Automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging

Abstract

This invention relates to the field of image analysis technology, specifically to an automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging. The method involves acquiring images of thyroid nodules; given that the invasive growth of cancer cells in malignant nodules can damage cell structure and induce grayscale mutations, the nodule region is segmented and divided into sub-regions. The grayscale similarity of sub-regions is analyzed to quantify local heterogeneity, yielding spatial grayscale performance indicators; then, the correlation between these indicators and the proportion of blood flow pixels is analyzed to establish a structural-blood flow abnormality association model, obtaining the blood flow supply matching degree; further, based on shape and edge grayscale fluctuation analysis, a structural disorder index is obtained, infiltration candidate areas are located, and blood flow is analyzed to obtain peripheral invasive blood flow indicators. Combining the correlation between these two indicators with the blood flow supply matching degree, a structure-blood flow coupling coefficient is obtained; finally, multiple indicators are integrated to obtain a malignancy risk score, and a classification neural network is trained to classify the risk level of samples, assisting doctors in making refined judgments.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, specifically to an automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound images. Background Technology

[0002] Currently, thyroid nodules have a high detection rate in the general population and are one of the most common diseases of the endocrine system. Differentiation between benign and malignant nodules is crucial for clinical decision-making. By integrating ultrasound modal information and leveraging the powerful feature learning capabilities of AI, a more accurate, objective, and efficient solution for analyzing the benign and malignant nature of thyroid nodules is provided.

[0003] Existing automatic classification methods for benign and malignant thyroid nodules mainly rely on the grayscale features of nodules in a single static image for discrimination. However, malignant nodules are usually accompanied by synergistic changes in structural disorder and abnormal blood flow, and this process is often dynamic. If this feature is ignored and only grayscale information in a single static image is relied upon, the state analysis of the nodules will be insufficient, and the deep correlation between modalities will not be explored, resulting in bias in the final classification decision. Summary of the Invention

[0004] To address the technical problem that malignant nodules are often accompanied by synergistic changes in structural disorder and abnormal blood flow, and that this process is often dynamic, ignoring this characteristic and relying solely on grayscale information from a single static image frame leads to insufficient analysis of the nodule's state, failing to uncover deep correlations between modalities, and resulting in biased final classification decisions, this invention aims to provide an automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging. The specific technical solution adopted is as follows: A series of nodule images of thyroid nodules with known risk levels are obtained in time; each nodule image is segmented to obtain the thyroid nodule region and divide it into sub-regions; Based on the similarity of grayscale performance among sub-regions in each nodule image, the spatial grayscale performance index of each sub-region is determined; blood flow pixels are obtained in each nodule image, and the blood supply matching degree of each nodule image is determined based on the correlation between the proportion of blood flow pixels in sub-regions and the changes in the spatial grayscale performance index of sub-regions. Based on the shape features of the thyroid nodule region and the gray-level fluctuations of edge pixels in each nodule image, a structural disorder index is determined. After performing morphological operations on the thyroid nodule region in each nodule image to obtain infiltration candidate regions, the distribution of blood flow pixels is analyzed to obtain a peripheral invasive blood flow index. The correlation between the peripheral invasive blood flow index and the structural disorder index of all nodule images is analyzed and combined with the blood flow supply matching degree to obtain the structure-blood flow coupling coefficient of the thyroid nodule sample. Based on the structural disorder index, spatial grayscale performance index, blood flow pixel distribution characteristics, structure-blood flow coupling coefficient, and peripheral invasive blood flow index of nodule images, the malignancy risk score of thyroid nodule samples is determined and used to train a classification neural network.

[0005] Furthermore, the method for obtaining the spatial grayscale performance index includes: In each nodule image, the average gray value of all pixels in each sub-region is obtained as the gray value feature. Determine the local neighborhood of each sub-region. Within the local neighborhood of each sub-region, calculate the absolute value of the difference between the gray-scale feature values ​​of the sub-region and each neighboring region, which serves as the gray-scale performance highlighting factor. The normalized value of the mean of the gray-scale performance highlighting factors of each sub-region and all neighboring regions serves as the spatial gray-scale performance index of each sub-region.

[0006] Furthermore, the method for acquiring the blood flow pixels includes: In each nodule image, the thyroid nodule region is processed using the HSV color space separation algorithm to separate the red and blue pixels, which are then used as blood flow pixels in each nodule image.

[0007] Furthermore, the method for obtaining the blood flow supply matching degree includes: In each nodule image, the ratio of blood flow pixels to the total number of pixels in each sub-region is calculated as the blood flow percentage of each sub-region; In each nodule image, the blood flow proportion and spatial grayscale index of all sub-regions are sorted according to the preset arrangement order of the sub-regions to obtain the blood flow proportion sequence and spatial grayscale sequence. The Pearson correlation coefficients of the blood flow percentage sequence and the spatial grayscale sequence are calculated and normalized, and used as the blood flow supply matching degree of each nodule image.

[0008] Furthermore, the method for obtaining the structural disorder index includes: In each nodule image, obtain the minimum circumcircle and maximum incircle of the thyroid nodule region; The Euclidean distance between the center coordinates of the minimum circumcircle and the maximum incircle is used as the center offset, and the ratio of the radii of the minimum circumcircle and the maximum incircle is used as the radius offset ratio. The normalized value of the sum of the radius offset ratio and the center offset is used as the structural disorder factor for each nodule image; In each nodule image, the gray-level gradient variance of the outermost pixels of the thyroid nodule region is used as the gray-level fluctuation factor. The normalized value of the product of the structural disorder factor and the gray-level fluctuation factor for each nodule image is used as the structural disorder index for each nodule image.

[0009] Furthermore, the method for obtaining the invasive blood flow index includes: The HSV color space separation algorithm is used to process the infiltration candidate region, separating the red and blue pixels as the blood flow pixels of the infiltration candidate region; The infiltration candidate region is segmented to obtain region blocks. In each region block, the normalized value of the proportion of blood flow pixels is used as the quantity factor, and the distribution entropy of the quantity factors of all region blocks is used as the quantity disorder factor. In the infiltration candidate region, the pixels that are immediately adjacent to the edge of the thyroid nodule region are taken as neighboring pixels, and the outermost edge pixels of the infiltration candidate region are taken as outer pixels. The proportion of blood flow pixels in neighboring pixels is calculated as the first proportion value, and the proportion of blood flow pixels in peripheral pixels is calculated as the second proportion value. The sum of the first proportion value and a preset constant is used as the denominator, and the second proportion value is used as the numerator. The resulting ratio is used as the edge blood flow gradient index. The value obtained by normalizing the sum of the marginal blood flow gradient index and the quantity disorder factor is used as the peripheral invasive blood flow index.

[0010] Furthermore, the method for obtaining the structure-blood flow coupling coefficient includes: The peripheral invasive blood flow indicators corresponding to all nodule images were arranged in chronological order to obtain the first sequence; The structural disorder factors corresponding to all nodule images are arranged in temporal order to obtain the second sequence; The absolute values ​​of the Spearman rank correlation coefficients of the first and second sequences were calculated as a coherent indicator for the thyroid nodule samples. The mean of the blood supply matching degree of all nodule images and the aforementioned synergistic index is used as the structure-blood flow coupling coefficient of the thyroid nodule sample.

[0011] Furthermore, the method for obtaining the malignancy risk score includes: In each nodule image, the distribution characteristics of blood flow pixels are analyzed to obtain the abnormal blood supply values ​​for each nodule image; The normalized sum of the abnormal blood flow values ​​and the surrounding invasive blood flow indicators of each nodule image was used as the abnormal blood flow performance factor of each nodule image, and the mean of the abnormal blood flow performance indicators of all nodule images was used as the abnormal blood flow performance indicator of the thyroid nodule sample. The mean of the spatial grayscale performance index of all sub-regions in each nodule image is multiplied by the structural disorder index of each nodule image, and the resulting product is used as the preliminary risk index of the thyroid nodule sample. The normalized value of the product of the structure-blood flow coupling coefficient, abnormal blood flow indicators, and preliminary risk index of the thyroid nodule sample is used as the malignancy risk score of the thyroid nodule sample.

[0012] Furthermore, the method for obtaining the abnormal blood supply values ​​includes: In each nodule image, the mean of the blood flow percentage of all sub-regions is calculated as the blood flow performance factor, and the coefficient of variation of the blood flow percentage of all sub-regions is used as the blood flow fluctuation factor. The normalized sum of the blood flow performance factor and the blood flow fluctuation factor is used as the abnormal blood supply value for each nodule image.

[0013] Furthermore, the process of training the classification neural network includes: Obtain the set of clinical characteristic parameters for each thyroid nodule sample; The clinical feature parameter set and malignancy risk score of thyroid nodule samples are concatenated into a feature vector and used as the input of a neural network. The risk level of the thyroid nodule samples is used as the output to train the neural network and obtain a trained classification neural network.

[0014] The present invention has the following beneficial effects: Acquiring multi-frame images of thyroid nodules with known risk levels over time helps analyze the morphological and blood flow changes during nodule tissue micromovement. Because malignant nodule cancer cells grow invasively, disrupting the original cell structure and causing abrupt changes in grayscale values, the thyroid nodule region can be segmented and divided into several sub-regions. Analyzing the similarity of grayscale performance between sub-regions quantifies local heterogeneity characteristics, yielding spatial grayscale performance indicators for each sub-region. Then, analyzing the correlation between the proportion of blood flow pixels in the sub-regions and changes in spatial grayscale performance indicators establishes a quantitative correlation model between structural disorder and blood flow abnormalities. This accurately identifies the unique "spatial abnormality and blood flow abnormality" synergistic features of malignant nodules in each image, obtaining the blood flow supply matching degree. Furthermore, based on nodule shape features and edge pixel grayscale fluctuation analysis, a structural disorder index is obtained, effectively analyzing the invasive edges of nodules. Simultaneously, morphological operations are used to locate invasive candidate regions and analyze blood flow distribution, yielding a peripheral invasive blood flow index for each nodule image. Next, the correlation between the peripheral invasive blood flow index and the structural disorder index for all nodule images is analyzed to reflect the spatiotemporal synergy between quantifying morphological invasion and vascular dynamics infiltration. This is combined with blood supply matching to obtain the structure-blood flow coupling coefficient for thyroid nodule samples, accurately capturing the invasive phenotype of malignant nodules and quantifying the co-evolutionary pattern of structure and blood flow during malignant transformation. Finally, by integrating the structural disorder index, spatial grayscale performance index, blood flow distribution features, structure-blood flow coupling coefficient, and peripheral invasive blood flow index, covering the entire chain of information from nodule morphology to function, a malignancy risk score for thyroid nodules is obtained. This score is then used to train a classification neural network to classify subsequent samples by risk level, thereby assisting doctors in making more refined judgments and improving efficiency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging, provided as an embodiment of the present invention; Figure 2 This is an image of a nodule and a schematic diagram of a thyroid nodule region provided in an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of the automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging provided by this invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of an automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging, according to an embodiment of the present invention. The method includes the following steps: Step S1: Obtain a series of nodule images in time for each thyroid nodule sample with a known risk level; segment each nodule image to obtain the thyroid nodule region and divide it into sub-regions.

[0021] Patients with thyroid nodules confirmed by pathology by doctors were selected as thyroid nodule samples, covering different risk levels (in this embodiment of the invention, the risk levels are divided into four levels: benign, low suspicion, high suspicion, and malignant, which are determined through doctor consultation and other methods). The patients were placed in a supine position with their necks tilted back to fully expose the thyroid region. High-resolution ultrasound equipment was used for ultrasound scanning. The operator systematically scanned both thyroid lobes and isthmus, identified and located the target thyroid nodule in the ultrasound image, adjusted the probe position and angle, and obtained an ultrasound image that clearly showed the maximum diameter of the nodule, including the complete outline of the nodule and the surrounding normal thyroid tissue, thereby obtaining the nodule image of each thyroid nodule sample in time sequence (the nodule image at each time point includes a B-mode image and a color Doppler image obtained by starting the color Doppler mode under the guidance of the baseline B-mode image).

[0022] Then, the pre-trained U-Net segmentation model is used to process each nodule image, identifying and segmenting the thyroid nodule region. Please refer to [link to relevant documentation]. Figure 2 It shows a nodule image and a schematic diagram of the thyroid nodule region in one embodiment of the present invention (where the white area is the thyroid nodule region).

[0023] It should be noted that the thyroid nodule regions in the B-mode image and the color Doppler image of the nodule at the same time are the same, and can be aligned using a rigid registration algorithm (a well-known technique, which will not be elaborated here).

[0024] Based on the segmented thyroid nodule region of each nodule image, adaptive meshing is performed according to the actual size of the nodule (the number of meshes can be determined according to the implementation scenario and is not limited here) to ensure that the physical size of each sub-region is not less than 1mm×1mm, thereby obtaining the sub-region corresponding to each nodule image, which is convenient for subsequent analysis of the local grayscale changes of pixels in each nodule image.

[0025] It should be noted that the collection and acquisition of personal information data in the embodiments of the present invention are all authorized by the relevant users, and the process does not violate relevant laws and regulations, nor does it violate public order and good morals.

[0026] Step S2: Based on the similarity of grayscale performance between sub-regions in each nodule image, determine the spatial grayscale performance index of each sub-region; obtain blood flow pixels in each nodule image, and determine the blood supply matching degree of each nodule image based on the correlation between the proportion of blood flow pixels in the sub-region and the change of the spatial grayscale performance index of the sub-region.

[0027] In clinical practice, malignant nodule cancer cells exhibit invasive growth, destroying the thyroid follicle structure and causing significant differences in cell density, leading to mutations in grayscale values ​​within the thyroid nodule area. Although benign nodules may have follicular dilation, cell proliferation is regulated by the body, and the follicle arrangement still maintains local order, resulting in a relatively uniform grayscale value distribution. Therefore, in each nodule image, we can first analyze the similarity of grayscale performance between sub-regions to determine the spatial grayscale performance index of each sub-region, which is used to reflect the grayscale spatial distribution characteristics of thyroid nodules, thereby characterizing their malignant nodule presentation.

[0028] Preferably, in one embodiment of the present invention, the method for obtaining the spatial grayscale performance index includes: In each nodule image, the average gray value of all pixels in each sub-region is obtained as the gray-level feature value, which represents the average gray-level performance level of each sub-region.

[0029] The local neighborhood of each sub-region is determined (in this embodiment of the invention, the sub-regions at the eight neighboring positions of each sub-region can be regarded as its neighborhood). In the local neighborhood of each sub-region, the absolute value of the difference between the gray-scale feature values ​​of the sub-region and each neighboring region is calculated. The larger the absolute value of the difference, the greater the difference between the gray-scale values ​​of the local neighborhoods, and the higher the possibility of a sudden change in the gray-scale distribution. Therefore, the absolute value of the difference is used as the gray-scale performance highlighting factor between the sub-region and each neighboring region.

[0030] Finally, the mean value of the gray-level performance highlighting factors of each sub-region and all neighboring regions is normalized and used as the spatial gray-level performance index of each sub-region. Based on the above analysis, it can be seen that the larger the spatial gray-level performance index, the more gray-level abrupt changes exist between the sub-region and its neighboring sub-regions, which corresponds to the pathological feature of "disordered cell arrangement" in malignant nodules. Normalization is a technique well-known to those skilled in the art. The normalization function can be linear normalization or standard normalization, such as maximum-minimum value normalization. The specific normalization method is not limited here.

[0031] Malignant nodules exhibit uncontrolled cancer cell proliferation and a significantly higher cellular metabolic rate than benign nodules. They require angiogenesis to meet their nutritional needs, resulting in higher internal blood flow density and a "disordered distribution and supply-demand imbalance." In contrast, the blood flow supply of benign nodules follows physiological regulation, with lower density and distribution matching the functional areas of the tissue. Therefore, blood flow pixels can be obtained from each nodule image, and the proportion of blood flow pixels in sub-regions can be analyzed to characterize blood flow density. Based on this, the correlation between the blood flow density and the changes in spatial grayscale performance indicators of sub-regions can be analyzed to determine the blood flow supply matching degree of each nodule image, reflecting the matching relationship between the internal blood flow density and structure of thyroid nodules.

[0032] Preferably, in one embodiment of the present invention, the method for obtaining blood flow supply matching degree includes: First, blood flow pixels are obtained: In each nodule image (color Doppler image), the thyroid nodule region is processed using the HSV color space separation algorithm, which can separate the red and blue blood flow pixels from the background tissue (usually represented as gray), thereby obtaining the blood flow pixels in each nodule image.

[0033] In each nodule image, the ratio of blood flow pixels to the total number of pixels in each sub-region is calculated as the blood flow percentage of each sub-region, which reflects the blood flow density in each sub-region.

[0034] In each nodule image, the blood flow proportion and spatial grayscale index of all sub-regions are sorted according to the preset arrangement order of the sub-regions (e.g., from top to bottom, from left to right, which can be set according to the implementation scenario and is not limited here). That is, the blood flow proportion sequence and spatial grayscale sequence are obtained directly based on the spatial position relationship of the sub-regions. At this time, the values ​​at the same position in the blood flow proportion sequence and spatial grayscale sequence represent the blood flow proportion and spatial grayscale index of the same sub-region.

[0035] Then, the Pearson correlation coefficients of the blood flow proportion sequence and the malignancy tendency contribution value sequence were calculated. A Pearson coefficient closer to 1 indicates a strong positive correlation between the two, meaning there is a consistent trend between the blood flow density and spatial grayscale abrupt changes. This suggests a statistically high correlation between abnormal blood flow supply and the degree of internal cell disorder in the nodule image, forming consistent malignant evidence of "structural disorder - abnormal blood supply." Conversely, a Pearson correlation coefficient closer to -1 indicates a weak positive correlation, suggesting the nodule is more likely to benign hyperplasia. Therefore, the normalized Pearson correlation coefficient was used as the blood flow supply matching degree for each nodule image. Based on the aforementioned analysis, a higher blood flow supply matching degree indicates a more significant malignant manifestation of the thyroid nodule. The normalization method here can be... function.

[0036] It should be noted that the HSV color space separation algorithm is a well-known technology, and the specific process will not be described in detail here.

[0037] Step S3: Based on the shape features of the thyroid nodule region and the grayscale fluctuation of edge pixels in each nodule image, determine the structural disorder index; after performing morphological operations on the thyroid nodule region in each nodule image to obtain infiltration candidate areas, analyze the distribution of blood flow pixels to obtain the peripheral invasive blood flow index; analyze the correlation between the peripheral invasive blood flow index and the structural disorder index of all nodule images, and combine them with the blood flow supply matching degree to obtain the structure-blood flow coupling coefficient of the thyroid nodule sample.

[0038] Benign thyroid nodules are mostly expansive in growth, and their shape is more closely round due to the physical constraints of the surrounding normal thyroid tissue. Malignant nodules, on the other hand, have an active invasive characteristic and can break through the constraints of the surrounding tissue, resulting in morphological deformation. Therefore, in each nodule image, the shape characteristics of the thyroid nodule region and the gray-scale fluctuation of the edge pixels can be analyzed to determine the structural disorder index.

[0039] Preferably, in one embodiment of the present invention, the method for obtaining the structural disorder index includes: In each nodule image (here, a B-mode image), the smallest circumcircle (the smallest circle that completely surrounds the thyroid nodule region, reflecting the maximum extent of the nodule's expansion) and the largest inscribed circle (the largest circle that is completely contained within the thyroid nodule region, reflecting the core dense region of the nodule) of the thyroid nodule region are obtained.

[0040] Under normal circumstances, when a thyroid nodule is ideally round, the similarity between the smallest circumscribed circle and the largest inscribed circle will be higher. However, malignant nodules often exhibit irregular expansion (invasive growth) due to uncontrolled growth, leading to significant differences between the smallest circumscribed circle and the largest inscribed circle. Therefore, the Euclidean distance between the center coordinates of the smallest circumscribed circle and the largest inscribed circle was calculated as the center offset. The center offset quantifies the degree of spatial misalignment between the smallest circumscribed circle and the largest inscribed circle. The larger the value, the more significant the asymmetry in the growth of the thyroid nodule, and the more significant the malignancy reflected in the morphological characteristics of the thyroid nodule. At the same time, the ratio of the radii of the smallest circumscribed circle and the largest inscribed circle was also calculated as the radius offset ratio. A radius offset ratio greater than 1 and closer to 1 indicates a closer conformity to the standard circle. Therefore, the larger the radius offset ratio, the greater the morphological difference between the smallest circumscribed circle and the largest inscribed circle, and the higher the likelihood of malignancy reflected in the morphology of the thyroid nodule.

[0041] Based on the aforementioned analysis, both the radius offset ratio and the center offset are positively correlated with the likelihood that a thyroid nodule is malignant. Therefore, the normalized sum of the radius offset ratio and the center offset is used as the structural disorder factor for each nodule image. The structural disorder factor reflects the degree of spatial invasion of the surrounding tissue by the thyroid nodule in the nodule image. The higher the value, the less the thyroid nodule is constrained by the surrounding tissue space, and the greater the invasive potential of the nodule.

[0042] Because malignant nodules often have blurred boundaries due to their invasive growth, the gray-level gradient variance of the outermost pixels in each nodule image is used as the gray-level fluctuation factor. The larger the gray-level fluctuation factor, the more drastic the gray-level changes at the edge of the thyroid nodule area, and the higher the probability of it being a malignant nodule.

[0043] Finally, the normalized product of the structural disorder factor and the gray-level fluctuation factor for each nodule image is used as the structural disorder index for each nodule image. Based on the aforementioned analysis, a larger structural disorder index indicates a greater potential for thyroid nodule regions to invade the physiological tissue space, and thus a higher risk. Normalization is a technique well-known to those skilled in the art. The normalization function can be linear normalization, standard normalization, or, for example, maximum-minimum normalization. The specific normalization method is not limited here.

[0044] During the invasive growth of malignant nodules, abnormal blood vessels that point towards the surrounding normal tissue and are randomly distributed can be induced. In color Doppler images, these blood vessels appear as blood flow regions connected to the nodule boundary, with large fluctuations in signal intensity and chaotic spatial distribution. Therefore, in each nodule image, morphological operations can be performed on the thyroid nodule region to obtain infiltration candidate regions. Then, the distribution of blood flow pixels in the infiltration candidate regions can be analyzed to obtain the surrounding invasive blood flow index, which can be used to reflect the malignant invasive characteristics of thyroid nodules.

[0045] Preferably, in one embodiment of the present invention, the method for obtaining peripheral invasive blood flow indicators includes: First, in each nodule image (here, a color Doppler image), based on the outline boundary of the thyroid nodule region, a morphological dilation algorithm is used to expand outward (in this embodiment, the expansion is 5mm; the specific expansion range can be adjusted according to the implementation scenario and is not limited here; the morphological dilation algorithm is a well-known technology and will not be described in detail here), thereby forming a ring-shaped infiltration candidate region.

[0046] Then, similarly, the HSV color space separation algorithm is used to process the infiltration candidate region, separating the red and blue pixels as the blood flow pixels of the infiltration candidate region. The infiltration candidate region is then adaptively segmented (the number of grids can be set according to the implementation scenario; in this embodiment, it is set to 16. The adaptive segmentation method can use superpixel segmentation, which is a well-known technique and will not be elaborated here), resulting in region blocks. It should be noted that during the segmentation process, it is necessary to ensure that each region block contains blood flow pixels.

[0047] In each region block, the normalized value of the proportion of blood flow pixels (each proportion divided by the sum of all proportions) is used as the quantity factor, and the distribution entropy of the quantity factors of all regions blocks is used as the quantity disorder factor. The larger the quantity disorder factor, the more uneven the blood flow distribution, that is, the stronger the spatial heterogeneity of the blood flow distribution, which is consistent with the disordered characteristics of angiogenesis caused by malignant infiltration.

[0048] Malignant nodules typically exhibit a halo-like characteristic with abundant peripheral blood flow and relatively necrotic or poor internal blood flow. Therefore, in the infiltration candidate region of each nodule image, pixels immediately adjacent to the edge of the thyroid nodule region (inner ring) are considered neighboring pixels, and pixels at the outermost edge of the candidate region (outer ring) are considered outer pixels. The proportion of blood flow pixels among the neighboring pixels is calculated as the first proportion value, and the proportion of blood flow pixels among the outer pixels is calculated as the second proportion value. The sum of the first proportion value and a preset constant is used as the denominator, and the second proportion value is used as the numerator. The resulting ratio is used as the edge blood flow gradient index. A larger edge blood flow gradient index indicates that the blood flow around the thyroid nodule exhibits a "high outer, low inner" halo distribution pattern, which also suggests malignant infiltration characteristics. The preset constant is used to prevent the denominator from being 0; here, a value of 0.001 can be used. The specific value can be adjusted according to the implementation scenario and is not limited here.

[0049] Finally, the sum of the marginal blood flow gradient index and the quantity disorder factor was normalized and used as the peripheral invasive blood flow index. Based on the aforementioned analysis, a larger peripheral invasive blood flow index indicates a synergistic amplification effect of "disordered distribution" and "high on the outside, low on the inside" characteristics in the blood flow around the thyroid nodule. Furthermore, a higher value indicates more significant peripheral invasive biological behavior and a higher degree of malignancy in the thyroid nodule. Normalization is a technique well-known to those skilled in the art. The normalization function can be linear normalization, standard normalization, or, for example, maximum-minimum normalization. Specific normalization methods are not limited here.

[0050] The "structural abnormalities" and "hemodynamic abnormalities" of malignant nodules usually originate from the same pathophysiological process, and there should be a deep correlation between the two. Therefore, we can verify the cross-modal consistency between the structural features under B-mode images and the blood flow features under color Doppler images, forming a logical closed loop. That is, we can analyze the correlation between the peripheral invasive blood flow indicators and structural disorder indicators of all nodule images, and combine them with the blood supply matching degree calculated in step S2 to obtain the structure-blood flow coupling coefficient of the thyroid nodule sample.

[0051] Preferably, in one embodiment of the present invention, the method for obtaining the structure-blood flow coupling coefficient includes: Based on the aforementioned steps, the peripheral invasive blood flow indices and structural disorder factors corresponding to each nodule image can be obtained. Therefore, the peripheral invasive blood flow indices corresponding to all nodule images are arranged in chronological order to obtain the first sequence, and the structural disorder factors corresponding to all nodule images are arranged in chronological order to obtain the second sequence.

[0052] Then, the absolute values ​​of the Spearman rank correlation coefficients of the first and second sequences are calculated as a coherence index for thyroid nodule samples. The coherence index quantifies the spatial coherence of the "morphological invasion" and "hemodynamic infiltration" of thyroid nodules. The larger the value, the higher the coherence, and the more significant the malignant manifestation.

[0053] Based on the analysis in step S2, it can be seen that the greater the blood flow supply matching degree, the more significant the malignancy of the thyroid nodule. Therefore, the mean of the blood flow supply matching degree of all nodule images and the aforementioned synergistic index is used as the structure-blood flow coupling coefficient of the thyroid nodule sample. The coefficient ranges from 0 to 1, which comprehensively reflects the pathophysiological consistency of the structure and blood flow characteristics of the thyroid nodule at both the "internal basis" and "external manifestation" levels. The larger the value, the higher the consistency of multimodal evidence and the higher the confidence of the malignancy judgment.

[0054] Step S4: Based on the structural disorder index, spatial grayscale performance index, blood flow pixel distribution characteristics, structure-blood flow coupling coefficient, and peripheral invasive blood flow index of the nodule image, determine the malignancy risk score of the thyroid nodule sample, which is used to train the classification neural network.

[0055] In this step, two main categories of features are integrated: structure-related indicators (spatial grayscale performance indicators, structural disorder indicators) and blood flow distribution-related indicators (peripheral invasive blood flow indicators, structure-blood flow coupling coefficient, and blood flow pixel distribution characteristics) to determine the malignancy risk score of thyroid nodule samples. This can more comprehensively and holistically characterize the malignancy tendency of thyroid nodule samples.

[0056] Preferably, in one embodiment of the present invention, the method for obtaining the malignancy risk score includes: First, we analyze the indicators related to blood flow: in each nodule image, we calculate the mean of the blood flow proportion of all sub-regions as the blood flow performance factor, and the coefficient of variation of the blood flow proportion of all sub-regions as the blood flow fluctuation factor.

[0057] The blood flow performance factor reflects the overall blood flow richness of the region corresponding to the thyroid nodule in each nodule image; a higher value indicates richer overall blood flow. The blood flow fluctuation factor reflects the uniformity of blood flow distribution in the region corresponding to the thyroid nodule in each nodule image; a higher value indicates more uneven distribution. The normalized sum of the blood flow performance factor and the blood flow fluctuation factor is used as the abnormal blood flow value for each nodule image. A higher abnormal blood flow value indicates rich overall blood flow within the thyroid nodule, but with extremely uneven distribution, exhibiting a typical malignant "focal hypervascularization" pattern. Normalization is a technique well-known to those skilled in the art. The normalization function can be linear normalization or standard normalization, such as maximum-minimum normalization; the specific normalization method is not limited here.

[0058] Since a larger peripheral invasive blood flow index in each nodule image indicates a more significant peripheral invasive biological behavior and a higher degree of malignancy in the thyroid nodule, the normalized sum of the abnormal blood flow value and the peripheral invasive blood flow index in each nodule image is used as the blood flow abnormality factor for each nodule image. The mean of the blood flow abnormality indicators of all nodule images is used as the blood flow abnormality indicator for the thyroid nodule sample. Based on the above analysis, it can be seen that a larger blood flow abnormality indicator is considered to be of higher abnormal risk.

[0059] Then, we analyzed the structure-related indicators: because the larger the spatial grayscale performance index of the sub-region in the nodule image, the higher the degree of correspondence with the pathological feature of "disordered cell arrangement" of malignant nodules; the larger the structural disorder index of the nodule image, the greater the potential for the thyroid nodule region to invade the physiological tissue space in the nodule image, which also indicates a higher risk. Therefore, we multiplied the mean of the spatial grayscale performance index of all sub-regions in each nodule image with the structural disorder index of each nodule image, and used the resulting product as the preliminary risk index of the thyroid nodule sample. The larger the preliminary risk index, the higher the degree of contribution of the external malignant tendency of the thyroid nodule.

[0060] Finally, because a larger structure-blood flow coupling coefficient indicates a high degree of consensus among multimodal evidence and the highest confidence level in malignancy assessment, the normalized value of the product of the structure-blood flow coupling coefficient, abnormal blood flow indicators, and preliminary risk index of thyroid nodule samples is used as the malignancy risk score for thyroid nodule samples. Normalization is a technique well-known to those skilled in the art, and the normalization function can be linear normalization, standard normalization, or, for example, maximum-minimum normalization. Specific normalization methods are not limited here.

[0061] Thus, the above process yields the malignancy risk score for all known risk levels of thyroid nodule samples, upon which a classification neural network can be trained.

[0062] Preferably, in one embodiment of the present invention, the process of obtaining the classification neural network includes: Obtain a set of clinical characteristic parameters for each thyroid nodule sample. The set of clinical characteristic parameters should include at least the following: patient age, gender (represented by the number 1 for males and the number 2 for females), nodule diameter, and calcification grade.

[0063] Then, the clinical feature parameter set and malignancy risk score of the thyroid nodule samples are concatenated into a feature vector. The high-dimensional feature vector is used as the input of the neural network, and the risk level of the thyroid nodule samples is used as the output, thereby training the neural network to obtain a trained classification neural network.

[0064] It should be noted that the training process of neural networks is a well-known technique, and the specific process will not be described in detail here; furthermore, in this embodiment of the present invention, the neural network can be a CNN.

[0065] After obtaining the trained classification neural network, the test sample can be analyzed based on the neural network to output its risk level. For test samples with a non-benign risk level, doctors need to make further judgments and analyses in a timely manner to determine whether it is benign or malignant.

[0066] In this embodiment of the invention, all numerical values ​​involved in the calculation have undergone data preprocessing to eliminate the influence of dimensions. The specific means of eliminating the influence of dimensions are well known to those skilled in the art and will not be limited or described in detail here.

[0067] In summary, acquiring multiple time-series ultrasound images of thyroid nodules with known risk levels helps analyze the morphological and blood flow changes in nodule tissue micromovements. Because malignant nodule cancer cells grow invasively, disrupting the original cell structure and causing abrupt changes in grayscale values, the thyroid nodule region can be segmented and divided into several sub-regions. Analyzing the similarity of grayscale performance between sub-regions quantifies local heterogeneity characteristics, yielding spatial grayscale performance indicators for each sub-region. Then, analyzing the correlation between the proportion of blood flow pixels in the sub-regions and changes in spatial grayscale performance indicators establishes a quantitative correlation model between structural disorder and blood flow abnormalities. This accurately identifies the unique "spatial abnormality and blood flow abnormality" synergistic features of malignant nodules in each image, obtaining the blood flow supply matching degree. Furthermore, based on nodule shape features and edge pixel grayscale fluctuation analysis, a structural disorder index is obtained, effectively analyzing the invasive edges of nodules. Simultaneously, morphological operations are used to locate invasive candidate regions and analyze blood flow distribution, yielding a peripheral invasive blood flow index for each nodule image. Next, the correlation between the peripheral invasive blood flow index and the structural disorder index for all nodule images is analyzed to reflect the spatiotemporal synergy between quantifying morphological invasion and vascular dynamics infiltration. This is combined with blood supply matching to obtain the structure-blood flow coupling coefficient for thyroid nodule samples, accurately capturing the invasive phenotype of malignant nodules and quantifying the co-evolutionary pattern of structure and blood flow during malignant transformation. Finally, by integrating the structural disorder index, spatial grayscale performance index, blood flow distribution features, structure-blood flow coupling coefficient, and peripheral invasive blood flow index, covering the entire chain of information from nodule morphology to function, a malignancy risk score for thyroid nodules is obtained. This score is then used to train a classification neural network to classify subsequent samples by risk level, thereby assisting doctors in making more refined judgments and improving efficiency.

[0068] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0069] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. An automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging, characterized in that, The method includes: A series of nodule images of thyroid nodules with known risk levels are obtained in time; each nodule image is segmented to obtain the thyroid nodule region and divide it into sub-regions; Based on the similarity of grayscale performance among sub-regions in each nodule image, the spatial grayscale performance index of each sub-region is determined; blood flow pixels are obtained in each nodule image, and the blood supply matching degree of each nodule image is determined based on the correlation between the proportion of blood flow pixels in sub-regions and the changes in the spatial grayscale performance index of sub-regions. Based on the shape features of the thyroid nodule region and the gray-level fluctuations of edge pixels in each nodule image, a structural disorder index is determined. After performing morphological operations on the thyroid nodule region in each nodule image to obtain infiltration candidate regions, the distribution of blood flow pixels is analyzed to obtain a peripheral invasive blood flow index. The correlation between the peripheral invasive blood flow index and the structural disorder index of all nodule images is analyzed and combined with the blood flow supply matching degree to obtain the structure-blood flow coupling coefficient of the thyroid nodule sample. Based on the structural disorder index, spatial grayscale performance index, blood flow pixel distribution characteristics, structure-blood flow coupling coefficient, and peripheral invasive blood flow index of nodule images, the malignancy risk score of thyroid nodule samples is determined and used to train a classification neural network.

2. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The method for obtaining the spatial grayscale performance index includes: In each nodule image, the average gray value of all pixels in each sub-region is obtained as the gray value feature. Determine the local neighborhood of each sub-region. Within the local neighborhood of each sub-region, calculate the absolute value of the difference between the gray-scale feature values ​​of the sub-region and each neighboring region, which serves as the gray-scale performance highlighting factor. The normalized value of the mean of the gray-scale performance highlighting factors of each sub-region and all neighboring regions serves as the spatial gray-scale performance index of each sub-region.

3. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The method for obtaining the blood flow pixels includes: In each nodule image, the thyroid nodule region is processed using the HSV color space separation algorithm to separate the red and blue pixels, which are then used as blood flow pixels in each nodule image.

4. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The method for obtaining the blood flow supply matching degree includes: In each nodule image, the ratio of blood flow pixels to the total number of pixels in each sub-region is calculated as the blood flow percentage of each sub-region; In each nodule image, the blood flow proportion and spatial grayscale index of all sub-regions are sorted according to the preset arrangement order of the sub-regions to obtain the blood flow proportion sequence and spatial grayscale sequence. The Pearson correlation coefficients of the blood flow percentage sequence and the spatial grayscale sequence are calculated and normalized, and used as the blood flow supply matching degree of each nodule image.

5. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The method for obtaining the structural disorder index includes: In each nodule image, obtain the minimum circumcircle and maximum incircle of the thyroid nodule region; The Euclidean distance between the center coordinates of the minimum circumcircle and the maximum incircle is used as the center offset, and the ratio of the radii of the minimum circumcircle and the maximum incircle is used as the radius offset ratio. The normalized value of the sum of the radius offset ratio and the center offset is used as the structural disorder factor for each nodule image; In each nodule image, the gray-level gradient variance of the outermost pixels of the thyroid nodule region is used as the gray-level fluctuation factor. The normalized value of the product of the structural disorder factor and the gray-level fluctuation factor for each nodule image is used as the structural disorder index for each nodule image.

6. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The method for obtaining the invasive blood flow index includes: The HSV color space separation algorithm is used to process the infiltration candidate region, separating the red and blue pixels as the blood flow pixels of the infiltration candidate region; The infiltration candidate region is segmented to obtain region blocks. In each region block, the normalized value of the proportion of blood flow pixels is used as the quantity factor, and the distribution entropy of the quantity factors of all region blocks is used as the quantity disorder factor. In the infiltration candidate region, the pixels that are immediately adjacent to the edge of the thyroid nodule region are taken as neighboring pixels, and the outermost edge pixels of the infiltration candidate region are taken as outer pixels. The proportion of blood flow pixels in neighboring pixels is calculated as the first proportion value, and the proportion of blood flow pixels in peripheral pixels is calculated as the second proportion value. The sum of the first proportion value and a preset constant is used as the denominator, and the second proportion value is used as the numerator. The resulting ratio is used as the edge blood flow gradient index. The value obtained by normalizing the sum of the marginal blood flow gradient index and the quantity disorder factor is used as the peripheral invasive blood flow index.

7. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The method for obtaining the structure-blood flow coupling coefficient includes: The peripheral invasive blood flow indicators corresponding to all nodule images were arranged in chronological order to obtain the first sequence; The structural disorder factors corresponding to all nodule images are arranged in temporal order to obtain the second sequence; The absolute values ​​of the Spearman rank correlation coefficients of the first and second sequences were calculated as a coherent indicator for the thyroid nodule samples. The mean of the blood supply matching degree of all nodule images and the aforementioned synergistic index is used as the structure-blood flow coupling coefficient of the thyroid nodule sample.

8. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 4, characterized in that, The method for obtaining the malignancy risk score includes: In each nodule image, the distribution characteristics of blood flow pixels are analyzed to obtain the abnormal blood supply values ​​for each nodule image; The normalized sum of the abnormal blood flow values ​​and the surrounding invasive blood flow indicators of each nodule image was used as the abnormal blood flow performance factor of each nodule image, and the mean of the abnormal blood flow performance indicators of all nodule images was used as the abnormal blood flow performance indicator of the thyroid nodule sample. The mean of the spatial grayscale performance index of all sub-regions in each nodule image is multiplied by the structural disorder index of each nodule image, and the resulting product is used as the preliminary risk index of the thyroid nodule sample. The normalized value of the product of the structure-blood flow coupling coefficient, abnormal blood flow indicators, and preliminary risk index of the thyroid nodule sample is used as the malignancy risk score of the thyroid nodule sample.

9. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 8, characterized in that, The method for obtaining abnormal blood supply values ​​includes: In each nodule image, the mean of the blood flow percentage of all sub-regions is calculated as the blood flow performance factor, and the coefficient of variation of the blood flow percentage of all sub-regions is used as the blood flow fluctuation factor. The normalized sum of the blood flow performance factor and the blood flow fluctuation factor is used as the abnormal blood supply value for each nodule image.

10. The automatic classification method for benign and malignant thyroid nodules based on multimodal ultrasound imaging according to claim 1, characterized in that, The process of training the classification neural network includes: Obtain the set of clinical characteristic parameters for each thyroid nodule sample; The clinical feature parameter set and malignancy risk score of thyroid nodule samples are concatenated into a feature vector and used as the input of a neural network. The risk level of the thyroid nodule samples is used as the output to train the neural network and obtain a trained classification neural network.

Images