A method and system for thyroid ultrasound image nodule detection based on pattern recognition

Abstract

This invention relates to the technical field of thyroid ultrasound image detection, and provides a method and system for thyroid ultrasound image nodule detection based on pattern recognition. The method includes: identifying a decisive local region that plays a crucial role in determining thyroid nodules in ultrasound image data based on fine feature data and stable feature data, and performing depth analysis to obtain depth analysis results; when an acoustic shadowing region exists within the decisive local region, inferring the structural characteristics of the acoustic shadowing region based on ultrasound imaging principles and surrounding tissue information, and performing structural feature enhancement processing on the decisive local region to update the depth analysis results; and generating and outputting auxiliary interpretation information based on the fine feature data, stable feature data, and depth analysis results. This invention improves the accuracy and stability of auxiliary interpretation information for thyroid ultrasound image nodule detection.

Description

Technical Field

[0001] This invention relates to the technical field of thyroid ultrasound image detection, specifically to a method and system for detecting thyroid ultrasound nodules based on pattern recognition. Background Technology

[0002] Currently, in thyroid ultrasound examinations, doctors typically need to manually mark suspicious areas and rely on their experience to determine the nature of the nodules. This approach often encounters challenges such as inconsistent diagnostic criteria and limited screening efficiency. To address these issues, a pattern recognition-based thyroid ultrasound image nodule detection system has been developed. This system learns the characteristics of nodules in ultrasound images and performs automatic classification, aiming to improve the standardization and efficiency of diagnosis.

[0003] However, when probe pressure, section angle selection, and differences in patient physiological activity are combined, a significant and unpredictable feature gap arises between the ultrasound images received by the system and the standardized image library used during training. The system begins to experience a precipitous decline in performance, manifesting as missed diagnoses of some real nodules because the nodule morphology is excessively distorted or partially obscured, making its features far removed from the "normal" nodules in the training data. Simultaneously, misdiagnosis of non-nodular structures (such as muscle bundles, vascular cross-sections, and glandular margins) also occurs, as the image features of these structures may coincidentally resemble the nodule features learned by the system under specific scanning conditions and patient states. This unstable detection result not only places unnecessary psychological burden and financial pressure of repeated examinations on patients, but more seriously, it greatly undermines clinicians' trust in this intelligent assisted diagnostic system. Doctors find that they have to spend more time manually reviewing every judgment of the system, and even need to re-evaluate the original images, which actually increases their workload. The intelligent system, originally intended to improve efficiency, becomes a burden requiring extra "care" in practical application.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application discloses a method and system for detecting thyroid ultrasound nodules based on pattern recognition, which aims to solve the problems of inconsistent detection standards, limited screening efficiency, and fluctuations in nodule recognition performance of existing thyroid ultrasound examinations in real clinical applications.

[0006] The technical solution of this application is as follows: In a first aspect, this application discloses a method for detecting thyroid ultrasound nodules based on pattern recognition, comprising the following steps: Acquire ultrasound image data from a thyroid detection instrument; Preliminary analysis of ultrasound image data was conducted to identify preliminary nodule regions based on regional geometric characteristics and the stability of suspected nodule regions in continuous image sequences under morphological changes caused by slight changes in probe pose or physiological micro-movements. Multi-level image processing is performed on the preliminary nodule region to extract features of varying levels of detail. Cross-level features are aligned for consistency, and features with discriminative power for nodule identification are enhanced to obtain refined feature data. Simultaneously, feature tracking is performed on continuous image sequences containing the preliminary nodule region to identify instantaneous positional changes or blurred states. Upon detection of such changes or blurring, stable feature data for the preliminary nodule region is extracted and integrated from the corresponding continuous image sequences. Discriminative power characterizes the degree of difference in feature value distribution across different nodule interpretation categories; features with a difference exceeding a preset threshold are identified as discriminative features. Based on fine and stable feature data, decisive local regions that play a crucial role in the judgment of thyroid nodules in ultrasound images are identified and subjected to depth analysis to obtain depth analysis results. When acoustic shadowing occurs in the decisive local region, the structural characteristics of the acoustic shadowing region are inferred based on the principles of ultrasound imaging and the information of the surrounding tissues. Structural feature enhancement processing is then performed on the decisive local region to update the depth analysis results. The acoustic shadowing region is the area where the local structure of the decisive local region is not fully displayed due to acoustic shadowing, artifacts, or attenuation. Based on detailed feature data, stable feature data, and in-depth analysis results, auxiliary interpretation information is generated and output.

[0007] This technical solution effectively addresses the issue of fluctuating nodule recognition performance in existing technologies due to differences in ultrasound technicians' operating habits, patient physiological conditions, and the complexity of anatomical structures. Through multi-level image processing, feature tracking, and in-depth analysis of acoustic shadowing areas, it significantly improves the accuracy and stability of thyroid ultrasound nodule detection, providing doctors with more reliable auxiliary interpretation information.

[0008] Secondly, this application also discloses a pattern recognition-based thyroid ultrasound image nodule detection system for performing pattern recognition-based thyroid ultrasound image nodule detection, including: The ultrasound image acquisition module is used to acquire ultrasound image data from the thyroid detection instrument; The nodule region identification module is used to perform preliminary analysis of ultrasound image data. Based on the geometric characteristics of the region and the stability of the morphological changes of the suspected nodule region in a continuous image sequence caused by slight changes in probe pose or physiological micro-movements, the module identifies the preliminary nodule region. The image processing execution module performs multi-level image processing on the preliminary nodule region to extract features of different levels of detail, performs consistent alignment processing on cross-level features, and enhances features with discriminative power for nodule identification to obtain refined feature data. Simultaneously, it performs feature tracking on continuous image sequences containing the preliminary nodule region to identify instantaneous positional changes or blurred display states of the preliminary nodule region. When instantaneous positional changes or blurred display states are detected, it acquires and integrates stable feature data of the preliminary nodule region from the corresponding continuous image sequence. Discriminative power is used to characterize the degree of difference in feature value distribution under different nodule interpretation categories; features with a difference degree higher than a preset threshold are determined to be features with discriminative power. The depth analysis execution module is used to identify and perform depth analysis on decisive local regions that play a crucial role in the judgment of thyroid nodules in ultrasound images, based on fine feature data and stable feature data, to obtain depth analysis results. When there are acoustic shadowing areas in the decisive local regions, the structural characteristics of the acoustic shadowing areas are inferred based on the principles of ultrasound imaging and the information of the surrounding tissues of the decisive local regions. Structural feature enhancement processing is then performed on the decisive local regions to update the depth analysis results. Among them, the acoustic shadowing areas are the areas where the local structures of the decisive local regions are not fully displayed due to acoustic shadowing, artifacts, or attenuation. The auxiliary interpretation information module is used to generate and output auxiliary interpretation information based on fine feature data, stable feature data, and in-depth analysis results.

[0009] This application provides a fully functional thyroid ultrasound imaging nodule detection system. Through modular design, it realizes functions such as ultrasound image acquisition, nodule region identification, multi-level image processing, depth analysis, and auxiliary interpretation information generation. It can systematically solve the problem of nodule identification performance fluctuation in the prior art and provide an efficient and accurate tool for clinical diagnosis.

[0010] Beneficial Effects: This application discloses a pattern recognition-based method for thyroid ultrasound image nodule detection. After acquiring ultrasound image data, a preliminary analysis is first performed to identify preliminary nodule regions based on regional geometric characteristics and the morphological stability of suspected nodule regions in a continuous image sequence. Subsequently, multi-level image processing is performed on the preliminary nodule regions to extract features of different levels of detail. Consistency alignment and discriminative enhancement processing are applied to cross-level features to obtain refined feature data. Simultaneously, stable feature data of the preliminary nodule regions is obtained through feature tracking to address instantaneous positional changes or display ambiguity caused by probe pose changes or physiological micro-movements. Based on this, decisive local regions that play a crucial role in thyroid nodule diagnosis are identified and subjected to depth analysis. Specifically, when acoustic shadowing exists in the decisive local region, this application infers the structural characteristics of the acoustic shadowing region based on ultrasound imaging principles and surrounding tissue information, and performs structural feature display enhancement processing to update the depth analysis results. Finally, based on the refined feature data, stable feature data, and depth analysis results, auxiliary interpretation information is generated and output.

[0011] Through the above technical solution, this application effectively solves the problem of fluctuating nodule recognition performance caused by differences in ultrasound technician operating habits, patient physiological states, and the complexity of anatomical structures in existing technologies. This method significantly improves the accuracy and robustness of thyroid nodule detection in complex clinical environments by incorporating considerations of nodule morphological stability in continuous image sequences, as well as depth analysis and structural characteristic inference of acoustically shadowed areas. Especially when acoustic shadowing, artifacts, or attenuation cause incomplete display of local structures, this application can provide more comprehensive and reliable nodule information through intelligent inference and enhancement processing, thereby providing doctors with more accurate auxiliary interpretation information, effectively improving the standardization and efficiency of diagnosis, overcoming the difficulties faced by existing systems in real clinical applications, and achieving unexpected technical results. Attached Figure Description

[0012] Figure 1 This is a flowchart of a method for detecting thyroid ultrasound nodules based on pattern recognition, according to one embodiment of the present invention. Figure 2 This is a flowchart of a method for detecting thyroid ultrasound nodules based on pattern recognition, according to another embodiment of the present invention. Figure 3 This is a system block diagram of a pattern recognition-based thyroid ultrasound image nodule detection system according to another embodiment of the present invention; Explanation of reference numerals in the attached figures: 1. A pattern recognition-based thyroid ultrasound image nodule detection system; 11. Ultrasound image acquisition module; 12. Nodule region recognition module; 13. Image processing execution module; 14. Depth analysis execution module; 15. Auxiliary interpretation information module. Detailed Implementation

[0013] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0014] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0015] This application proposes a pattern recognition-based method for thyroid ultrasound image nodule detection, combined with... Figure 1 As shown, it includes: S1, acquire ultrasound image data from the thyroid detection instrument; S2, perform preliminary analysis of ultrasound image data, and identify preliminary nodule regions based on regional geometric characteristics and the stability of morphological changes in suspected nodule regions in continuous image sequences caused by slight changes in probe pose or physiological micro-movements. S3 performs multi-level image processing on the preliminary nodule region to extract features of different levels of detail, performs consistency alignment processing on cross-level features, and enhances features with discriminative power for nodule identification to obtain refined feature data; simultaneously, it performs feature tracking on continuous image sequences containing the preliminary nodule region to identify instantaneous position changes or blurred display states of the preliminary nodule region, and when instantaneous position changes or blurred display states are identified, it obtains and integrates stable feature data of the preliminary nodule region from the corresponding continuous image sequence; wherein, discriminative power is used to characterize the degree of difference in the value distribution of features under different nodule interpretation categories, and features with a degree of difference higher than a preset threshold are determined to be features with discriminative power; S4, based on fine feature data and stable feature data, identifies the decisive local regions that play a crucial role in determining thyroid nodules in ultrasound image data and performs depth analysis to obtain depth analysis results; when there is an acoustic shadowing area in the decisive local region, the structural characteristics of the acoustic shadowing area are inferred based on the ultrasound imaging principle and the information of the surrounding tissues of the decisive local region, and structural feature display enhancement processing is performed on the decisive local region to update the depth analysis results; where the acoustic shadowing area is the area where the local structure of the decisive local region is not fully displayed due to acoustic shadowing, artifacts or attenuation; S5 generates and outputs auxiliary interpretation information based on fine feature data, stable feature data, and in-depth analysis results.

[0016] To better understand the technical solution proposed in this application, some key terms involved are explained first. Ultrasound image data refers to the raw image information acquired by a thyroid detection instrument, usually stored in digital form, including pixel grayscale values, pixel spatial locations, and image frame sequence information. Preliminary nodule region refers to the suspected nodule region identified through preliminary image analysis in ultrasound images. This region may contain real nodules, artifacts, or normal tissue structures. Discriminative power refers to the degree of difference in the value distribution of a feature across different nodule interpretation categories. When a feature shows a significant difference between benign and malignant nodules, its discriminative power is high and contributes significantly to nodule identification. Fine feature data refers to the set of features extracted from the preliminary nodule region through multi-level image processing. This feature set may include texture features, shape features, edge features, and internal echo structure features. Stable feature data refers to the stable feature information obtained in a continuous image sequence through feature tracking and multi-frame integration techniques. This type of feature can reflect the stable performance of the preliminary nodule region under dynamic changing conditions. The decisive local region refers to a local image area that has a crucial impact on the judgment result during the nodule diagnosis process. This region usually contains imaging information closely related to the pathological structure. The acoustic shadowing region refers to the area where local structures are not fully displayed due to physical effects of ultrasound imaging such as acoustic shadowing, echo attenuation, or artifacts. Auxiliary interpretation information refers to comprehensive information generated by the system based on the analysis results to assist doctors in diagnosis, such as the probability of nodule malignancy, key observation areas, and image enhancement results. The implementation environment of this application can be a computer system integrating an image processing unit, a data storage unit, and a user interface, or a dedicated medical imaging device.

[0017] The core of the pattern recognition-based thyroid ultrasound image nodule detection method proposed in this application lies in the in-depth analysis of ultrasound images through multi-dimensional feature extraction and dynamic information integration, so as to improve the accuracy of nodule identification and judgment in complex clinical environments.

[0018] After acquiring ultrasound image data from a thyroid detection instrument, preliminary analysis is required to identify potential nodule regions. One approach involves using a grayscale threshold-based image segmentation method. By setting a grayscale threshold, regions with grayscale values ​​significantly higher or lower than those of surrounding tissue are identified as suspected nodule regions. Another approach utilizes edge detection algorithms for image boundary identification. For example, the Canny or Sobel operator can be used to extract image edge information, and the preliminary nodule region can be determined based on the degree of edge closure and regional shape characteristics.

[0019] After identifying the initial nodule regions, multi-level image processing is required to extract fine-grained feature data. One approach involves using wavelet transform to decompose the initial nodule regions at multiple scales, thereby extracting image features at different scales. For example, at lower scales, overall nodule morphology and boundary features can be obtained, while at higher scales, microscopic texture and internal speckle structure features can be acquired. Simultaneously, principal component analysis or linear discriminant analysis can be used to perform consistency alignment of cross-level features, thereby reducing redundant information and enhancing discriminative features. For features with significant discriminative value in nodule diagnosis, such as microcalcification features or boundary ambiguity features, feature fusion or weight enhancement methods can be used to increase their contribution to subsequent analysis.

[0020] Furthermore, feature tracking analysis needs to be performed on the continuous image sequence containing the initial nodule region. In one implementation, an optical flow algorithm can be used to track the motion changes of the initial nodule region in consecutive image frames, and a Kalman filter algorithm can be used to smooth the positional changes of the region. When the system detects an instantaneous positional change in the initial nodule region or a blurred image display, such as a significant shift in the region center between consecutive frames or unclear region boundaries, multi-frame image data of the corresponding region can be obtained from the continuous image sequence, and these data can be integrated through averaging or weighted averaging to obtain stable feature data. This multi-frame integration method can effectively reduce the impact of instantaneous noise and motion artifacts on feature extraction.

[0021] After obtaining detailed and stable feature data, it is necessary to identify and analyze the decisive local regions that play a crucial role in nodule diagnosis. One approach is to use saliency detection algorithms to identify the local regions within the initial nodule area that contribute most to the diagnostic results. For example, when microcalcifications are present within the nodule, the microcalcified region and its surrounding tissue structures often constitute a decisive local region. Subsequently, deep model analysis can be performed on this decisive local region, such as using convolutional neural networks to extract high-level semantic features, thereby obtaining more abstract image structural features and generating corresponding deep analysis results.

[0022] When a decisive local area is obscured by acoustic shadowing, it is necessary to infer the structural characteristics of this area based on ultrasound imaging principles and information about the surrounding tissue structure. For example, acoustic attenuation often occurs behind hyperechoic structures, causing the underlying tissue structure to appear as a hypoechoic or absent area in the image. The system can infer the possible structural information present in the acoustic shadowing area by combining the morphological characteristics of the surrounding visible tissue and the pattern of echo intensity changes. For example, when the acoustic shadowing area is adjacent to a hyperechoic calcification point and the surrounding tissue boundary is irregular, it can be inferred that there may be calcification-related lesions in this area. Subsequently, the inferred structural information can be visualized and enhanced using local contrast enhancement or pseudo-color encoding methods, thereby updating the depth analysis results.

[0023] After completing the above analysis, the system can generate and output auxiliary interpretation information based on refined feature data, stable feature data, and deep analysis results. In one implementation, these features can be input into a classification model, such as a support vector machine or random forest model, to calculate the probability that the nodule is benign or malignant. Simultaneously, the system can also generate visual maps of nodule boundaries, internal structural features, and acoustic shadowing regions, presenting this information to doctors in the form of image overlays or diagnostic reports, thus providing doctors with more comprehensive auxiliary diagnostic information.

[0024] Optional, combined Figure 2 As shown, when a sound-shadowing area exists in the decisive local region, the steps for inferring the structural characteristics of the sound-shadowing area based on the principles of ultrasound imaging and information about the surrounding tissues include: A1, acquire image data of the decisive local region; A2 analyzes the gray-level distribution and microstructural heterogeneity within the sound shadow region in the image data of the decisive local region, as well as the consistency of the gradient direction of the sound shadow edge. It also evaluates the uniformity of the visible nodular tissue texture adjacent to the sound shadow and obtains the uncertainty score of the sound shadow region, sound shadow edge, and surrounding visible nodular tissue. Among them, microstructural heterogeneity is used to characterize the statistical difference in speckle texture, microcalcification point echo, and fine echo separation structure at the small neighborhood scale within the sound shadow region. A3. Based on the visible part of the nodule that is not obscured by the sound shadow in the image data of the decisive local region, geometric features and topological features are extracted as the structural characteristics of the sound shadow occluded region, and the deviation of the visible part of the nodule relative to the nodule morphology template library is calculated to obtain the morphological deviation index. A4, integrating uncertainty score and morphological deviation index, generates a local diagnostic confidence map; wherein, the local diagnostic confidence map is a two-dimensional spatial distribution map formed by assigning diagnostic confidence values ​​to pixels or local blocks in a decisive local region. The diagnostic confidence values ​​are calculated based on fine feature data, stable feature data and uncertainty score, and are used to characterize the reliability of the nodule interpretation conclusion of the corresponding local region. A5, based on the local diagnostic confidence map, displays the confidence level of nodule auxiliary judgment on ultrasound images; A6 generates clinical intervention suggestions to assist doctors in making judgments based on the distribution and uncertainty score of the local diagnostic confidence map.

[0025] Specifically, after acquiring image data of the decisive local region, further analysis of the gray-scale distribution characteristics and microstructural heterogeneity within the acoustic shadow region is required. Microstructural heterogeneity can be understood as the degree of statistical difference in statistical features such as speckle texture variations, microcalcified echo points, and fine echo septa within the acoustic shadow region at a small scale. Since ultrasound shadows obscure some tissue structures, making it difficult to directly observe deep tissue information, statistical analysis of the gray-scale distribution and microstructural heterogeneity within the acoustic shadow region can indirectly infer whether potential pathological structures exist beneath the acoustic shadow. Simultaneously, the gradient direction consistency of the acoustic shadow edge region is analyzed to determine the morphological stability of the acoustic shadow edge; and the uniformity of the texture of visible nodular tissue adjacent to the acoustic shadow is evaluated. This information is used together to generate a comprehensive uncertainty score for the acoustic shadow region, its edges, and surrounding visible nodular tissue. This score quantifies the degree of uncertainty of structural information within these regions.

[0026] Furthermore, based on the visible portion of the nodule in the decisive local region that is not obscured by acoustic shadowing, its geometric and topological features are extracted as important bases for inferring the structural characteristics of the acoustic shadowing region. Geometric features can include information such as the overall shape, size, and boundary regularity of the nodule; topological features can include regional connectivity, the number of internal pores, and the distribution relationship of local structures. By analyzing these features, a structural description of the visible portion of the nodule can be obtained, serving as a reference for inferring the potential structure of the acoustic shadowing region. Simultaneously, by comparing the morphological features of the visible portion of the nodule with a pre-established nodule morphology template library, the corresponding morphological deviation index can be calculated. This morphological deviation index reflects the degree of deviation of the nodule morphology from typical benign or malignant morphological templates; a large deviation usually indicates abnormal characteristics in the nodule morphology.

[0027] Subsequently, the uncertainty score and morphological deviation index are fused to generate a local diagnostic confidence map. The local diagnostic confidence map is a two-dimensional spatial distribution map formed by assigning a diagnostic confidence value to each pixel or local image patch within a decisive local region. This diagnostic confidence value is calculated by comprehensively considering information such as fine feature data, stable feature data, and uncertainty scores of the acoustic shadowing region. Its main function is to characterize the reliability of different local regions in the nodule interpretation results. In this way, the contribution degree of different regions to the diagnostic conclusion and the reliability of information can be intuitively displayed in the spatial dimension.

[0028] After generating a local diagnostic confidence map, this map can be overlaid on the original ultrasound image, allowing physicians to visually observe the distribution of diagnostic confidence at different locations within the nodule area. Simultaneously, the system can combine the spatial distribution of the confidence map with uncertainty scores to identify areas with potential diagnostic uncertainty and generate corresponding clinical intervention suggestions. For example, for areas with low confidence and high uncertainty scores, the system can suggest further examinations, such as elastography, fine-needle aspiration biopsy, or regular follow-up observation, thus providing auxiliary reference for clinical decision-making.

[0029] In a preferred embodiment, the above process can be illustrated by the following example. For instance, when detecting nodules on a thyroid ultrasound image, the system identifies a decisive local region, but some structures within this region are significantly obscured by acoustic shadowing, making it difficult to clearly display the deep boundaries and internal structures. First, the system acquires image data of the decisive local region and performs statistical analysis on the grayscale distribution within the shadowed region. The analysis results show that the region exhibits a certain degree of grayscale variation and speckle texture, while also detecting possible microcalcified echo points and fine echo septa. Subsequently, the system evaluates the consistency of the gradient direction at the edge of the shadow and analyzes the uniformity of the visible nodule tissue texture adjacent to the shadow. Based on the above analysis results, the system calculates the corresponding uncertainty score. For example, when the internal microstructure heterogeneity of the shadow is high and the edge morphology is relatively irregular, the uncertainty score will be determined to be at a high level.

[0030] After completing the uncertainty assessment, the system further extracts geometric and topological features from the visible portion of the nodule that is not obscured by acoustic shadows. For example, the system can calculate features such as the ellipticity of the nodule, the smoothness of its boundaries, and the presence of internal anechoic regions, and match these features with a pre-defined library of benign and malignant nodule morphological templates to calculate a morphological deviation index. When the nodule boundary is irregular and its morphological features are close to those of a malignant template, the morphological deviation index will be relatively high.

[0031] Subsequently, the system fuses the uncertainty score and the morphological deviation index to generate a local diagnostic confidence map. In this confidence map, the diagnostic confidence of the acoustically shadowed region is generally low due to incomplete information; while the visible nodule portion, which is not shadowed, has a relatively high diagnostic confidence due to clearer structural information. For example, when the uncertainty score of the acoustically shadowed region is high and the morphological deviation index of the visible nodule portion is high, the system may determine that the region has a potential malignancy risk, but due to the insufficient information caused by acoustic shadowing, the overall diagnostic confidence remains low.

[0032] Ultimately, the system can display the distribution of local diagnostic confidence on ultrasound images using color coding or heatmaps, and provide corresponding auxiliary interpretation information in the diagnostic report. For example, the system may indicate "high structural uncertainty in the acoustic shadowing area; further evaluation of the underlying tissue structure is recommended using elastography or biopsy." In this way, even when acoustic shadowing leads to incomplete image information, the system can still provide doctors with more comprehensive and reliable diagnostic references, thereby improving the accuracy and decision support capabilities of ultrasound diagnosis of thyroid nodules.

[0033] Optionally, when a sound-shadowing area exists in the decisive local region, the step of inferring the structural characteristics of the sound-shadowing area based on the principles of ultrasound imaging and information about the surrounding tissues of the decisive local region includes: The system identifies whether the occlusion or incomplete display of the sound-shadow occlusion area is caused by dynamic factors. The dynamic factors include at least one or more of the following: changes in probe pose, patient physiological micro-movements, and time-varying sound-shadow occlusion. The dynamic factors are used to trigger the detection of inter-frame displacement, changes in boundary sharpness, or changes in texture consistency in a continuous image sequence, so as to generate time-varying characterization parameters corresponding to the dynamic factors and to update stable feature data, uncertainty scores, and local diagnostic confidence maps. Specifically, identifying whether the obscuring or incomplete display of a sound-shadowed region is caused by dynamic factors involves analyzing a continuous sequence of ultrasound images to detect changes in the position, shape, boundary sharpness, or internal texture consistency of the sound-shadowed region across different frames. Dynamic factors can be understood as external or internal disturbances that cause momentary or non-fixed obscuring of the sound-shadowed region in the ultrasound image. For example, changes in probe pose refer to minor adjustments made by the doctor to the probe position or angle during the examination; patient physiological micro-movements refer to minor body movements caused by physiological activities such as breathing and heartbeat; and time-varying sound-shadow obscuring refers to dynamic obscuring caused by changes in the sound propagation path or tissue characteristics. These dynamic factors trigger the detection of inter-frame displacement, changes in boundary sharpness, or changes in texture consistency across a continuous image sequence. For example, by calculating the centroid displacement of the sound-shadowed region between adjacent frames, the average distance change of boundary pixels, or the similarity of local texture features (such as the gray-level co-occurrence matrix), time-varying characterization parameters corresponding to the dynamic factors can be generated. These parameters were then used to update stable feature data, uncertainty scores, and local diagnostic confidence plots to reflect the impact of dynamic changes on these data, thereby improving the accuracy of subsequent analyses.

[0034] When identified as being caused by dynamic factors, the instantaneous motion and deformation of the acoustic shadowing area are tracked in a continuous image sequence, and the instantaneous acoustic properties of the tissues surrounding the acoustic shadowing area are analyzed. When occlusion or incomplete display is identified as being caused by dynamic factors, real-time tracking of the occluded area is necessary. This can be achieved through techniques such as optical flow, feature point matching, or deep learning models, precisely capturing the instantaneous positional changes and shape deformations of the occluded area in a continuous image sequence. Simultaneously, analyzing the instantaneous acoustic properties of the surrounding tissues—for example, by measuring changes in the acoustic attenuation coefficient, echo intensity, or sound velocity—can reveal their acoustic performance under dynamic conditions. This helps to more accurately understand the formation mechanism of sound shadows and their impact on nodular structures.

[0035] Based on instantaneous motion, deformation, and instantaneous acoustic characteristics, adjust the parameters of the ultrasound imaging principle; In practical applications, the parameters of ultrasound imaging principles can be dynamically adjusted based on the tracked instantaneous motion and deformation, as well as the analyzed instantaneous acoustic characteristics. For example, the emission frequency, focal depth, gain setting, dynamic range, or time gain compensation (TGC) curve of the ultrasound probe can be adjusted. The aim is to optimize ultrasound image acquisition, making the display of acoustically obscured areas and surrounding tissues clearer and more stable in dynamic environments, thereby providing higher-quality raw data for subsequent structural characteristic inference.

[0036] By combining the adjusted ultrasound imaging principle parameters with instantaneous surrounding tissue information, the structural characteristics of the acoustic shadowing area can be inferred in real time. Furthermore, after the parameters of the ultrasound imaging principle are adjusted, the structural characteristics of the acoustically shadowed area can be inferred in real time by combining these adjusted parameters with the real-time acquired information of the surrounding tissue. For example, by substituting the adjusted imaging parameters into the sound propagation model and combining them with the acoustic properties of the surrounding tissue (such as acoustic impedance, elastic modulus, etc.), the acoustic attenuation and scattering within the acoustically shadowed area can be simulated more accurately, thereby inferring the density, hardness, or internal microstructure of the shadowed area.

[0037] The stability of the structural characteristics of the sound-shadow occlusion region between consecutive frames of a continuous image sequence is evaluated, and the structural characteristics of the sound-shadow occlusion region of consecutive frames are fused based on the evaluated stability.

[0038] The evaluation of the stability of the structural characteristics of the occlusion region between consecutive frames in a continuous image sequence refers to comparing the consistency or variability of the inferred structural characteristics between different frames. For example, the similarity or difference of inferred structural features (such as texture features and geometric features) between adjacent frames can be calculated. The purpose is to determine the reliability of the inference results under the influence of dynamic factors. When the evaluation results show high stability of the structural characteristics, fusion strategies such as averaging, weighted averaging, or majority voting can be used to integrate the inference results of consecutive frames to obtain a more robust and accurate structural characteristic of the occlusion region. This fusion process helps eliminate noise or transient artifacts that may exist in a single frame image, improving the overall confidence of the inference.

[0039] Optionally, the steps for evaluating the structural characteristics of sound-shadow occlusion regions between consecutive frames of a continuous image sequence include: Feature points are extracted from the structural characteristics inferred in consecutive frames to form multiple sets of local feature points; Calculate the relative position change and local deformation degree of each local feature point set between consecutive frames to obtain local motion deformation parameters; The local motion deformation parameters are compared with the physiological micro-motion range to obtain the range comparison results; When the range comparison result indicates that the local motion deformation parameter is within the physiological micro-motion range, the structural characteristics of the corresponding decisive local area's sound shadow occlusion region are determined to be stable; when the range comparison result indicates that it exceeds the physiological micro-motion range, the structural characteristics of the corresponding decisive local area's sound shadow occlusion region are determined to be unstable. Based on the determination of the structural characteristics of each decisive local region, the stability of the structural characteristics inferred between consecutive frames is comprehensively evaluated.

[0040] Specifically, feature point extraction of structural characteristics inferred from consecutive frames refers to identifying and extracting salient and traceable image feature points, such as corner points, edge points, or texture feature points, from consecutive frames of ultrasound images, targeting the acoustically occluded areas and their surrounding decisive local regions. These feature points are organized into multiple local feature point sets, with the aim of providing quantifiable tracking targets for subsequent motion and deformation analysis.

[0041] This involves calculating the relative positional changes and local deformation degree of each set of local feature points across consecutive frames to obtain local motion deformation parameters. This can be understood as using image processing and computer vision techniques, such as optical flow, feature matching algorithms, or deformation field estimation, to quantify the displacement and shape changes of these feature point sets over time. These parameters reflect the dynamic behavior of the occlusion region across consecutive frames, including translation, rotation, scaling, and more complex local deformations.

[0042] In practical applications, comparing local motion deformation parameters with the physiological micromotion range involves comparing the calculated motion deformation parameters with a pre-set threshold range that represents the patient's physiological micromotions (such as breathing, heartbeat, or slight probe vibration). This physiological micromotion range can be established through statistical analysis of a large number of ultrasound image sequences under normal physiological conditions or through clinical experience. Its purpose is to distinguish between meaningless changes caused by physiological factors and meaningful changes caused by structural changes in the nodule itself or significant probe movement.

[0043] Furthermore, when the range comparison results indicate that the local motion deformation parameters are within the physiological micro-motion range, the structural characteristics of the corresponding decisive local region's sound shadow occlusion area are determined to be stable; when the range comparison results indicate that the parameters exceed the physiological micro-motion range, the structural characteristics of the corresponding decisive local region's sound shadow occlusion area are determined to be unstable. This means that if the dynamic changes of the sound shadow occlusion area are within the physiologically permissible micro-motion range, its internal structural characteristics are considered relatively stable, and reliable inferences and fusion can be made; conversely, if the changes exceed this range, its structural characteristics are considered unstable, and further analysis or prompting the physician to pay attention may be necessary.

[0044] Therefore, based on the determination of the structural characteristics of each decisive local region, a comprehensive evaluation of the stability of the inferred structural characteristics across consecutive frames is conducted. This involves determining the stability of the structural characteristics of the sound-shadow occlusion area within each decisive local region, and then combining the determination results of all relevant regions to perform an overall stability assessment. For example, a comprehensive judgment on whether the structural characteristics of the entire nodule or its key parts are sufficiently stable across consecutive frames can be derived by statistically analyzing the proportion of stable regions, assessing the degree of unstable regions, or combining other contextual information.

[0045] Optionally, the step of extracting feature points from the inferred structural characteristics in consecutive frames to form multiple sets of local feature points includes: Perform a decisive local region texture complexity and gray-level gradient change analysis on the structural characteristics to obtain the texture entropy value and local gray-level gradient intensity; The analysis of decisive local region texture complexity and gray-level gradient changes in structural characteristics involves using image processing algorithms to analyze image data of decisive local regions to quantify the complexity of their texture and the drasticness of gray-level value changes. Texture entropy can be understood as a statistical measure describing the randomness or irregularity of image texture; a higher texture entropy value usually indicates a more complex or disordered texture. Local gray-level gradient intensity reflects the rate and direction of gray-level value changes within a local area of ​​the image and is typically used for edge and structure detection. Its purpose is to provide quantitative foundational data for subsequent sensitivity assessment.

[0046] Based on texture entropy and local gray-level gradient intensity, sensitivity assessment is performed on decisive local regions to obtain preliminary sensitive regions. Specifically, sensitivity assessment refers to determining which parts of a critical local region are more sensitive or important for the diagnostic information of a nodule, based on indicators such as texture entropy value and local gray-level gradient intensity. For example, regions with high texture entropy values ​​and high gray-level gradient intensities may contain rich structural information and have high indicative significance for the morphology, boundaries, and other characteristics of the nodule, and are therefore assessed as initially sensitive regions. Its purpose is to initially identify regions that may contain key diagnostic information.

[0047] For decisive local regions with texture complexity below the preset value, grayscale gradient changes less than the preset degree, or no pathological significance, the priority of feature point extraction is reduced. Lowering the priority of feature point extraction means giving lower weights or excluding regions with flat textures, insignificant grayscale variations, or those that, based on prior medical knowledge, do not conform to the pathological characteristics of nodules, during subsequent feature point extraction. For example, background tissue, homogeneous cystic regions, or known artifact regions may be considered pathologically insignificant. The aim is to reduce interference from irrelevant or low-value feature points and improve the diagnostic relevance of the feature point set.

[0048] After the priority reduction operation is performed, preliminary sensitive areas are screened based on the adjusted priority and combined with prior knowledge of nodule pathology. In practical applications, after the priority reduction operation, the system further refines the screening of initially sensitive areas based on the priority assigned to each region and in conjunction with known prior knowledge of nodule pathology (e.g., malignant nodules often exhibit irregular borders, heterogeneous internal echoes, and microcalcifications). For example, even if a region has a high texture entropy value, its priority may be adjusted if its morphology highly matches the typical characteristics of a benign cyst. The aim is to ensure that the final extracted set of feature points has greater diagnostic value and clinical significance.

[0049] Feature points are extracted from the initially sensitive regions after screening to form multiple sets of local feature points.

[0050] Therefore, representative feature points are extracted from the initial sensitive regions after priority adjustment and screening based on prior pathological knowledge, using appropriate feature point detection algorithms (such as SIFT, SURF, ORB, etc.), and these feature points are organized into multiple local feature point sets. The purpose is to provide high-quality input for subsequent calculation of local motion deformation parameters and stability assessment.

[0051] Optionally, the step of performing a decisive local region texture complexity and gray-level gradient change analysis on the structural characteristics to obtain the texture entropy value and local gray-level gradient intensity may include the following operations: An adaptive filter is applied to the decisive local region, and the shape and size of the filter kernel are adjusted according to the gray-level distribution characteristics of the decisive local region to enhance the contrast of the micro-calcification region and suppress background noise, thus obtaining the filtered image. The adaptive filter can dynamically adjust its filtering parameters, such as the shape and size of the filter kernel, based on local image characteristics. Specifically, this filter can intelligently adjust its range and intensity based on the gray-level distribution characteristics of pixels within a decisive local region, such as statistical measures like the mean, variance, kurtosis, or skewness of gray-level values. Its purpose is to effectively enhance the visual contrast of micro-calcification areas while preserving important image details, making them more prominent in ultrasound images, and simultaneously suppressing background noise that may interfere with nodule identification, thereby obtaining a clearer and more interpretable filtered image.

[0052] Multi-directional edge detection is performed on the filtered image to extract edge intensities in different directions, and the directional consistency of edge intensities within the decisive local region is calculated. Furthermore, after obtaining the filtered image, various multi-directional edge detection algorithms, such as Sobel, Prewitt, Canny, or Gabor filters, can be used to process the image. These algorithms can capture edge information in the image from different angles and scales, thereby extracting edge intensity in multiple directions, such as horizontal, vertical, and diagonal. Subsequently, by analyzing the correlation or difference between these edge intensities in different directions, the directional consistency of edge intensity within a decisive local region can be calculated. Regions with high directional consistency typically indicate clear and continuous structural boundaries, while regions with low directional consistency may indicate complex, ambiguous, or heterogeneous structures.

[0053] Weighted processing of gray-level gradient changes in decisive local regions is performed based on directional consistency. Therefore, the directional consistency calculated above can be used to weight the gray-level gradient changes in decisive local regions. Specifically, in regions with high directional consistency at the edges, higher weights can be assigned to the gray-level gradient changes to emphasize these clear structural boundaries; while in regions with low directional consistency, the weights can be reduced to decrease the influence of noise or blurred regions on the gradient calculation. This weighting process helps to more accurately reflect the true gray-level gradient changes inside or at the edges of nodules, making them more representative of the structural characteristics of nodules.

[0054] In the process of texture complexity analysis, the gray-level histogram of the decisive local region is read and analyzed to calculate the kurtosis coefficient and skewness coefficient, and to obtain the gray-level distribution irregularity. Simultaneously, when performing texture complexity analysis, the gray-level histograms of decisive local regions can be read and analyzed in depth. The gray-level histogram reflects the distribution of the number of pixels at different gray levels in an image. By calculating the kurtosis and skewness coefficients of the histogram, the shape characteristics of the gray-level distribution can be quantified. The kurtosis coefficient describes the sharpness or flatness of the gray-level distribution, while the skewness coefficient describes the symmetry of the gray-level distribution. These statistics can effectively characterize the irregularity of the gray-level distribution within decisive local regions, such as the presence of multiple gray-level peaks or whether gray-level values ​​are concentrated on one side, thus indirectly reflecting the complexity of the texture.

[0055] By combining the weighted grayscale gradient changes and the irregularity of grayscale distribution, we obtain the texture entropy value and the local grayscale gradient intensity.

[0056] Finally, the gray-level gradient changes after directional consistency weighting are combined with the gray-level distribution irregularities obtained through gray-level histogram analysis. This combination can be achieved through weighted summation, feature fusion, or machine learning models. This combined processing allows for a more comprehensive and accurate quantification of the texture entropy and local gray-level gradient intensity of decisive local regions. Texture entropy can be understood as the complexity or randomness of the image texture, while local gray-level gradient intensity reflects the drasticness of local brightness changes in the image. These parameters are crucial for subsequent sensitivity assessment and feature point extraction.

[0057] Optionally, the steps of adjusting the shape and size of the filter kernel based on the grayscale distribution characteristics of the decisive local region include: Statistical analysis of the gray-level distribution in the decisive local region yields the peak value, valley value, and width of the gray-level distribution. Based on the peak value, valley value, and width, identify the transition zone of grayscale overlap and determine the range of the transition zone; Calculate the grayscale distance between the pixel and the peak and valley values ​​within the transition zone, and generate a tendency weight for the pixel to belong to the calcified region or normal tissue. The shape and size of the filter kernel within the transition band are locally adjusted based on the bias weights to enhance the contrast of the calcified region and suppress noise in normal tissue while preserving the structural information of the transition band.

[0058] Specifically, statistical analysis of the grayscale distribution in a decisive local area involves constructing a grayscale histogram for that area and analyzing its statistical characteristics, such as calculating the mode, median, mean, and variance of the grayscale values. This helps identify the main clustering points (peaks), troughs (valleys), and the range (width) of the grayscale distribution. Peaks typically represent the grayscale centers of the main tissue types within the area, while troughs may indicate boundaries between different tissue types or sparse grayscale areas.

[0059] Furthermore, based on the peaks, valleys, and widths of the obtained grayscale distribution, transition zones with overlapping grayscale values ​​can be identified. These transition zones refer to areas on the grayscale histogram where the grayscale distributions of different tissue types (e.g., calcified areas and normal tissue) overlap. By analyzing the relationship between peaks and valleys, the starting and ending grayscale values ​​of these overlapping areas can be determined, thus precisely defining the range of the transition zone. For example, a threshold can be set; when the grayscale value between two peaks is lower than this threshold, the area is considered a transition zone.

[0060] After identifying the transition zone, for each pixel within the transition zone, the grayscale distance between its grayscale value and preset peak values ​​(e.g., grayscale peak values ​​representing calcified regions) and valley values ​​(e.g., grayscale valley values ​​representing normal tissue) is calculated. Based on these distances, a bias weight indicating whether a pixel belongs to a calcified region or normal tissue can be generated. The bias weight can be a value between 0 and 1; for example, the closer a pixel is to the peak value of the calcified region, the higher its bias weight indicating it belongs to the calcified region, and vice versa. This can be achieved using the reciprocal of the distance, a Gaussian function, or other non-linear mapping methods.

[0061] Finally, based on the generated bias weights, the shape and size of the filter kernel within the transition band are locally adjusted. Specifically, for pixels tending towards calcified regions, filter kernel parameters that emphasize edge enhancement and contrast improvement can be used; for example, increasing the sharpness of the filter kernel or using a smaller filter kernel to highlight minute structures. Conversely, for pixels tending towards normal tissue, filter kernel parameters that emphasize noise suppression and smoothing can be used; for example, using a larger filter kernel to smooth background noise. The purpose of this local adjustment is to selectively enhance the contrast of calcified regions and effectively suppress background noise in normal tissue while preserving key structural information (such as tiny calcified edges) within the transition band, avoiding over-smoothing or over-sharpening.

[0062] Optionally, the step of calculating the grayscale distance between a pixel and its peak and valley values ​​within the transition band, and generating a bias weight indicating whether a pixel belongs to a calcified region or normal tissue, may include the following: Within the transition zone, acquire grayscale values, local texture features, and spatial location information for each pixel; The grayscale value, local texture features and spatial location information are fused to form a multi-dimensional feature vector; Calculate the distance between the multidimensional feature vector and the preset feature centers of the calcified region and the feature centers of normal tissue; The distance generates a weight that indicates whether a pixel belongs to a calcified region or normal tissue.

[0063] Within the transition zone, acquiring the grayscale value, local texture features, and spatial location information for each pixel refers to extracting the grayscale value, features reflecting the texture characteristics of the surrounding local area (e.g., texture direction, roughness, contrast), and the pixel's specific location in the image coordinate system for each pixel within the identified grayscale overlapping transition zone region in the ultrasound image. This information aims to provide a comprehensive data foundation for subsequent classification and judgment.

[0064] Furthermore, fusing grayscale values, local texture features, and spatial location information to form a multidimensional feature vector can be understood as integrating the aforementioned information into a unified mathematical vector. This multidimensional feature vector can comprehensively reflect the grayscale performance of pixels, local structural features, and their contextual relationships in the image, thus providing richer and more discriminative information for distinguishing calcified regions from normal tissue.

[0065] Subsequently, the distance between the multidimensional feature vector and the preset feature centers of the calcified region and normal tissue is calculated. This involves comparing the multidimensional feature vector formed by each pixel with the typical feature representations (i.e., feature centers) of the calcified region and normal tissue, which are learned in advance through a large number of samples. This distance can be Euclidean distance, Mahalanobis distance, or other suitable similarity measures, the purpose of which is to quantify the degree of similarity between the current pixel and the two target regions.

[0066] Finally, the pixel's tendency weight for belonging to calcified or normal tissue is generated based on distance. This means assigning a numerical value to each pixel based on the calculated distance, representing the probability or tendency of that pixel belonging to calcified or normal tissue. For example, the closer the pixel is to the feature center of the calcified region, the higher its tendency weight for belonging to the calcified region, and vice versa. These tendency weights will be used for subsequent refined analysis and diagnosis.

[0067] Optionally, before the step of applying the adaptive filter to the decisive local region, the method further includes: Image data of decisive local regions are preprocessed to identify and label potential artifact or acoustic shadow regions; Structural characteristic analysis is performed on artifact or sound shadow regions to obtain the corresponding boundary shape, internal texture pattern and connection characteristics with surrounding tissues, and the structural characteristic analysis results are compiled. Based on the structural characteristic analysis results, the source is determined to distinguish whether the artifact region or the acoustic shadow region originates from external interference or from the internal structure of the nodule. When the source is determined to be external interference, the filter kernel parameters are adjusted to suppress the contrast enhancement processing of the artifact or acoustic shadow region and enhance the corresponding noise suppression; when the source is determined to be internal structure of the nodule, the filter kernel parameters are maintained or enhanced to enhance the contrast of the calcified region.

[0068] Specifically, preprocessing of image data for decisive local regions aims to automatically detect and locate potential artifact or acoustic shadowing regions in the image using image analysis techniques. This can be achieved through various image segmentation or pattern recognition algorithms, such as threshold-based segmentation, region growing, edge detection combined with morphological operations, or semantic segmentation of the image using deep learning models, thereby accurately identifying these abnormal regions. The identified regions are labeled for subsequent processing. Further, structural characteristic analysis of the identified artifact or acoustic shadowing regions involves in-depth analysis of the image features of these regions. This includes extracting the geometry of their boundaries, internal texture patterns, and their connections to surrounding normal tissue or nodular structures. These structural characteristics are compiled into structural characteristic analysis results, providing a basis for subsequent source determination. Source determination based on structural characteristic analysis results aims to distinguish whether the artifact or acoustic shadowing region is caused by external interference from ultrasound equipment or operation (e.g., lateral artifacts, reverberation artifacts) or by the internal structure of the nodule itself (e.g., acoustic shadowing caused by microcalcification or coarse calcification). For example, artifacts originating from external interference may have regular geometric shapes, specific texture patterns related to the probe direction, and unnatural connections with surrounding tissues; while acoustic shadows originating from the internal structures of nodules are usually closely related to the morphology and internal echo characteristics of the nodule, and may be accompanied by attenuation behind the nodule. In practical applications, when artifact or acoustic shadow regions are determined to originate from external interference, the filter kernel parameters are adjusted. Specifically, contrast enhancement processing for these regions is suppressed, because enhancing the contrast of artifacts does not help in nodule diagnosis and may even introduce misleading information. At the same time, corresponding noise suppression is enhanced to eliminate or reduce the impact of external interference on image quality. For example, stronger smoothing filters or specific artifact removal algorithms can be used. Conversely, when the acoustic shadows are determined to originate from the internal structures of the nodule, such as acoustic shadows caused by calcification, the filter kernel parameters are maintained or enhanced. This is because calcification is an important pathological feature of thyroid nodules, and although the acoustic shadows it causes obscure some information, they are themselves important diagnostic evidence. In this case, maintaining or enhancing the filter kernel parameters helps to better display the contrast of the calcified area, thereby assisting doctors in judging the internal structure of the nodule.

[0069] This application proposes a pattern recognition-based thyroid ultrasound image nodule detection system for performing pattern recognition-based thyroid ultrasound image nodule detection, combined with... Figure 3 As shown, the pattern recognition-based thyroid ultrasound image nodule detection system 1 includes: The ultrasound image acquisition module 11 is used to acquire ultrasound image data from the thyroid detection instrument; The nodule region identification module 12 is used to perform preliminary analysis of ultrasound image data. Based on the geometric characteristics of the region and the stability of the morphological changes caused by slight changes in probe pose or physiological micro-movements in the suspected nodule region in a continuous image sequence, the preliminary nodule region is identified. The image processing execution module 13 is used to perform multi-level image processing on the preliminary nodule region to extract features of different levels of refinement, perform consistent alignment processing on cross-level features, and enhance features with discriminative power for nodule identification to obtain refined feature data. At the same time, it performs feature tracking on the continuous image sequence containing the preliminary nodule region to identify instantaneous position changes or blurred display states of the preliminary nodule region. When instantaneous position changes or blurred display states are identified, it obtains and integrates stable feature data of the preliminary nodule region from the corresponding continuous image sequence. Among them, the discriminative power is used to characterize the degree of difference in the value distribution of features under different nodule interpretation categories. Features with a difference degree higher than a preset threshold are determined to be features with discriminative power. The depth analysis execution module 14 is used to identify and perform depth analysis on the decisive local regions that play a crucial role in the judgment of thyroid nodules in ultrasound image data based on fine feature data and stable feature data, and obtain depth analysis results. When there is an acoustic shadowing area in the decisive local region, the structural characteristics of the acoustic shadowing area are inferred based on the ultrasound imaging principle and the information of the surrounding tissues of the decisive local region, and structural feature display enhancement processing is performed on the decisive local region to update the depth analysis results. Among them, the acoustic shadowing area is the area where the local structure of the decisive local region is not fully displayed due to acoustic shadowing, artifacts or attenuation. The auxiliary interpretation information module 15 is used to generate and output auxiliary interpretation information based on fine feature data, stable feature data and deep analysis results.

[0070] To better understand the technical solution proposed in this application, some key terms involved are explained first. Ultrasound image data refers to the raw image information acquired by a thyroid detection instrument, usually stored in digital form, including pixel grayscale values, pixel spatial locations, and image frame sequence information. Preliminary nodule region refers to the suspected nodule region identified through preliminary analysis in the ultrasound image. This region may contain real nodules, artifacts, or normal tissue structures. Discriminative power refers to the degree of difference in the value distribution of a feature across different nodule interpretation categories. When a feature shows a significant difference between benign and malignant nodules, its discriminative power is high and contributes significantly to nodule identification. Fine feature data refers to the set of features extracted from the preliminary nodule region through multi-level image processing. This set may include texture features, shape features, edge features, and grayscale distribution features. Stable feature data refers to the stable feature information obtained in a continuous image sequence through feature tracking and multi-frame integration techniques. This type of feature reflects the stable performance of the preliminary nodule region under dynamic changing conditions. The decisive local region refers to a local image area that has a crucial impact on the interpretation result during nodule diagnosis. This region typically contains the most important pathological structural information of the nodule. The acoustic shadowing region refers to the area where local tissue structures are not fully displayed due to physical phenomena in ultrasound imaging such as acoustic shadowing, artifacts, or acoustic attenuation. Auxiliary interpretation information refers to comprehensive information generated by the system based on the analysis results to assist doctors in diagnosis, such as the probability of nodule malignancy, key areas of focus, and structural enhancement display results. The implementation environment of this application is typically a computer system or dedicated medical device integrating an image processing unit, a data storage unit, and a user interface.

[0071] The thyroid ultrasound image nodule detection system based on pattern recognition proposed in this application is based on the collaborative work of multiple functional modules to achieve multi-dimensional in-depth analysis of ultrasound image data, thereby improving the accuracy and stability of nodule detection and interpretation in complex clinical environments.

[0072] The ultrasound image acquisition module is used to establish a data connection with the thyroid detection instrument and to receive or read ultrasound image data. This module can be understood as a data interface unit whose main function is to provide raw image data input to the system. In specific implementations, this module can interact with the ultrasound equipment through standard medical imaging communication protocols, such as using medical imaging communication protocols to achieve real-time image data transmission, or loading historical ultrasound image data from storage media via file reading.

[0073] The nodule region identification module performs preliminary analysis on the acquired ultrasound image data to identify potential nodule regions. This module can be understood as a preliminary screening unit, its main function being to quickly locate regions where nodules may exist within a large amount of image information. In its implementation, this module can call upon image processing algorithm libraries, such as grayscale thresholding, edge detection, or texture analysis algorithms, to scan and analyze local regions in the image, thereby identifying suspected nodule regions with specific geometric features or dynamic stability. This step effectively reduces the data volume required for subsequent in-depth analysis, thus improving the overall system processing efficiency.

[0074] The image processing execution module performs multi-level image processing and dynamic feature tracking on the initial nodule region. This module can be understood as a multi-dimensional feature extraction and dynamic analysis unit. Its main function is to extract feature information of the nodule region at different scales and obtain stable feature data through continuous image sequence analysis. For example, this module can use wavelet transform or Gabor filters to extract texture features at different scales and obtain nodule boundary features through edge analysis algorithms. Simultaneously, this module can also use optical flow algorithms or Kalman filtering algorithms to track the positional changes of the nodule region in consecutive frames, thereby identifying image changes caused by slight probe movement or patient physiological motion. By integrating multiple frames of images, a more stable and reliable feature representation can be obtained, thus improving the robustness of nodule identification.

[0075] The depth analysis module is used to identify decisive local regions based on fine-grained and stable feature data, and to perform depth analysis on them. When a decisive local region is obscured by acoustic shadowing, this module can also infer the structural characteristics of the obscured area based on ultrasound imaging principles and surrounding tissue structure information. For example, by analyzing the morphological features of the acoustic shadowing region, changes in echo intensity, and the continuity of the surrounding tissue structure, it can infer the possible tissue structures beneath the acoustic shadowing. Simultaneously, this module can also perform structural feature enhancement processing on the decisive local region, such as visualizing the inferred structural information through local contrast enhancement or pseudo-color encoding, thereby updating the depth analysis results. This processing method can still provide a relatively reliable basis for structural judgment even when image information is incomplete.

[0076] The auxiliary interpretation information module is used to transform the system analysis results into auxiliary diagnostic information that is easy for doctors to understand. This module can be understood as a result presentation and interaction unit, whose main function is to generate structured diagnostic information and visualize it. For example, this module can generate a diagnostic report that includes the probability of nodule malignancy, nodule boundary features, internal structural features, and acoustic shadowing region inference results. It can also overlay annotation information on the original ultrasound image, such as displaying decisive local areas and their diagnostic confidence distribution by color coding or region highlighting.

[0077] The overall workflow of this system is as follows: First, the ultrasound image acquisition module acquires thyroid ultrasound image data as the basic input for system analysis. Then, the nodule region identification module performs preliminary analysis of the images, identifying preliminary nodule regions based on regional geometric features and dynamic stability. Next, the image processing module performs multi-level feature extraction and continuous image sequence tracking on the preliminary nodule regions to obtain refined and stable feature data. Then, the depth analysis module identifies decisive local regions based on the aforementioned features, and in the presence of acoustic shadowing, infers the structural characteristics of the obscured region using ultrasound imaging principles and surrounding tissue information, while simultaneously enhancing the display of structural features. Finally, the auxiliary interpretation information module generates auxiliary diagnostic information based on the system analysis results and presents it to the doctor in the form of image overlay or reports.

[0078] Compared with existing technologies, the system of this application has significant advantages in processing thyroid ultrasound images in complex clinical environments. Traditional thyroid ultrasound diagnosis relies heavily on physician experience, and diagnostic results can easily vary between different physicians. Furthermore, some existing automated systems are prone to performance fluctuations when faced with complex situations such as differences in ultrasound technician operation, subtle patient movements, and acoustic shadowing. This application achieves multi-dimensional feature extraction, dynamic stability analysis, and acoustic shadowing region structure inference through a modular system architecture, enabling the system to provide reliable auxiliary diagnostic results even with incomplete information or poor image quality.

[0079] Especially in handling acoustically obscured areas, this application combines ultrasound imaging principles with information from surrounding tissue structures to infer the structure of the obscured area and then visualizes it through enhanced structural feature display. This overcomes the shortcomings of traditional systems that struggle to make effective judgments when acoustically obscured area information is missing. Through this technical solution, the system can provide more stable, comprehensive, and reliable auxiliary interpretation information in complex clinical environments, significantly improving the accuracy and robustness of thyroid ultrasound nodule detection and providing clinicians with more effective diagnostic support.

[0080] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for detecting thyroid ultrasound nodules based on pattern recognition, characterized in that, include: Acquire ultrasound image data from a thyroid detection instrument; A preliminary analysis of the ultrasound image data was performed, and preliminary nodule regions were identified based on the regional geometric characteristics and the stability of the suspected nodule region in a continuous image sequence under the influence of slight changes in probe pose or physiological micro-movements. Multi-level image processing is performed on the preliminary nodule region to extract features of different levels of detail. Cross-level features are aligned for consistency, and features with discriminative power for nodule identification are enhanced to obtain refined feature data. Simultaneously, feature tracking is performed on a continuous image sequence containing the preliminary nodule region to identify instantaneous positional changes or blurred display states. When such instantaneous positional changes or blurred display states are detected, stable feature data of the preliminary nodule region is obtained and integrated from the corresponding continuous image sequence. Discriminative power is used to characterize the degree of difference in feature value distribution under different nodule interpretation categories; features with a difference degree higher than a preset threshold are determined to be discriminative features. Based on the refined feature data and the stable feature data, a decisive local region that plays a crucial role in determining thyroid nodules in ultrasound image data is identified and subjected to depth analysis to obtain depth analysis results. When the decisive local region is obscured by acoustic shadowing, the structural characteristics of the obscured region are inferred based on the principles of ultrasound imaging and the information of the surrounding tissues. Structural feature enhancement processing is then performed on the decisive local region to update the depth analysis results. The obscured region is defined as the area where the local structure of the decisive local region is not fully displayed due to acoustic shadowing, artifacts, or attenuation. Based on the refined feature data, the stable feature data, and the deep analysis results, auxiliary interpretation information is generated and output.

2. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 1, characterized in that, When an acoustic shadowing region exists within the decisive local region, the steps for inferring the structural characteristics of the acoustic shadowing region based on ultrasound imaging principles and information about the surrounding tissues of the decisive local region include: Acquire image data of the decisive local region; The gray-level distribution and microstructural heterogeneity within the acoustic shadow region, as well as the consistency of the gradient direction at the acoustic shadow edge, are analyzed in the image data of the decisive local region. The uniformity of the visible nodular tissue texture adjacent to the acoustic shadow is also evaluated. Uncertainty scores for the acoustic shadow region, acoustic shadow edge, and surrounding visible nodular tissue are obtained. The microstructural heterogeneity is used to characterize the statistical differences in speckle texture, microcalcification echo, and fine echo separation structure at the small neighborhood scale within the acoustic shadow region. Based on the visible portion of the nodules that are not obscured by sound shadows in the image data of the decisive local region, geometric and topological features are extracted as the structural characteristics of the sound shadow occlusion region, and the degree of deviation of the visible portion of the nodules relative to the nodule morphology template library is calculated to obtain the morphological deviation index. By integrating the uncertainty score and the morphological deviation index, a local diagnostic confidence map is generated; wherein, the local diagnostic confidence map is a two-dimensional spatial distribution map formed by assigning diagnostic confidence values ​​to pixels or local blocks in the decisive local region, and the diagnostic confidence values ​​are calculated based on the fine feature data, the stable feature data and the uncertainty score, and are used to characterize the reliability of the nodule interpretation conclusion of the corresponding local region; Based on the aforementioned local diagnostic confidence map, the confidence level for nodule-assisted judgment is displayed on the ultrasound image; Based on the distribution of the local diagnostic confidence map and the uncertainty score, clinical intervention suggestions are generated to assist doctors in making judgments.

3. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 2, characterized in that, When an acoustic shadowing region exists within the decisive local region, the steps for inferring the structural characteristics of the acoustic shadowing region based on ultrasound imaging principles and information about the surrounding tissues of the decisive local region include: Identify whether the occlusion or incomplete display of the sound-shadow occlusion area is caused by dynamic factors; wherein, the dynamic factors include at least one or more of the following: probe pose change, patient physiological micro-movement, and time-varying sound-shadow occlusion. The dynamic factors are used to trigger the detection of inter-frame displacement, boundary sharpness change, or texture consistency change in a continuous image sequence to generate time-varying characterization parameters corresponding to the dynamic factors, and to update stable feature data, uncertainty scores, and local diagnostic confidence maps. When identified as being caused by dynamic factors, the instantaneous motion and deformation of the acoustic shadowing area are tracked in a continuous image sequence, and the instantaneous acoustic characteristics of the tissues surrounding the acoustic shadowing area are analyzed. Based on the instantaneous motion, the deformation, and the instantaneous acoustic characteristics, adjust the ultrasonic imaging principle parameters; The structural characteristics of the acoustic shadowing area are inferred in real time by combining the adjusted ultrasound imaging principle parameters with instantaneous surrounding tissue information; The stability of the structural characteristics of the sound-shadow occlusion region between consecutive frames of a continuous image sequence is evaluated, and the structural characteristics of the sound-shadow occlusion region of the consecutive frames are fused based on the evaluated stability.

4. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 3, characterized in that, The step of evaluating the stability of the structural characteristics of the sound-shadow occlusion region between consecutive frames of a continuous image sequence includes: Feature points are extracted from the structural characteristics inferred in consecutive frames to form multiple sets of local feature points; Calculate the relative position change and local deformation degree of each local feature point set between consecutive frames to obtain local motion deformation parameters; The local motion deformation parameters are compared with the physiological micro-motion range to obtain the range comparison results; When the range comparison result indicates that the local motion deformation parameter is within the physiological micro-motion range, the structural characteristics of the corresponding decisive local area's sound shadow occlusion region are determined to be stable; when the range comparison result indicates that it exceeds the physiological micro-motion range, the structural characteristics of the corresponding decisive local area's sound shadow occlusion region are determined to be unstable. Based on the determination of the structural characteristics of each decisive local region, the stability of the structural characteristics inferred between consecutive frames is comprehensively evaluated.

5. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 4, characterized in that, The step of extracting feature points from the inferred structural characteristics in consecutive frames to form multiple sets of local feature points includes: Perform a decisive local region texture complexity and gray-level gradient change analysis on the structural characteristics to obtain the texture entropy value and local gray-level gradient intensity; Based on the texture entropy value and the local gray-level gradient intensity, the sensitivity of the decisive local region is evaluated to obtain the preliminary sensitive region. For decisive local regions with texture complexity below the preset value, grayscale gradient changes less than the preset degree, or no pathological significance, the priority of feature point extraction is reduced. After the priority reduction operation is performed, the preliminary sensitive areas are screened based on the adjusted priority and combined with prior knowledge of nodule pathology. Feature points are extracted from the preliminarily sensitive regions after screening to form multiple sets of local feature points.

6. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 5, characterized in that, The steps of performing decisive local region texture complexity and gray-level gradient change analysis on structural characteristics to obtain texture entropy values ​​and local gray-level gradient intensity include: An adaptive filter is applied to the decisive local region, and the shape and size of the filter kernel are adjusted according to the gray-level distribution characteristics of the decisive local region to enhance the contrast of the micro-calcification region and suppress background noise, thus obtaining the filtered image. Multi-directional edge detection is performed on the filtered image to extract edge intensities in different directions, and the directional consistency of edge intensities within the decisive local region is calculated. Weighted processing of gray-level gradient changes in decisive local regions is performed based on directional consistency. In the process of texture complexity analysis, the gray-level histogram of the decisive local region is read and analyzed to calculate the kurtosis coefficient and skewness coefficient, and to obtain the gray-level distribution irregularity. By combining the weighted grayscale gradient changes and the irregularity of grayscale distribution, we obtain the texture entropy value and the local grayscale gradient intensity.

7. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 6, characterized in that, The steps of adjusting the shape and size of the filter kernel based on the grayscale distribution characteristics of a decisive local region include: Statistical analysis of the gray-level distribution in the decisive local region yields the peak value, valley value, and width of the gray-level distribution. Based on the peak value, valley value, and width, identify the transition zone of grayscale overlap and determine the range of the transition zone; Calculate the grayscale distance between the pixel and the peak and valley values ​​within the transition zone, and generate a tendency weight for the pixel to belong to the calcified region or normal tissue. The shape and size of the filter kernel within the transition band are locally adjusted according to the aforementioned bias weights to enhance the contrast of the calcified region and suppress normal tissue noise while preserving the structural information of the transition band.

8. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 7, characterized in that, The steps of calculating the grayscale distance between a pixel and its peak and valley values ​​within the transition band, and generating a bias weight indicating whether a pixel belongs to a calcified region or normal tissue, include: Within the transition zone, acquire grayscale values, local texture features, and spatial location information for each pixel; The grayscale value, local texture features and spatial location information are fused to form a multi-dimensional feature vector; Calculate the distance between the multidimensional feature vector and the preset feature center of the calcified region and the feature center of normal tissue; The distance is used to generate a weight that indicates whether a pixel belongs to a calcified region or normal tissue.

9. The method for detecting thyroid ultrasound nodules based on pattern recognition according to claim 6, characterized in that, Before the step of applying the adaptive filter to the deterministic local region, the method further includes: Image data of decisive local regions are preprocessed to identify and label potential artifact or acoustic shadow regions; Structural characteristic analysis is performed on the artifact region or acoustic shadow region to obtain the corresponding boundary shape, internal texture pattern and connection characteristics with surrounding tissues, and the structural characteristic analysis results are obtained by sorting them out. Based on the analysis results of the aforementioned structural characteristics, the source is determined to distinguish whether the artifact region or the acoustic shadow region originates from external interference or from the internal structure of the nodule. When the source is determined to be external interference, the filter kernel parameters are adjusted to suppress the contrast enhancement processing of the artifact or acoustic shadow region and enhance the corresponding noise suppression; when the source is determined to be internal structure of the nodule, the filter kernel parameters are maintained or enhanced to enhance the contrast of the calcified region.

10. A pattern recognition-based thyroid ultrasound image nodule detection system, used to perform pattern recognition-based thyroid ultrasound image nodule detection, characterized in that, include: The ultrasound image acquisition module is used to acquire ultrasound image data from the thyroid detection instrument; The nodule region identification module is used to perform preliminary analysis of the ultrasound image data. Based on the geometric characteristics of the region and the stability of the morphological changes of the suspected nodule region in a continuous image sequence caused by slight changes in probe pose or physiological micro-movements, the module identifies the preliminary nodule region. The image processing execution module performs multi-level image processing on the preliminary nodule region to extract features of different levels of detail, performs consistency alignment processing on cross-level features, and enhances features with discriminative power for nodule identification to obtain refined feature data. Simultaneously, it performs feature tracking on a continuous image sequence containing the preliminary nodule region to identify instantaneous positional changes or blurred display states of the preliminary nodule region. When an instantaneous positional change or blurred display state is identified, it acquires and integrates stable feature data of the preliminary nodule region from the corresponding continuous image sequence. Discriminative power is used to characterize the degree of difference in feature value distribution under different nodule interpretation categories; features with a difference degree higher than a preset threshold are determined to be features with discriminative power. The depth analysis execution module is used to identify and perform depth analysis on the decisive local regions that play a crucial role in determining thyroid nodules in ultrasound image data, based on the fine feature data and the stable feature data, to obtain depth analysis results. When the decisive local region is obscured by acoustic shadowing, the module infers the structural characteristics of the obscured region based on ultrasound imaging principles and the surrounding tissue information, and performs structural feature enhancement processing on the decisive local region to update the depth analysis results. The obscured region is defined as the area where the local structure of the decisive local region is not fully displayed due to acoustic shadowing, artifacts, or attenuation. The auxiliary interpretation information module is used to generate and output auxiliary interpretation information based on the fine feature data, the stable feature data and the deep analysis results.

Images