CT image data analysis method and system based on artificial intelligence

Abstract

The invention provides a CT image data analysis method and system based on artificial intelligence. Effective coding and analysis of a region of interest in target CT image data are realized. By generating target CT interest lesion features and distributing corresponding coordinate indexes for the target CT interest lesion features, lesion areas can be accurately positioned and analyzed. Furthermore, through a coordinate neighborhood enhancement technology, the surrounding environment of the lesion area can be expanded and analyzed, so that lesion feature information is enriched. The standardized conversion process not only generates standardized image features containing detailed lesion information, but also extracts target semantic features corresponding to the standardized image features, so that the accuracy and reliability of lesion diagnosis are remarkably improved. And finally, feature annotation is performed on the target CT image data by using the rich feature information, high-quality annotated CT image data is generated, and data support is provided for deep analysis and auxiliary diagnosis of medical images.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a CT image data analysis method and system based on artificial intelligence. Background Art

[0002] In the field of medical image analysis, CT (computed tomography) images are an important diagnostic tool that provides doctors with rich internal structure information. However, traditional CT image data analysis methods often rely on doctors' experience and manual annotation, which is not only time-consuming and labor-intensive, but may also be affected by human factors, resulting in limited accuracy and consistency of diagnostic results. Summary of the invention

[0003] In view of this, the purpose of this application is to provide a CT image data analysis method and system based on artificial intelligence.

[0004] According to the first aspect of the present application, a method for analyzing CT image data based on artificial intelligence is provided, the method comprising:

[0005] Based on the CT image data analysis instruction, the target CT image data is encoded with a region of interest, and features of the target CT lesion of interest are generated;

[0006] Based on the coordinates of each pixel point in the target CT image data in the target CT image data, a coordinate index is assigned to all pixel points in the target CT image data to generate a target interest coordinate index set corresponding to the target CT interest lesion feature;

[0007] According to the set neighborhood enhancement range, performing at least one round of coordinate neighborhood enhancement on the target interest coordinate index set;

[0008] After generating a neighborhood enhancement coordinate index set in each round of coordinate neighborhood enhancement, a target neighborhood enhancement region feature corresponding to the neighborhood enhancement coordinate index set is subjected to a standardized transformation to generate a standardized image feature including the target CT lesion feature of interest and a target semantic feature corresponding to the standardized image feature; the standardized image feature includes: a knowledge point lesion feature corresponding to at least one knowledge point in a lesion diagnosis knowledge network under a target lesion diagnosis label;

[0009] The target CT image data is feature-labeled according to the standardized image features and the target semantic features to generate labeled CT image data.

[0010] According to the second aspect of the present application, a smart medical system is provided, which includes a machine-readable storage medium and a processor, wherein the machine-readable storage medium stores machine-executable instructions, and when the processor executes the machine-executable instructions, the smart medical system implements the aforementioned artificial intelligence-based CT image data analysis method.

[0011] According to the third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned artificial intelligence-based CT image data analysis method is implemented.

[0012] According to any one of the above aspects, the technical effect of the present application is:

[0013] The embodiment of the present application realizes the effective encoding and analysis of the region of interest in the target CT image data. By generating the target CT lesion feature of interest and assigning the corresponding coordinate index to it, the lesion area can be accurately located and analyzed. Furthermore, through the coordinate neighborhood enhancement technology, the surrounding environment of the lesion area can be expanded and analyzed, thereby enriching the lesion feature information. The standardized conversion process not only generates standardized image features containing detailed lesion information, but also extracts the corresponding target semantic features, which significantly improves the accuracy and reliability of lesion diagnosis. Finally, the target CT image data is feature-annotated using this rich feature information to generate high-quality annotated CT image data, which provides data support for in-depth analysis and auxiliary diagnosis of medical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 A schematic diagram of the process of the artificial intelligence-based CT image data analysis method provided in the embodiment of the present application is shown;

[0016] Figure 2 A schematic diagram of the component structure of the intelligent medical system provided in an embodiment of the present application for implementing the above-mentioned artificial intelligence-based CT image data analysis method is shown. DETAILED DESCRIPTION

[0017] The embodiments of the present application are described below in conjunction with the drawings in the present application. It should be understood that the implementation methods described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0018] It will be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present application refer to that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude the implementation as other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the technical field. It should be understood that when an element is "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element, or may refer to the element and the other element establishing a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling, and the term "and / or" used herein indicates at least one of the items defined by the term, such as "A and / or B" may be implemented as "A", or as "B", or as "A and B".

[0019] In order to make the purpose, technical solution and advantages of the present application clearer, the implementation mode of the present application will be further described in detail with reference to the accompanying drawings. The technical solution of the embodiment of the present application and the technical effect produced by the technical solution of the present application are explained below by describing several exemplary implementation modes. It should be pointed out that the following implementation modes can refer to, draw lessons from or combine with each other, and the same terms, similar features and similar implementation steps in different implementation modes are not described repeatedly.

[0020] Figure 1 The flowchart of the CT image data analysis method and system based on artificial intelligence provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps of the CT image data analysis method based on artificial intelligence of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the CT image data analysis method based on artificial intelligence include:

[0021] Step S110, based on the CT image data analysis instruction, the region of interest of the target CT image data is encoded to generate the target CT lesion of interest feature.

[0022] In this embodiment, it is assumed that the server receives a CT image data analysis instruction from a medical institution, requiring analysis of a patient's lung CT image data. The server first obtains the target CT image data, which is a two-dimensional matrix composed of a series of gray values, representing a tomographic image of the patient's lungs.

[0023] The server uses a pre-trained deep learning model to identify regions of interest (ROIs) in images. For example, for a lung CT image, areas where nodules or shadows appear may be identified as ROIs. The model extracts and encodes features for these areas.

[0024] Assume that in this set of lung CT images, the server identifies a lung nodule area, and generates the target CT lesion feature of interest by encoding the texture, shape, gray value distribution and other features of this area. For example, this lesion feature may include information such as the size of the nodule, the degree of irregularity of the shape, the mean and variance of the gray value, etc.

[0025] Step S120, based on the coordinates of each pixel point in the target CT image data in the target CT image data, a coordinate index is assigned to all pixel points in the target CT image data to generate a target interest coordinate index set corresponding to the target CT interest lesion feature.

[0026] After completing step S110 , the server continues processing. For each pixel in the target CT image data, the server assigns a coordinate index according to its position in the image (for example, represented by two-dimensional coordinates (x, y)).

[0027] Assuming that the target CT image data is a 512×512 pixel image, the server will assign a unique index value from 0 to 512×512-1 to each pixel.

[0028] Then, the server determines the coordinate index of the corresponding pixel points based on the previously generated target CT lesion of interest feature, and combines these index values ​​into a target interest coordinate index set. For example, if the target CT lesion of interest feature corresponds to a nodule area of ​​50×50 pixels, the server will extract the coordinate index of all pixels in this area to form a target interest coordinate index set.

[0029] Step S130: performing at least one round of coordinate neighborhood enhancement on the target interest coordinate index set according to the set neighborhood enhancement range.

[0030] In this embodiment, the neighborhood enhancement range set by the server is 3 pixels. Based on this, the server performs coordinate neighborhood enhancement on the target interest coordinate index set.

[0031] Taking a pixel with a coordinate index value of (100,200) as an example, during the first round of coordinate neighborhood enhancement, the server will also include the coordinate indexes of the pixels within the 3-pixel range around it (for example, the pixels within the range of (97,197) to (103,203)) into the new set.

[0032] Assume that there are initially 1000 coordinate indexes in the target interest coordinate index set. After the first round of neighborhood enhancement, the new set may increase to 1500 coordinate indexes.

[0033] If multiple rounds of enhancement are required, the server will continue in the same manner, and each round will include the pixel coordinate indexes within a specified range around the coordinate index after the previous round of enhancement.

[0034] Step S140, after generating a neighborhood enhancement coordinate index set in each round of coordinate neighborhood enhancement, the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set are standardized and transformed to generate standardized image features including the target CT lesion of interest features and target semantic features corresponding to the standardized image features. The standardized image features include: knowledge point lesion features corresponding to at least one knowledge point in the lesion diagnosis knowledge network under the target lesion diagnosis label.

[0035] After each round of coordinate neighborhood enhancement generates a neighborhood enhancement coordinate index set, the server will perform a standardized transformation on the corresponding target neighborhood enhancement area features.

[0036] Assume that in the area corresponding to the neighborhood enhancement coordinate index set generated after the first round of enhancement, the server extracts some new features, such as density changes of surrounding tissues, blood vessel distribution, etc.

[0037] The server first extracts the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set from the target CT image data, and then performs regularized semantic mapping on these features.

[0038] For example, the server compares and maps these features with the semantic features in the pre-established feature knowledge base. If it is found that these features match the semantic feature in the knowledge base that indicates "high probability of early lung cancer", then this semantic feature will be output as the target semantic feature, and the current target CT lesion of interest feature will be output as the standardized image feature.

[0039] If they do not match, the server will decide the subsequent operation based on the current round number. If the current round number is less than the set neighborhood enhancement range, for example, the current round is the first round, the server will output the newly generated reference semantic features as the target semantic features, and output the target neighborhood enhancement area features as the standardized image features, and then continue to the next round of coordinate neighborhood enhancement.

[0040] If the current number of rounds has reached the set neighborhood enhancement range, for example, it has reached the third round and the features still do not match, the server will also output the reference semantic features as the target semantic features and the target neighborhood enhancement area features as the standardized image features, thus ending the coordinate neighborhood enhancement of the target interest coordinate index set.

[0041] Step S150 , feature-labeling the target CT image data according to the standardized image features and the target semantic features, to generate labeled CT image data.

[0042] In this embodiment, the server performs feature annotation on the target CT image data based on the generated standardized image features and target semantic features.

[0043] In one way, the server determines a first setting template for embedding the target semantic feature. Assume that the first setting template displays the relevant semantic description in a specific font and color in the upper right corner of the image. The server embeds the target semantic feature into the target CT image data based on the first setting template to generate the first temporary image data.

[0044] For example, if the target semantic feature is "pulmonary nodules, suspected to be benign", the server will display this description in yellow font in the upper right corner of the image to generate the first temporary image data. Then, through the set first CT enhanced tracking and annotation strategy, the standardized image features and target semantic features in the first temporary image data are CT enhanced tracking and annotation. For example, use specific lines or color boxes to highlight the nodule area, and annotate the nodule area with information such as "nodule size: XX mm" next to it, and finally generate annotated CT image data.

[0045] Another way is that the server directly replaces the standardized image features in the target CT image data with the corresponding target semantic features to generate the second temporary image data. For example, some numerical values ​​representing nodule features are directly replaced with the description of "benign nodule". Then, through the set second CT enhanced tracking and annotation strategy, the target semantic features in the second temporary image data are CT enhanced tracking and annotation to generate annotated CT image data.

[0046] In this way, the processing and annotation of the target CT image data are completed, providing clearer and easier-to-understand information for medical diagnosis.

[0047] In addition, assume another scenario is the processing of brain CT image data.

[0048] In step S110, the CT image data analysis instruction received by the server is for the ischemic area of ​​the brain. Through model analysis, a possible ischemic area is determined, and the region of interest is encoded, and the target CT lesion of interest features are generated, such as the size, position, gray value change, etc. of the ischemic area.

[0049] In step S120, coordinate indexes are assigned to all pixel points according to the coordinates of each pixel in the brain CT image, and a target interest coordinate index set corresponding to the target CT interest lesion feature is generated.

[0050] In step S130, the neighborhood enhancement range is set to 2 pixels. For each coordinate index in the target interest coordinate index set, the server incorporates the pixel coordinate indexes within the 2 pixel range around it into a new set and performs multiple rounds of enhancement.

[0051] In step S140, after each round of enhancement, the server extracts the features of the neighborhood enhancement area, such as the morphology of the surrounding blood vessels, the density change of brain tissue, etc., and performs regularized semantic mapping. If it matches the "mild ischemia" semantic feature in the knowledge base, the corresponding standardized image features and target semantic features are output; if it does not match, whether to continue enhancement and processing is determined according to the number of rounds.

[0052] In step S150, the server annotates the brain CT image data in one of the above two ways, for example, directly annotating "mild ischemic area" on the image and highlighting it with a specific color, or replacing the relevant numerical features with semantic descriptions before annotating.

[0053] Let’s assume a scenario for liver CT image data.

[0054] In step S110, the analysis instruction is to find the tumor area in the liver. The server identifies and encodes the features of the tumor area through the model, and generates target CT lesion features of interest, such as the boundary clarity and internal echo of the tumor.

[0055] In step S120, coordinate indexes are assigned to the pixels of the liver CT image to form a target interest coordinate index set.

[0056] In step S130, the neighborhood enhancement range is set to 4 pixels, and the target interest coordinate index set is enhanced.

[0057] In step S140, semantic mapping and judgment are performed on the enhanced regional features to determine the standardized image features and the target semantic features.

[0058] In step S150, target semantic features such as "suspected malignant tumor" are annotated on the liver CT image in an appropriate manner to generate annotated CT image data.

[0059] Based on the above steps, the embodiment of the present application realizes the effective encoding and analysis of the region of interest in the target CT image data. By generating the target CT lesion feature of interest and assigning the corresponding coordinate index to it, the lesion area can be accurately located and analyzed. Furthermore, through the coordinate neighborhood enhancement technology, the surrounding environment of the lesion area can be expanded and analyzed, thereby enriching the lesion feature information. The standardized conversion process not only generates standardized image features containing detailed lesion information, but also extracts the corresponding target semantic features, which significantly improves the accuracy and reliability of lesion diagnosis. Finally, the target CT image data is feature-annotated using this rich feature information to generate high-quality annotated CT image data, which provides data support for in-depth analysis and auxiliary diagnosis of medical images.

[0060] In a possible implementation, step S140 includes:

[0061] Step S141: extracting the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set from the target CT image data.

[0062] Step S142: performing regularized semantic mapping on the target neighborhood enhanced region features to generate reference semantic features of the target neighborhood enhanced region features.

[0063] Step S143, when the reference semantic feature matches the basic semantic feature, the basic semantic feature is output as the target semantic feature, and the target CT lesion of interest feature is output as the standardized image feature. The basic semantic feature is a semantic feature generated by performing the regularized semantic mapping on the target CT lesion of interest feature.

[0064] Step S144, when the reference semantic feature does not match the basic semantic feature, output the reference semantic feature as the target semantic feature, and output the target neighborhood enhanced region feature as the standardized image feature.

[0065] In this embodiment, illustratively, during the processing of lung CT image data, the server first completes multiple rounds of coordinate neighborhood enhancement of the target interest coordinate index set. Next, the server needs to perform a standardized transformation on the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set to generate standardized image features containing the target CT interest lesion features and the corresponding target semantic features.

[0066] In the first step, the server extracts the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set from the target CT image data.

[0067] Assume that after three rounds of coordinate neighborhood enhancement, the area corresponding to the neighborhood enhancement coordinate index set generated in the latest round contains the tissue information of the lung nodule and its surrounding area. The server extracts the target neighborhood enhancement area features by reading the pixel values, grayscale values, texture and other features corresponding to these coordinate indexes in the target CT image data. For example, the server obtains that the grayscale value distribution range of the pixels in the area is between [50,200], the texture is rough to a certain extent, and the shape is irregular.

[0068] In the second step, the server performs regularized semantic mapping on the target neighborhood enhanced region features to generate reference semantic features of the target neighborhood enhanced region features.

[0069] In order to achieve regularized semantic mapping, the server first deconstructs the feature space of the target neighborhood enhancement area features. It decomposes the gray value distribution, texture features, shape features, etc. into multiple independent first deconstructed CT lesion features of interest. For example, the gray value distribution is deconstructed into features such as "gray value mean is 120" and "gray value variance is 30"; texture features are deconstructed into features such as "high texture roughness"; shape features are deconstructed into features such as "shape irregularity is 0.7".

[0070] At the same time, the server performs feature space deconstruction on the semantic features in the set feature knowledge base to generate multiple second deconstructed CT lesion features of interest. Assuming that the feature knowledge base stores various semantic features about lung nodules, such as "benign nodule features" may be deconstructed into features such as "grayscale value average between 80-150", "low texture roughness", and "high shape regularity"; "malignant nodule features" may be deconstructed into features such as "grayscale value average greater than 150", "high texture roughness", and "high shape irregularity".

[0071] The server obtains characteristic attributes of the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature, such as gray value mean, variance, texture roughness, shape irregularity, etc.

[0072] Based on these feature attributes, the server performs morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature to generate a deconstructed feature morphological matching degree. Before performing the morphological comparison, the server determines the morphological matching feature of the first deconstructed CT lesion of interest feature according to a preset morphological matching feature sequence. For example, the preset morphological matching feature sequence specifies the focus on the gray value mean and shape irregularity.

[0073] For the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature with the same characteristic attribute, the server calculates the first deviation between them. Assuming that for the attribute of gray value mean, the gray value mean of the first deconstructed CT lesion of interest feature is 120, while the gray value mean range of the second deconstructed CT lesion of interest feature (such as benign nodule feature) is 80-150, then the calculated first deviation is smaller.

[0074] At the same time, the server determines a second deviation between the morphological matching feature (such as gray value mean and shape irregularity) having the same characteristic attributes and the second deconstructed CT lesion of interest feature.

[0075] The deconstructed feature morphology matching degree is determined according to the reciprocal corresponding to the smallest deviation between the first deviation and the second deviation. If the first deviation is smaller, then its reciprocal is determined as the deconstructed feature morphology matching degree.

[0076] Next, the server performs a semantic comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature based on the feature attributes, and generates a semantic matching degree of the deconstructed features.

[0077] The server encodes the first deconstructed CT lesion of interest feature, the second deconstructed CT lesion of interest feature, and the morphological matching feature respectively, and generates a first encoding vector of the first deconstructed CT lesion of interest feature, a second encoding vector of the second deconstructed CT lesion of interest feature, and a third encoding vector of the morphological matching feature, for example, using one-hot encoding or other suitable encoding methods.

[0078] Then, the server determines a first correlation coefficient between the first code vector and the second code vector, and determines a second correlation coefficient between the third code vector and the second code vector. By comparing the two correlation coefficients, considering their size, direction and other factors, the semantic matching degree of the deconstructed feature is determined according to certain rules.

[0079] Finally, the server performs a fusion calculation on the deconstructed feature morphological matching degree and the deconstructed feature semantic matching degree, and then calculates the inverse, generating the feature distance between the target neighborhood enhanced region feature and the semantic feature. The fusion calculation can use weighted average or other appropriate methods to comprehensively consider the degree of morphological and semantic matching.

[0080] The server outputs the semantic feature with the lowest feature distance as the reference semantic feature of the target neighborhood enhancement region feature. Assume that after calculation, the semantic feature with the smallest distance to the current target neighborhood enhancement region feature is "suspected early malignant nodule".

[0081] In the third step, the server determines whether the reference semantic feature matches the basic semantic feature.

[0082] The basic semantic feature is a semantic feature generated by regularizing the semantic mapping of the initial target CT lesion of interest. Assuming that the basic semantic feature is "suspected benign nodule" and the currently generated reference semantic feature is "suspected early malignant nodule", the two do not match.

[0083] In this case, the server outputs the reference semantic features as the target semantic features and outputs the target neighborhood enhanced region features as the standardized image features.

[0084] The server will record the judgment results and the generated target semantic features and standardized image features for the subsequent generation and use of annotated CT image data.

[0085] Therefore, the target neighborhood enhanced regional features in the lung CT image data can be accurately standardized and transformed, providing valuable information for subsequent feature labeling and medical diagnosis.

[0086] Throughout the entire processing process, the server makes full use of deep learning and data analysis technology to perform fine processing and comparison of a large amount of feature data to improve the accuracy and reliability of diagnosis. At the same time, as the amount of processed data increases, the server can also continuously optimize and improve methods such as feature deconstruction and matching calculation, further enhancing its support for lung disease diagnosis.

[0087] In a possible implementation, step S144 includes:

[0088] Gets the current round number of the coordinate neighborhood enhancement.

[0089] When the reference semantic feature does not match the basic semantic feature and the current round number is less than the neighborhood enhancement range, the reference semantic feature is output as the target semantic feature, and the target neighborhood enhancement area feature is output as the standardized image feature, thereby ending the coordinate neighborhood enhancement of the target interest coordinate index set.

[0090] Alternatively, when the reference semantic feature does not match the basic semantic feature and the current round number is less than the neighborhood enhancement range, the reference semantic feature is output as the target semantic feature, and the target neighborhood enhancement area feature is output as the standardized image feature, and the next round of coordinate neighborhood enhancement of the target interest coordinate index set is continued until the current round number is equal to the neighborhood enhancement range, at which time the coordinate neighborhood enhancement of the target interest coordinate index set is terminated.

[0091] In the process of processing lung CT image data, the server has completed the generation of reference semantic features of the target neighborhood enhanced regional features, and matched and judged them with the basic semantic features.

[0092] Assume that the neighborhood enhancement range set by the server is 5 rounds. The current round number of coordinate neighborhood enhancement obtained by the current server is 3. At this time, the reference semantic feature is "pulmonary nodules have a certain malignant tendency", while the basic semantic feature is "pulmonary nodules tend to be benign", and the two do not match.

[0093] Since the current round number 3 is less than the set neighborhood enhancement range 5, the server first outputs the reference semantic feature "pulmonary nodules have a certain malignant tendency" as the target semantic feature, and outputs the current target neighborhood enhancement region feature as the standardized image feature. Then, the server decides to continue to perform the next round of coordinate neighborhood enhancement on the target interest coordinate index set.

[0094] In the fourth round of coordinate neighborhood enhancement, the server again performs a standardized transformation on the target neighborhood enhancement region features corresponding to the newly generated neighborhood enhancement coordinate index set to generate new reference semantic features. Assuming that the reference semantic feature of this round is "increased malignancy possibility of lung nodules", while the basic semantic feature is still "pulmonary nodules tend to be benign", the two still do not match.

[0095] At this time, because the current round number 4 is still less than the set neighborhood enhancement range 5, the server again outputs the reference semantic feature of "increased possibility of malignancy of lung nodules" as the target semantic feature, and outputs the new target neighborhood enhancement area feature as the standardized image feature, and then continues with the fifth round of coordinate neighborhood enhancement.

[0096] When it comes to the fifth round, the reference semantic feature generated by the server is "highly suspected malignant lung nodules", and the basic semantic feature is still "lung nodules tend to be benign", and the two still do not match.

[0097] Since the current round number 5 is equal to the set neighborhood enhancement range, the server no longer continues to perform coordinate neighborhood enhancement. Instead, it determines the reference semantic feature of "highly suspected malignant pulmonary nodules" as the final target semantic feature, and determines the target neighborhood enhancement regional feature of this round as the final standardized image feature, ending the coordinate neighborhood enhancement of the target interest coordinate index set.

[0098] During the whole process, the server continuously obtains the current round number and decides whether to end or continue coordinate neighborhood enhancement based on its relationship with the neighborhood enhancement range, so as to analyze the lesion characteristics in the lung CT images as comprehensively and accurately as possible, providing a more reliable basis for subsequent diagnosis and treatment.

[0099] Through multiple rounds of neighborhood enhancement and semantic feature analysis, the server can dig deeper into the characteristic information of lung nodules. For example, during the neighborhood enhancement process, the server may discover more detailed features such as density changes in the tissue around the nodule and abnormal blood vessel distribution, which can help to more accurately determine the nature of the nodule.

[0100] At the same time, the target semantic features and standardized image features generated in each round are recorded and stored by the server for subsequent comparison and verification with the doctor's diagnosis results, continuously optimizing the server's analysis model and algorithm, and improving the accuracy and reliability of lung disease diagnosis.

[0101] Therefore, by strictly following the set rules and processes to perform coordinate neighborhood enhancement and semantic feature analysis, strong support is provided for the diagnosis and treatment of lung diseases.

[0102] In a possible implementation, step S142 includes:

[0103] Step S1421, performing feature space deconstruction on the target neighborhood enhanced region feature to generate at least one first deconstructed CT lesion of interest feature.

[0104] Step S1422, performing feature space deconstruction on the semantic features in the set feature knowledge base to generate at least one second deconstructed CT lesion of interest feature.

[0105] Step S1423, acquiring feature attributes of the first deconstructed CT lesion of interest feature and feature attributes of the second deconstructed CT lesion of interest feature.

[0106] Step S1424: performing a morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature according to the feature attribute, and generating a deconstructed feature morphological matching degree.

[0107] Step S1425: performing a semantic comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature according to the feature attribute, and generating a deconstructed feature semantic matching degree.

[0108] Step S1426, the deconstructed feature morphological matching degree and the deconstructed feature semantic matching degree are fused and calculated, and then the inverse is obtained to generate a feature distance between the target neighborhood enhanced region feature and the semantic feature.

[0109] Step S1427: outputting the semantic feature with the lowest feature distance as the reference semantic feature of the target neighborhood enhanced region feature.

[0110] In this embodiment, when processing lung CT image data, the server first obtains the target neighborhood enhancement region features, assuming that this region is a certain range of tissue surrounding a lung nodule and its surroundings.

[0111] The server performs feature space deconstruction on the target neighborhood enhancement area features to generate multiple first deconstructed CT lesion features of interest. For example, the server deconstructs "grayscale mean is 150" and "grayscale standard deviation is 20" from the grayscale value distribution; deconstructs "texture roughness is high" from the texture feature; and deconstructs "nodule shape irregularity is 0.6" from the shape feature.

[0112] At the same time, the server performs feature space deconstruction on the semantic features in the set feature knowledge base to generate multiple second deconstructed CT lesion features of interest. Assume that the semantic features of "benign nodules" in the feature knowledge base are deconstructed into "grayscale value mean between 80-120", "grayscale value standard deviation is small", "texture roughness is low", "nodule shape regularity is high", etc.; the semantic features of "malignant nodules" are deconstructed into "grayscale value mean greater than 150", "grayscale value standard deviation is large", "texture roughness is high", "nodule shape irregularity is high", etc.

[0113] The server obtains the characteristic attributes of the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature. For the gray value mean, its characteristic attribute is a specific value; for the texture roughness, the characteristic attribute may be a level describing the roughness; for the shape irregularity, the characteristic attribute is a value indicating the degree of irregularity.

[0114] Based on these characteristic attributes, the server performs a morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature to generate a deconstructed feature morphological matching degree. The server first determines the morphological matching feature of the first deconstructed CT lesion of interest feature based on a preset morphological matching feature sequence. For example, the preset morphological matching feature sequence stipulates that the gray value mean and shape irregularity are focused on.

[0115] For the gray value mean, assuming that the gray value mean in the first deconstructed CT lesion of interest feature is 150, and the gray value mean range corresponding to the semantic feature of "benign nodule" is 80-120, the first deviation is large; the gray value mean corresponding to the semantic feature of "malignant nodule" is greater than 150, and the first deviation is small. For shape irregularity, assuming that the shape irregularity in the first deconstructed CT lesion of interest feature is 0.6, the shape irregularity corresponding to the semantic feature of "benign nodule" is usually low, and the second deviation is large; the shape irregularity corresponding to the semantic feature of "malignant nodule" is usually high, and the second deviation is small. The server determines the deconstructed feature morphological matching degree based on the inverse of the smallest deviation between the first deviation and the second deviation. In this example, since the deviation from the semantic feature of "malignant nodule" is small, its inverse is determined as the deconstructed feature morphological matching degree.

[0116] Then, based on the feature attributes, the server performs a semantic comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature to generate a semantic matching degree of the deconstructed feature. The server encodes the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature with the same feature attributes to generate corresponding encoding vectors. For example, for the gray value mean, the encoding corresponding to the "benign nodule" semantic feature is [0,1,0], the encoding corresponding to the "malignant nodule" semantic feature is [1,0,0], and the encoding of the current first deconstructed CT lesion of interest feature is [1,0,0]. The server determines the first correlation coefficient between the first encoding vector and the second encoding vector of the "benign nodule" semantic feature, and the second correlation coefficient between the first encoding vector and the second encoding vector of the "malignant nodule" semantic feature. By comparing the two correlation coefficients, assuming that the correlation with the "malignant nodule" semantic feature is higher, the server determines the semantic matching degree of the deconstructed feature based on this.

[0117] Then, the server performs a fusion calculation on the morphological matching degree of the deconstructed features and the semantic matching degree of the deconstructed features. For example, the server uses a weighted average method, giving the morphological matching degree a weight of 0.4 and the semantic matching degree a weight of 0.6 for calculation. After the calculation is completed, the inverse is calculated to generate the feature distance between the target neighborhood enhanced area feature and each semantic feature. Assume that the feature distance with the semantic feature of "benign nodule" is 0.8, and the feature distance with the semantic feature of "malignant nodule" is 0.3.

[0118] Finally, the server outputs the semantic feature of “malignant nodule” corresponding to the lowest feature distance (0.3 in this case) as the reference semantic feature of the target neighborhood enhanced region feature.

[0119] Through such a detailed and comprehensive analysis process, the server can more accurately generate semantic features with reference value for the target neighborhood enhancement regional features, providing strong support for subsequent medical diagnosis and treatment decisions. As more lung CT image data is processed, the server continues to accumulate experience and optimize algorithms, and the accuracy and reliability of the reference semantic features it generates will continue to improve.

[0120] In subsequent processing, the server will match this reference semantic feature with the basic semantic feature and take corresponding measures based on the matching results, such as continuing to enhance the neighborhood or ending the enhancement of the neighborhood, so as to gradually improve the analysis and annotation of the lung CT images.

[0121] At the same time, the server will continue to update and enrich the feature knowledge base, incorporating more types of lung lesion features and corresponding semantic descriptions to improve the ability to analyze various complex lung lesions. For example, adding feature deconstruction and semantic features of lesions such as lung inflammation and pulmonary fibrosis will enable a more comprehensive consideration of various possible lesions when processing lung CT images.

[0122] In addition, the server will evaluate and verify each generated reference semantic feature. If the doctor gives a different diagnosis based on the final annotated CT image data, the server will incorporate this feedback information into the learning process and adjust the methods and parameters of feature deconstruction, matching calculation, and fusion calculation to continuously improve the accuracy and reliability of the generated reference semantic features.

[0123] In a possible implementation, before step S1424, the method further includes: determining the morphological matching features of the first deconstructed CT lesion of interest features according to a preset morphological matching feature sequence.

[0124] Step S1424 includes:

[0125] Step S1424-1, determining a first deviation between the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature having the same characteristic attribute.

[0126] Step S1424-2, determining a second deviation between the morphological matching feature and the second deconstructed CT lesion of interest feature having the same characteristic attribute.

[0127] Step S1424-3, determining the deconstructed feature morphology matching degree according to the inverse of the smallest deviation between the first deviation and the second deviation.

[0128] In this embodiment, in the process of processing lung CT images, the server first determines the morphological matching features of the first deconstructed CT lesion of interest feature based on a set preset morphological matching feature sequence before performing morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature based on feature attributes to generate a deconstructed feature morphological matching degree.

[0129] Assume that the preset morphological matching feature sequence stipulates that for the analysis of lung nodules, the focus is on the three features of gray value mean, nodule shape irregularity, and texture roughness. Then the server determines the parts corresponding to these three key features from the first deconstructed CT lesion of interest features obtained previously as morphological matching features.

[0130] Next, the server performs a morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature based on the feature attributes, and generates a deconstructed feature morphological matching degree.

[0131] For example, for the same feature attribute of gray value mean, the gray value mean in the first deconstructed CT lesion of interest feature is 180, while the gray value mean in the second deconstructed CT lesion of interest feature (assuming it is a benign nodule feature) is usually between 100 and 150. The server calculates the difference between the two and determines the first deviation. Assume that the calculated first deviation is 0.4.

[0132] At the same time, for the gray value mean in the morphological matching feature, assuming that its value is also 180, it is compared with the gray value mean range in the second deconstructed CT lesion of interest feature to determine the second deviation. Assume that the calculated second deviation is 0.5.

[0133] Since the first deviation 0.4 is smaller than the second deviation 0.5, the server selects the reciprocal of the first deviation 0.4, ie, 1 / 0.4=2.5, to determine the deconstructed feature morphology matching degree.

[0134] For another example, for the characteristic attribute of nodule shape irregularity, the shape irregularity in the first deconstructed CT lesion of interest feature is 0.7, while the shape irregularity in the second deconstructed CT lesion of interest feature (assuming it is a malignant nodule feature) is usually greater than 0.6. The server calculates the first deviation to be 0.1.

[0135] For the nodule shape irregularity in the morphological matching feature, assuming its value is 0.75, compared with the shape irregularity range in the second deconstructed CT interest lesion feature, the second deviation is obtained to be 0.2.

[0136] Since the first deviation 0.1 is smaller than the second deviation 0.2, the server selects the reciprocal of the first deviation 0.1, ie, 1 / 0.1=10, as the deconstruction feature morphology matching degree of this part.

[0137] For the feature attribute of texture roughness, assuming that the texture roughness in the first deconstructed CT lesion of interest feature is higher, with a value of 0.8, while the texture roughness in the second deconstructed CT lesion of interest feature (assuming it is an inflammation feature) is usually between 0.2 and 0.5. The server calculates the first deviation to be 0.6.

[0138] For the texture roughness in the morphological matching feature, assuming its value is 0.7, it is compared with the texture roughness range in the second deconstructed CT lesion of interest feature, and the second deviation is obtained to be 0.5.

[0139] Since the second deviation 0.5 is smaller than the first deviation 0.6, the server selects the reciprocal of the second deviation 0.5, ie, 1 / 0.5=2, as the deconstruction feature morphology matching degree of this part.

[0140] By comprehensively considering the deconstructed feature morphological matching of the three characteristic attributes, namely, the gray value mean, nodule shape irregularity, and texture roughness, the server can obtain the final comprehensive deconstructed feature morphological matching through some weighted average or comprehensive evaluation method.

[0141] Through such detailed and precise morphological comparison and deviation calculation, the server can more accurately evaluate the morphological similarities and differences between the first deconstructed CT lesion of interest features and the second deconstructed CT lesion of interest features, providing an important foundation and basis for the subsequent generation of accurate reference semantic features and effective image analysis.

[0142] In actual processing, the server may process a large amount of lung CT image data at the same time, and perform such rigorous morphological comparison and calculation for the target neighborhood enhancement area features in each image. Moreover, with the accumulation of data and optimization of algorithms, the server will become more and more accurate and efficient in determining morphological matching features, calculating deviations, and generating final morphological matching degrees.

[0143] At the same time, the server will also compare and verify the morphological matching results generated each time with the actual medical diagnosis results. If deviations or inaccuracies are found, the server will automatically adjust the selection of morphological matching features, the calculation method of deviation, and the setting of weights, etc., to continuously improve the accuracy and reliability of morphological matching.

[0144] In addition, the server will also take into account the possible differences in lung CT image data collected by different medical institutions and different equipment. Through a large amount of data training and model optimization, its morphological comparison method will have a wider range of applicability and robustness, and can adapt to various complex and diverse lung CT imaging situations.

[0145] In a possible implementation, step S1425 includes:

[0146] Step S1425-1, encode the first deconstructed CT lesion of interest feature, the second deconstructed CT lesion of interest feature and the morphological matching feature with the same characteristic attributes respectively, and generate a first encoding vector of the first deconstructed CT lesion of interest feature, a second encoding vector of the second deconstructed CT lesion of interest feature and a third encoding vector of the morphological matching feature.

[0147] Step S1425-2, determine a first correlation coefficient between the first coding vector and the second coding vector, and determine a second correlation coefficient between the third coding vector and the second coding vector.

[0148] Step S1425-3: Determine the semantic matching degree of the deconstructed feature according to the first correlation coefficient and the second correlation coefficient.

[0149] In this embodiment, when processing lung CT images, the server needs to perform semantic comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature according to feature attributes to generate a semantic matching degree of the deconstructed features.

[0150] First, the server encodes the first deconstructed CT lesion of interest feature, the second deconstructed CT lesion of interest feature and the morphological matching feature respectively, assuming that the same feature attributes include gray value mean, nodule shape irregularity and texture roughness.

[0151] For the characteristic attribute of grayscale value mean, the server encodes the grayscale value mean in the first deconstructed CT lesion of interest feature as the number 50 (assuming that this is a specific encoding method), and generates an element in the first encoding vector; for the grayscale value mean in the second deconstructed CT lesion of interest feature (assuming that it is a specific benign nodule feature), it is encoded as 30, and the corresponding element in the second encoding vector is generated; for the grayscale value mean in the morphological matching feature, it is encoded as 45, and the corresponding element in the third encoding vector is generated.

[0152] For the characteristic attribute of nodule shape irregularity, the irregularity in the first deconstructed CT lesion of interest feature is encoded as 0.7, generating another element in the first encoding vector; the irregularity in the second deconstructed CT lesion of interest feature is encoded as 0.5, generating the corresponding element in the second encoding vector; the irregularity in the morphological matching feature is encoded as 0.6, generating the corresponding element in the third encoding vector.

[0153] For the characteristic attribute of texture roughness, the roughness in the first deconstructed CT lesion of interest feature is encoded as 0.8, generating another element in the first encoding vector; the roughness in the second deconstructed CT lesion of interest feature is encoded as 0.6, generating the corresponding element in the second encoding vector; the roughness in the morphological matching feature is encoded as 0.7, generating the corresponding element in the third encoding vector.

[0154] In this way, the server generates a first coding vector [50, 0.7, 0.8] of the first deconstructed CT lesion of interest feature, a second coding vector [30, 0.5, 0.6] of the second deconstructed CT lesion of interest feature, and a third coding vector [45, 0.6, 0.7] of the morphological matching feature.

[0155] Next, the server determines a first correlation coefficient between the first code vector and the second code vector. The correlation coefficient between the first code vector [50, 0.7, 0.8] and the second code vector [30, 0.5, 0.6] is calculated by a relevant mathematical calculation method (such as the Pearson correlation coefficient), and is assumed to be 0.4.

[0156] At the same time, the server determines a second correlation coefficient between the third code vector and the second code vector. Similarly, using a corresponding calculation method, the correlation coefficient between the third code vector [45, 0.6, 0.7] and the second code vector [30, 0.5, 0.6] is obtained, which is assumed to be 0.5.

[0157] Then, the server determines the semantic matching degree of the deconstructed feature according to the first correlation coefficient 0.4 and the second correlation coefficient 0.5.

[0158] Assume that the rule set by the server is that if the first correlation coefficient is greater than the second correlation coefficient, then the deconstruction feature semantic matching degree is equal to the first correlation coefficient; if the second correlation coefficient is greater than the first correlation coefficient, then the deconstruction feature semantic matching degree is equal to the second correlation coefficient.

[0159] In this example, since the second correlation coefficient 0.5 is greater than the first correlation coefficient 0.4, the server determines that the semantic matching degree of the deconstructed feature is 0.5.

[0160] Through such a process of encoding, correlation coefficient calculation and matching degree determination, the server can quantify the similarity between the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature at the semantic level, providing an important basis for subsequent analysis and judgment.

[0161] In actual processing, the server may be faced with a large amount of complex lung CT image data. For each target neighborhood enhancement area feature in the image, such precise semantic comparison and calculation are required.

[0162] As more data is processed, the server will continuously optimize the encoding method and correlation calculation method to improve the accuracy and reliability of the semantic matching of the deconstructed features.

[0163] At the same time, the server will also combine the semantic matching degree of the generated deconstructed features with other evaluation indicators (such as morphological matching degree, etc.) to comprehensively judge the matching degree between the target neighborhood enhanced area features and the known semantic features, so as to generate reference semantic features more accurately.

[0164] For example, if the morphological match is high but the semantic match is low, the server may further analyze the differences in feature attributes or readjust the deconstruction and encoding methods to improve the overall matching accuracy.

[0165] In addition, the server will also consider the unique semantic features and encoding methods that different types of lung diseases may have. For example, for lung inflammation and lung tumors, the semantic meaning and importance of features such as grayscale value, shape and texture may be different. The server will perform targeted processing and analysis based on the specific disease type.

[0166] In a possible implementation, after step S1427, the method further includes:

[0167] The target neighborhood enhanced region feature and the reference semantic feature corresponding to the target neighborhood enhanced region feature are added to a set semantic feature library.

[0168] Step S142 includes:

[0169] Step S1421, searching for reference semantic features of the target neighborhood enhanced region features from the semantic feature library.

[0170] Step S1422: When the reference semantic feature is not found, regularized semantic mapping is performed on the target neighborhood enhanced region feature to generate the reference semantic feature.

[0171] In the process of processing lung CT images, the server completes the step of outputting the semantic feature with the minimum feature distance as the reference semantic feature of the target neighborhood enhanced region feature.

[0172] For example, after a series of calculations and comparisons, the server determines that the reference semantic feature corresponding to the target neighborhood enhanced region feature of the current lung nodule is "the lung nodule has early signs of malignant transformation". After completing this output, the server will add the target neighborhood enhanced region feature and the corresponding reference semantic feature "the lung nodule has early signs of malignant transformation" to the set semantic feature library. This semantic feature library is like a constantly accumulating and enriching knowledge base for subsequent quick query and reference.

[0173] When a new lung CT image is subsequently processed, when a similar target neighborhood enhancement region feature is encountered again, the server will first search for the corresponding reference semantic feature from this semantic feature library. Assuming that a similar lung nodule neighborhood enhancement region feature also appears in the newly processed lung CT image, the server searches in the semantic feature library.

[0174] If the server successfully finds the matching target neighborhood enhanced regional features, it can directly obtain the corresponding reference semantic features, such as "lung nodules show early signs of malignancy", thereby saving the computing time and resources for re-regularizing the semantic mapping.

[0175] However, if the current target neighborhood enhancement area feature is not found in the semantic feature library, the server will perform a series of operations such as detailed feature space deconstruction of the target neighborhood enhancement area feature, comparison with the semantic features in the feature knowledge base, and calculation of morphological and semantic matching according to the previous regularized semantic mapping process, and finally generate a new reference semantic feature.

[0176] For example, for a new lung nodule target neighborhood enhancement region feature, its grayscale value distribution, texture features, and shape features are different from the features previously added to the library. The server does not find a match in the semantic feature library, so it starts regularized semantic mapping.

[0177] The server first performs feature space deconstruction on the new target neighborhood enhanced region feature, decomposing it into multiple specific first deconstructed CT lesion features of interest, such as "grayscale value average is 160", "texture roughness is high", "shape irregularity is 0.7", etc.

[0178] Then, the server performs the same feature space deconstruction on the semantic features in the set feature knowledge base to generate a plurality of second deconstructed CT lesion features of interest.

[0179] Next, the server obtains the feature attributes of the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature, and performs morphological comparison and semantic comparison based on these feature attributes to calculate the morphological matching degree and the semantic matching degree.

[0180] By fusion calculation and inverse, the feature distance between the target neighborhood enhanced area feature and each semantic feature is generated, and finally the semantic feature with the minimum feature distance is determined, such as "benign changes in lung nodules suspected to be caused by inflammation", and it is used as the new reference semantic feature.

[0181] At the same time, the server will again add the new target neighborhood enhanced regional features and the generated reference semantic features to the semantic feature library, continuously enriching the content of the library so that accurate reference semantic features can be obtained more quickly when processing similar lung CT images in the future.

[0182] In actual applications, as the number of lung CT images processed continues to increase, the semantic feature library will become larger and more complete. The server can use this continuously accumulated knowledge base to improve processing efficiency and accuracy, and provide more timely and reliable support for medical diagnosis.

[0183] For example, in a day's work, the server may process hundreds of lung CT images. For the first few images, it may be necessary to perform a complete regularized semantic mapping to generate reference semantic features and add them to the library. However, as the processing progresses, more and more similar features can directly find corresponding reference semantic features in the library, greatly speeding up the processing speed.

[0184] In addition, the server will regularly organize and optimize the semantic feature library, remove some outdated or inaccurate records, and ensure that the information in the library always has high quality and reference value.

[0185] In addition, if the medical field's understanding and diagnostic standards of lung diseases change, the server will be able to update the feature knowledge base and semantic feature library accordingly to adapt to new medical needs and technological developments.

[0186] In summary, by adding the enhanced regional features of the target neighborhood and their reference semantic features to the semantic feature library and making rational use of the information in the library, the server can continuously improve the processing capabilities and service quality of lung CT images, and play a more important role in the diagnosis and treatment of lung diseases.

[0187] In a possible implementation, before step S142, the method further includes:

[0188] Get the set cleaning feature sequence.

[0189] When the target neighborhood enhancement region feature belongs to the cleaning feature in the cleaning feature sequence, the basic semantic feature is output as the target semantic feature, the target CT lesion of interest feature is output as the standardized image feature, and the current coordinate neighborhood enhancement step for the target interest coordinate index set is skipped.

[0190] Before regularizing semantic mapping of the target neighborhood enhancement region features in the lung CT image, the server first obtains a set cleaning feature sequence. This cleaning feature sequence is set in advance based on experience and professional knowledge, and contains some specific features that are considered not to require complex processing and enhancement.

[0191] Assume that the cleaning feature sequence includes the feature of "micro calcification points on the edge of the lung". When the server analyzes the current lung CT image, it finds that the target neighborhood enhancement region feature belongs to the cleaning feature of "micro calcification points on the edge of the lung".

[0192] In this case, the server will directly output the previously obtained basic semantic features as target semantic features. For example, the basic semantic features may be "common benign calcification manifestations at the edge of the lungs". At the same time, the server will output the original target CT lesion of interest features as standardized image features.

[0193] In addition, the server will skip the coordinate neighborhood enhancement step for the current target interest coordinate index set. This is because the features in the cleaned feature sequence are considered to have clear diagnostic significance or do not require further enhancement analysis.

[0194] For example, when processing a lung CT image, the server analyzes a specific target area and determines that its feature is a tiny calcification point on the edge of the lung. After comparing it with the cleaning feature sequence, the server confirms that this is a cleaning feature.

[0195] Therefore, the server no longer performs coordinate neighborhood enhancement on this area, avoiding unnecessary consumption of computing resources and time. It directly outputs "common benign calcification manifestations at the edge of the lung" as the target semantic feature, and the originally extracted features of this calcification point area, such as the size and density of the calcification point, are output as the target CT lesion of interest features as standardized image features.

[0196] In another scenario, suppose the cleaning feature sequence also contains the feature of "small lung nodules of a specific size and regular shape". When the server encounters a target neighborhood enhancement region feature that meets this feature during analysis, the above operation will also be taken.

[0197] For example, the server detects a small lung nodule with a diameter of about 3 mm and a regular shape, which matches the description in the cleaned feature sequence. The server then outputs the corresponding basic semantic features, such as "benign and stable small lung nodule" as the target semantic feature, and uses the detailed features of this small nodule as the standardized image feature, while skipping the coordinate neighborhood enhancement step.

[0198] By obtaining a cleaning feature sequence and taking a corresponding concise processing approach when the target neighborhood enhancement area feature belongs to a cleaning feature, the server can improve processing efficiency and provide doctors with valuable diagnostic information more quickly.

[0199] In actual medical applications, such a mechanism enables the server to process large amounts of lung CT image data more intelligently and efficiently. For example, in a day's work, the server may need to process hundreds of lung CT images. By screening the cleaning feature sequence, the server can quickly process areas with clear features that do not require complex analysis, allocating more computing resources and time to key areas that require in-depth analysis and enhancement.

[0200] In addition, the cleaning feature sequence is not static. As medical research progresses and new diagnostic criteria emerge, the server can update and optimize the cleaning feature sequence based on the latest expertise and clinical experience. For example, if new research finds that a certain type of lung lesion no longer requires complex enhanced analysis, it can be added to the cleaning feature sequence.

[0201] At the same time, the server will also record the status of each cleaning feature processing for subsequent data analysis and evaluation. For example, statistics on the frequency of cleaning features, their distribution in different types of patients, etc., provide data support for further optimization of diagnostic processes and algorithms.

[0202] In short, by acquiring and applying the cleaning feature sequence, the server can process lung CT image data more efficiently while ensuring diagnostic accuracy, providing stronger support for medical diagnosis.

[0203] In a possible implementation, step S150 includes:

[0204] Step S151, determine a first setting template for embedding the target semantic feature, embed the target semantic feature into the target CT image data based on the first setting template, generate first temporary image data, and perform CT enhancement tracking and annotation on the standardized image features in the first temporary image data and the target semantic features corresponding to the standardized image features through a set first CT enhancement tracking and annotation strategy to generate the annotated CT image data.

[0205] Alternatively, in step S152, the standardized image features in the target CT image data are replaced with the target semantic features corresponding to the standardized image features to generate second temporary image data, and the target semantic features in the second temporary image data are CT enhanced tracking and labeled by using a set second CT enhanced tracking and labeling strategy to generate the labeled CT image data.

[0206] In this embodiment, when processing lung CT image data, the server performs feature annotation based on previously generated standardized image features and target semantic features to generate annotated CT image data.

[0207] First, the server determines a first setting template for embedding the target semantic feature. Assume that the first setting template stipulates that the target semantic feature is embedded in the image with a specific font, color and position. For example, the font is selected as Kaiti, the color is red, and the position is in the upper right corner of the image.

[0208] Based on this first set template, the server embeds the target semantic feature into the target CT image data to generate the first temporary image data. For example, the target semantic feature is "pulmonary nodules, suspected early malignant lesions", and the server displays this feature in red regular font in the upper right corner of the image according to the template to generate the first temporary image data.

[0209] Then, the server performs CT enhanced tracking and annotation on the standardized image features and the corresponding target semantic features in the first temporary image data through the set first CT enhanced tracking and annotation strategy. Assume that the first CT enhanced tracking and annotation strategy includes drawing a clear yellow border around the nodule area and annotating the specific values ​​of the standardized image features such as the size and irregularity of the nodule on the border. The server operates according to this strategy to generate the final annotated CT image data.

[0210] In another case, the server replaces the standardized image features in the target CT image data with the corresponding target semantic features to generate the second temporary image data. For example, some detailed grayscale values ​​and texture descriptions of nodules in the original standardized image features are replaced with the target semantic feature of "pulmonary nodules, suspected early malignant lesions".

[0211] Next, the server performs CT enhanced tracking and annotation on the target semantic features in the second temporary image data through the set second CT enhanced tracking and annotation strategy. Assume that the second CT enhanced tracking and annotation strategy is to add a green arrow next to the target semantic feature, and explain in detail the diagnostic basis and related suggestions of the semantic feature at the location pointed by the arrow, such as "Based on the morphology and growth rate of the nodule, further pathological biopsy is recommended." The server executes this strategy to generate annotated CT image data.

[0212] For example, when processing a lung CT image, the server first determines the target semantic feature as "there is an inflammation area in the lung, which requires close observation". According to the first set template, this feature is displayed in blue Song font in the upper left corner of the image to generate the first temporary image data. Then, through the first CT enhanced tracking and annotation strategy, the inflammation area is outlined with orange lines, and the range and density of the inflammation area are annotated.

[0213] Alternatively, for another lung CT image, the server replaces the complex description of lung texture changes in the standardized image features with the target semantic feature "abnormal lung texture, possibly related to smoking" to generate the second temporary image data. Then, based on the second CT enhanced tracking and annotation strategy, a purple annotation box is added below the target semantic feature to explain in detail the manifestation of abnormal texture and the possible impact of smoking.

[0214] In actual work, the server may flexibly choose to use the first or second annotation method according to different needs and image characteristics. For example, if the lesion features in the image are relatively complex and need to display both detailed standardized image features and concise and clear target semantic features, the first method will be selected; if the lesion features are relatively simple and clear, and more emphasis is placed on highlighting semantic features, the second method will be selected.

[0215] At the same time, the server will continue to optimize annotation strategies and templates to improve the clarity and readability of annotations. For example, according to doctors' feedback and clinical practice requirements, the font size, color contrast, and annotation position are adjusted so that the annotated CT image data can more effectively assist doctors in making diagnosis and treatment decisions.

[0216] In addition, the server will also consider the specific requirements of different medical institutions and departments and customize various annotation templates and strategies. For example, some hospitals may pay more attention to the simplicity of annotation, while some research institutions may require more detailed and accurate annotation information.

[0217] In summary, by determining the template, embedding or replacing features, and applying corresponding annotation strategies, the server can provide accurate, clear, and targeted annotations for lung CT images, bringing convenience and support to medical work.

[0218] Figure 2 A smart medical system 100 provided in an embodiment of the present application is shown, including a processor 1001, a memory 1003 and a program code stored in the memory 1003, and the processor 1001 executes the above program code to implement the steps of the CT image data analysis method based on artificial intelligence.

[0219] Figure 2 The smart medical system 100 shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the smart medical system 100 may also include a transceiver 1004, which may be used for data interaction between the smart medical system and other smart medical systems, such as data transmission and / or data reception. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the smart medical system 100 does not constitute a limitation on the embodiments of the present application.

[0220] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0221] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc.

[0222] The memory 1003 can be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to have or store program code and can be read by a computer, without limitation herein.

[0223] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0224] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0225] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders based on demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenario, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0226] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A CT image data analysis method based on artificial intelligence, characterized in that: The method comprises: Based on the CT image data analysis instruction, the target CT image data is encoded with a region of interest, and features of the target CT lesion of interest are generated; Based on the coordinates of each pixel point in the target CT image data in the target CT image data, a coordinate index is assigned to all pixel points in the target CT image data to generate a target interest coordinate index set corresponding to the target CT interest lesion feature; According to the set neighborhood enhancement range, performing at least one round of coordinate neighborhood enhancement on the target interest coordinate index set; After generating a neighborhood enhancement coordinate index set in each round of coordinate neighborhood enhancement, a target neighborhood enhancement region feature corresponding to the neighborhood enhancement coordinate index set is subjected to a standardized transformation to generate a standardized image feature including the target CT lesion feature of interest and a target semantic feature corresponding to the standardized image feature; the standardized image feature includes: a knowledge point lesion feature corresponding to at least one knowledge point in a lesion diagnosis knowledge network under a target lesion diagnosis label; The target CT image data is feature-labeled according to the standardized image features and the target semantic features to generate labeled CT image data.

2. The method for analyzing CT image data based on artificial intelligence according to claim 1, characterized in that: The step of performing a standardized transformation on the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set to generate a standardized image feature including the target CT lesion of interest features and a target semantic feature corresponding to the standardized image feature includes: Extracting the target neighborhood enhancement region features corresponding to the neighborhood enhancement coordinate index set from the target CT image data; Performing regularized semantic mapping on the target neighborhood enhanced region features to generate reference semantic features of the target neighborhood enhanced region features; When the reference semantic feature matches the basic semantic feature, the basic semantic feature is output as the target semantic feature, and the target CT lesion of interest feature is output as the standardized image feature; wherein the basic semantic feature is a semantic feature generated by performing the regularized semantic mapping on the target CT lesion of interest feature; When the reference semantic feature does not match the basic semantic feature, the reference semantic feature is output as the target semantic feature, and the target neighborhood enhanced region feature is output as the standardized image feature.

3. The artificial intelligence-based CT image data analysis method according to claim 2, characterized in that: When the reference semantic feature does not match the basic semantic feature, outputting the reference semantic feature as the target semantic feature, and outputting the target neighborhood enhanced region feature as the standardized image feature, comprises: Obtaining the current round number of the coordinate neighborhood enhancement; When the reference semantic feature does not match the basic semantic feature and the current round number is less than the neighborhood enhancement range, the reference semantic feature is output as the target semantic feature, and the target neighborhood enhancement area feature is output as the standardized image feature, and the coordinate neighborhood enhancement of the target interest coordinate index set is terminated; Alternatively, when the reference semantic feature does not match the basic semantic feature and the current round number is less than the neighborhood enhancement range, the reference semantic feature is output as the target semantic feature, and the target neighborhood enhancement area feature is output as the standardized image feature, and the next round of coordinate neighborhood enhancement of the target interest coordinate index set is continued until the current round number is equal to the neighborhood enhancement range, at which time the coordinate neighborhood enhancement of the target interest coordinate index set is terminated.

4. The method for analyzing CT image data based on artificial intelligence according to claim 2, characterized in that: The step of performing regularized semantic mapping on the target neighborhood enhanced region feature to generate a reference semantic feature of the target neighborhood enhanced region feature includes: Performing feature space deconstruction on the target neighborhood enhanced region feature to generate at least one first deconstructed CT lesion of interest feature; Deconstructing the semantic features in the set feature knowledge base in feature space to generate at least one second deconstructed CT lesion of interest feature; Acquire a feature attribute of the first deconstructed CT lesion of interest feature and a feature attribute of the second deconstructed CT lesion of interest feature; According to the feature attribute, performing morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature to generate a deconstructed feature morphological matching degree; According to the feature attribute, semantically comparing the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature to generate a deconstructed feature semantic matching degree; The deconstructed feature morphological matching degree and the deconstructed feature semantic matching degree are fused and calculated, and then the inverse is obtained to generate a feature distance between the target neighborhood enhanced region feature and the semantic feature; The semantic feature with the lowest feature distance is output as the reference semantic feature of the target neighborhood enhanced region feature.

5. The artificial intelligence-based CT image data analysis method according to claim 4, characterized in that: Before performing morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature according to the feature attribute to generate a deconstructed feature morphological matching degree, the method further includes: Determining the morphological matching features of the first deconstructed CT lesion of interest features according to a preset morphological matching feature sequence; The step of performing morphological comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature according to the feature attribute to generate a deconstructed feature morphological matching degree includes: Determining a first deviation between the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature having the same characteristic attribute; Determining a second deviation between the morphological matching feature and the second deconstructed CT lesion of interest feature having the same characteristic attribute; The deconstructed feature morphology matching degree is determined according to the inverse of the smallest deviation between the first deviation and the second deviation.

6. The artificial intelligence-based CT image data analysis method according to claim 5, characterized in that: The step of performing semantic comparison on the first deconstructed CT lesion of interest feature and the second deconstructed CT lesion of interest feature according to the feature attribute to generate a deconstructed feature semantic matching degree includes: Encoding the first deconstructed CT lesion of interest feature, the second deconstructed CT lesion of interest feature, and the morphological matching feature respectively, which have the same characteristic attributes, to generate a first encoding vector of the first deconstructed CT lesion of interest feature, a second encoding vector of the second deconstructed CT lesion of interest feature, and a third encoding vector of the morphological matching feature; Determine a first correlation coefficient between the first code vector and the second code vector, and determine a second correlation coefficient between the third code vector and the second code vector; The semantic matching degree of the deconstructed feature is determined according to the first correlation coefficient and the second correlation coefficient.

7. The method for analyzing CT image data based on artificial intelligence according to claim 4, characterized in that: After outputting the semantic feature with the lowest feature distance as the reference semantic feature of the target neighborhood enhanced region feature, the method further includes: Adding the target neighborhood enhanced region feature and the reference semantic feature corresponding to the target neighborhood enhanced region feature to a set semantic feature library; The step of performing regularized semantic mapping on the target neighborhood enhanced region feature to generate a reference semantic feature of the target neighborhood enhanced region feature includes: Searching for reference semantic features of the target neighborhood enhanced regional features from the semantic feature library; When the reference semantic feature is not found, regularized semantic mapping is performed on the target neighborhood enhanced region feature to generate the reference semantic feature.

8. The method for analyzing CT image data based on artificial intelligence according to claim 2, characterized in that: Before performing regularized semantic mapping on the target neighborhood enhanced region feature to generate a reference semantic feature of the target neighborhood enhanced region feature, the method further includes: Obtain the set cleaning feature sequence; When the target neighborhood enhancement region feature belongs to the cleaning feature in the cleaning feature sequence, the basic semantic feature is output as the target semantic feature, the target CT lesion of interest feature is output as the standardized image feature, and the current coordinate neighborhood enhancement step for the target interest coordinate index set is skipped.

9. The method for analyzing CT image data based on artificial intelligence according to any one of claims 1 to 8, characterized in that: The step of performing feature annotation on the target CT image data according to the standardized image features and the target semantic features to generate annotated CT image data includes: Determining a first setting template for embedding the target semantic feature; embedding the target semantic feature into the target CT image data based on the first setting template to generate first temporary image data; Performing CT enhanced tracking and annotation on the standardized image features in the first temporary image data and the target semantic features corresponding to the standardized image features through a set first CT enhanced tracking and annotation strategy to generate the annotated CT image data; Alternatively, the standardized image feature in the target CT image data is replaced with the target semantic feature corresponding to the standardized image feature to generate second temporary image data; By setting a second CT enhanced tracking and annotation strategy, CT enhanced tracking and annotation are performed on the target semantic features in the second temporary image data to generate the annotated CT image data.

10. A smart medical system, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the artificial intelligence-based CT image data analysis method described in any one of claims 1 to 9 is implemented.

Images