latency to 30-40ms, and reduces hardware deployment costs by 80%.
[0031] Enhancing explainability and human-machine trust:
[0032] It provides logical tracing based on medical dictionary coding, which has increased the acceptance rate of doctors' AI conclusions from about 50% to over 85%, and improved the Kappa value of junior doctors' consistency in reading images from 0.65 to 0.82. Attached Figure Description
[0033] Figure 1. Overall flowchart of the ultrasound-assisted diagnostic method integrating expert system and deep learning according to the present invention.
[0034] Figure 2 is a schematic diagram of the structure of the multimodal data preprocessing module of the present invention.
[0035] Figure 3 illustrates the execution logic of the expert system logic verification rule engine of this invention.
[0036] Figure 4. Decision tree flowchart for dynamic fusion and confidence calibration of the present invention.
[0037] Figure 5. Schematic diagram of the auxiliary diagnostic visualization interface generated by the present invention. Detailed Implementation
[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0039] Case Study 1: Postoperative Scar Correction in Breast Nodule Assessment
[0040] Step S1:
[0041] The program receives static images of breast nodules from ultrasound device A. It reads the DICOM header file, identifying the device model as "GE Logiq E9" and the probe frequency as "12MHz". The system then calls a pre-trained CycleGAN generator to map the image style to a standard template, eliminating the speckle noise enhancement effect specific to device A.
[0042] Step S2:
[0043] A standardized image was input to a deep learning model. The YOLOv8 network detected an irregular hypoechoic area with blurred boundaries at the 3 o'clock position of the left breast. The Vision Transformer output a malignancy probability P_dl=0.82.
[0044] Step S3:
[0045] The system retrieved the patient's electronic medical record and extracted the field "Past Medical History: 2 years after local resection of left breast fibroadenoma". This data triggered the "Postoperative Scar Identification" subroutine in the rule engine. The program determined that the morphological description (irregular hypoechoic area) and medical history (postoperative changes) had a high correlation, and set the intervention weight W_expert=0.4.
[0046] Step S4:
[0047] Since the divergence between P_dl and W_expert is greater than the preset threshold of 0.3, the system initiates Bayesian fusion calculation, combines prior knowledge (the probability of malignancy of postoperative scar is extremely low), and outputs the final malignancy probability P_final=0.21.
[0048] Step S5:
[0049] The interface displays a heat map of the designated area, and a pop-up message appears: "Please note: The patient has a clear history of local excision surgery, and the current imaging characteristics tend to indicate postoperative scarring (BI-RADS 2). It is recommended to conduct a follow-up examination with elastography." A structured report is automatically generated.
[0050] Implementation Case 2:
[0051] Differentiation between colloid shrinkage and microcalcification in thyroid nodules
[0052] Step S1:
[0053] The deep learning model identified punctate strong echoes within thyroid cystic nodules, labeled them as "microcalcifications," and output a TI-RADS 4-class probability of 0.75.
[0054] Step S2:
[0055] In step S2, the segmentation network simultaneously outputs that 95% of the nodule's internal area is an "anechoic" region. The rule engine triggers the logic: "IF nodule. Cystic component > 90% AND strong echo. Morphology = punctate THEN Corrected diagnosis = concentrated colloid." The system determines that this is a "comet tail" artifact caused by colloid condensation rather than psammophyll microcalcification.
[0056] Step S3:
[0057] The system forcibly downgraded the TI-RADS classification from Category 4 to Category 2 and output the message: "Hint: Strong punctate echoes in the anechoic area, consider concentrated colloids, it is recommended to observe the comet tail sign in conjunction with dynamic video."
[0058] Implementation Case 3: Edge Computing Deployment Scenarios
[0059] The system software package is deployed on a laptop configured with an Intel i5-1135G7 processor, 16GB of RAM, and no dedicated graphics card. When processing a single frame of 1024x768 resolution ultrasound image, step S2 only runs a lightweight detection branch, and step S3 is a memory database query, with a total time of 36ms. This enables real-time video stream-assisted analysis and meets the screening needs of primary hospitals.
Claims
1. A method for ultrasound-assisted diagnosis of breast and thyroid glands integrating expert systems and deep learning, characterized in that, Includes the following steps: Acquire the ultrasound medical images to be detected and the corresponding structured text data of the electronic medical records; The ultrasound medical images are subjected to standardized preprocessing to eliminate image style differences caused by different acquisition devices. The trained deep learning model is used to extract features from the standardized image and output the morphological quantification parameters of the lesion and the preliminary benign and malignant classification probability. Based on a preset medical knowledge rule engine, and combined with the morphological quantification parameters and the structured text data, the preliminary benign and malignant classification probabilities are logically verified to generate intervention weights. Calculate the discrepancy between the preliminary benign / malignant classification probability and the intervention weight (formula D=|PDL−WES|), and execute a dynamic fusion strategy based on the discrepancy to output the final auxiliary diagnostic result. Calculate the absolute difference of the discrepancy.
2. The method according to claim 1, characterized in that, The standardized preprocessing of ultrasound medical images specifically includes the following. Parse the DICOM header file to obtain device fingerprint information. A pre-trained image style transfer generative adversarial network is invoked to map the ultrasound medical image to a preset standard image style domain. The method according to claim 1, wherein the deep learning model comprises a parallel object detection branch, a semantic segmentation branch, and a feature classification branch; The target detection branch is used to locate the lesion bounding box; The semantic segmentation branch is used to generate a pixel-level segmentation mask and calculate the aspect ratio, quantified value of edge irregularity, and microcalcification density parameter. The feature classification branch outputs a preliminary benign / malignant probability vector based on the Vision Transformer architecture.
3. The method according to claim 1, characterized in that, The deep learning model includes parallel object detection branches, semantic segmentation branches, and feature classification branches; The target detection branch is used to locate the lesion bounding box; The semantic segmentation branch is used to generate a pixel-level segmentation mask and calculate the aspect ratio, quantified value of edge irregularity, and microcalcification density parameter. The feature classification branch outputs a preliminary benign / malignant probability vector based on the Vision Transformer architecture.
4. The method according to claim 1, characterized in that, The logical verification based on the preset medical knowledge rule engine specifically includes: Establish a rule database that includes BI-RADS dictionary encoding, TI-RADS dictionary encoding, and prior probabilities of clinical guidelines; When a preset logical conflict pattern is detected between the combination of the morphological quantization parameters and the structured text data, the corresponding expert system intervention weight calculation formula is triggered.
5. The method according to claim 4, characterized in that, The logical conflict patterns include at least the following: The electronic medical record contains a field for breast surgery history, and the image features an irregular hypoechoic area; Alternatively, the ultrasound image features include a cystic anechoic area that accounts for more than a threshold, and punctate strong echoes detected within it.
6. The method according to claim 1, characterized in that, The step of executing the dynamic fusion strategy based on the degree of divergence specifically includes: If the degree of divergence is less than the preset threshold, the preliminary benign / malignant classification probability of the deep learning model is directly output. If the degree of divergence is greater than or equal to the preset threshold, the Bayesian inference engine is activated, the intervention weight is used as the prior probability to correct the initial benign / malignant classification probability, and the posterior probability is calculated as the final auxiliary diagnostic result.
7. The method according to claim 1, characterized in that, It also includes the following steps: Generate and display a class activation heatmap to explain the decision-making process of the deep learning model; Highlight the specific medical terminology dictionary code and logical description text that triggered the logic verification; The morphological quantification parameters, heatmap, and auxiliary diagnostic conclusions are automatically filled into the structured report template.
8. The method according to any one of claims 1 to 7, characterized in that, The preliminary benign / malignant classification probability includes the probability of breast imaging report and data system grading or the probability of thyroid imaging report and data system grading.
9. An ultrasound-assisted diagnostic system integrating expert systems and deep learning, characterized in that, include: The data acquisition module is used to acquire ultrasound medical images and structured text data; Image normalization module is used to eliminate style differences in images from different devices; The deep learning inference module is used to output morphological parameters and preliminary classification probabilities; The expert system logic engine is used to generate intervention weights based on medical rules; A fusion decision module is used to calculate the degree of divergence and execute the dynamic fusion strategy described in claim 6 based on the degree of divergence. The output module is used to generate visual evidence chains and structured reports.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 8.
11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.
