Deep learning-based gallbladder adipoma image analysis method and deep learning-based gallbladder adipoma image analysis system

Abstract

The invention relates to a method and a system for analyzing a cholestenoma image based on deep learning, and provides an efficient image processing mode for cholestenoma, which comprises the steps of quantum feature screening of image data, topological segmentation, generative adversarial network enhancement, graph convolutional network spatial correlation analysis, fuzzy logic classification and the like. Through quantum state conversion and feature screening, contrast and resolution of image data are significantly improved, noise is removed, and key features are extracted. Topological data segmentation is combined with multi-scale processing, and the focus area boundary is effectively recognized. The generative adversarial network is used for enhancing lesion feature data, and the graph convolutional network is used for further analyzing spatial correlation features to ensure that description of a lesion region is more refined. And finally, the existence probability and the classification result of the focus are obtained through fuzzy logic classification. The method aims at improving the information extraction density of the bile adipoma image.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a cholesteatoma image analysis method and system based on deep learning. Background Art

[0002] Cholesteatoma is a complex complication of chronic otitis media, characterized by abnormal proliferation of keratinized epithelium in the middle ear, which can lead to erosion and destruction of surrounding bone, and in severe cases may cause hearing loss or even facial nerve paralysis. Therefore, timely and accurate diagnosis and evaluation of cholesteatoma is crucial. Currently, medical imaging such as computed tomography (CT) and magnetic resonance imaging (MRI) are widely used for the detection and evaluation of cholesteatoma. However, the interpretation of these imaging data is highly dependent on the physician's experience level, and there are problems of strong subjectivity and prone to misdiagnosis and missed diagnosis, especially in the identification of small lesions in the early stage of cholesteatoma. How to use intelligent technology to improve the accuracy and efficiency of image analysis has become one of the urgent problems to be solved in the field of cholesteatoma research.

[0003] In terms of image analysis, deep learning has made significant progress in recent years. Its automatic feature extraction and efficient classification capabilities have attracted much attention in the application of medical images. For cholesteatoma image analysis, traditional manual segmentation and feature engineering methods are not only time-consuming and labor-intensive, but also often fail to achieve satisfactory results when dealing with individual differences in images of different patients. Deep learning, especially technology based on convolutional neural networks (CNNs), has the ability to automatically extract multi-level features and can significantly improve the detection accuracy of cholesteatoma. However, there are still some shortcomings in current research. For example, the training process of the model requires a large amount of labeled data, while the labeling of medical images is often time-consuming and requires expert participation, and the lack of data has caused certain limitations on the generalization ability of the model. In addition, the "black box" characteristics of deep learning models make their decision-making process difficult to explain, limiting their trust and promotion in clinical applications.

[0004] In the existing image analysis systems, the development of dedicated deep learning models for cholesteatoma is still immature. Most current studies focus on the identification of general otitis media lesions, while there is a lack of systematic research on the detailed analysis of specific lesions such as cholesteatoma. At the same time, due to the diversity of cholesteatoma morphology and imaging features, the recognition accuracy of existing models may be reduced when facing different types of cholesteatoma, especially cases with blurred boundaries and complex morphology are more likely to be misjudged by the model. Therefore, how to develop a more targeted deep learning model that can process image data from different imaging devices and have good generalization while ensuring high accuracy is still a key issue that needs to be addressed in current cholesteatoma image analysis research. Summary of the invention

[0005] The main purpose of the present invention is to provide a cholesteatoma image analysis method and system based on deep learning to solve the technical problems of insufficient extraction of cholesteatoma lesion features, difficulty in identifying early small lesions, serious noise interference during processing, and inaccurate description of complex lesion spatial structure in existing medical image analysis. In addition, the traditional method has insufficient generalization ability for image processing of different types of cholesteatoma and cannot effectively handle the situation of blurred lesion boundaries and complex morphology, which further limits the reliability of the analysis and the application effect.

[0006] To achieve the above-mentioned purpose, the present invention provides a cholesteatoma image analysis method based on deep learning, comprising the following steps: acquiring the image data of a target patient, and performing quantum feature screening processing on the image data of the target patient to obtain image feature data after preliminary screening; performing topological data segmentation processing on the image feature data after preliminary screening to obtain topological feature data of the lesion area; performing generative adversarial network enhancement processing on the topological feature data of the lesion area to obtain enhanced lesion feature data; performing graph convolutional network spatial association analysis processing on the enhanced lesion feature data to obtain spatial association features of the lesion; performing adaptive fuzzy logic classification processing on the spatial association features of the lesion to obtain the probability value of the image data matching the cholesteatoma sample image.

[0007] The present invention provides a cholesteatoma image analysis system based on deep learning, comprising: an acquisition unit, used to acquire image data of a target patient, and perform quantum feature screening processing on the image data of the target patient to obtain image feature data after preliminary screening; a segmentation unit, used to perform topological data segmentation processing on the image feature data after preliminary screening to obtain topological feature data of a lesion area; an enhancement unit, used to perform generative adversarial network enhancement processing on the topological feature data of the lesion area to obtain enhanced lesion feature data; an association analysis unit, used to perform graph convolutional network spatial association analysis processing on the enhanced lesion feature data to obtain spatial association features of the lesion; and a classification unit, used to perform adaptive fuzzy logic classification processing on the spatial association features of the lesion to obtain a probability value that the image data matches a cholesteatoma sample image.

[0008] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0009] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0010] The cholesteatoma image analysis method and system based on deep learning provided by the present invention successfully realizes multi-level feature extraction, refined segmentation and enhanced processing of cholesteatoma lesions through quantum feature screening, topological data segmentation, generative adversarial network enhancement, graph convolution network spatial association analysis, adaptive fuzzy logic classification and other steps in the deep learning model, making the extraction of lesion features more comprehensive and accurate. This method significantly improves the recognition accuracy of complex lesion areas, especially enhances the recognition ability of early small lesions, and at the same time has stronger robustness and generalization in the processing of multiple types of lesions, and ultimately improves the accuracy of image analysis and classification reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is a schematic diagram of a cholesteatoma image analysis method based on deep learning in one embodiment of the present invention; Figure 2 is a structural diagram of a cholesteatoma image analysis device based on deep learning in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0012] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0014] Reference Figure 1 The embodiment of the present invention provides a cholesteatoma image analysis method based on deep learning, comprising the following steps: S1, acquiring the image data of a target patient, and performing quantum feature screening processing on the image data of the target patient to obtain image feature data after preliminary screening.

[0015] S2, performing topological data segmentation processing on the image feature data after preliminary screening to obtain topological feature data of the lesion area.

[0016] S3, performing generative adversarial network enhancement processing on the topological feature data of the lesion area to obtain enhanced lesion feature data.

[0017] S4, performing graph convolutional network spatial correlation analysis on the enhanced lesion feature data to obtain spatial correlation features of the lesion.

[0018] S5, performing adaptive fuzzy logic classification processing on the spatial correlation features of the lesion to obtain a probability value of the image data matching the cholesteatoma sample image.

[0019] In order to improve the accuracy and efficiency of cholesteatoma image analysis, an embodiment of the present invention provides a cholesteatoma image analysis method based on deep learning. First, by acquiring the image data of the target patient and performing quantum feature screening processing on it, the contrast and resolution of the image are enhanced while removing random noise. This can more clearly highlight the detailed features of the lesion area, which is helpful for subsequent feature extraction. For example, through quantum state conversion, the edges of the lesion area in the image can be enhanced, making these features more obvious and easy to distinguish in subsequent processing.

[0020] Next, the image feature data after preliminary screening needs to be further processed by topological data segmentation to accurately identify the topological features of the lesion area. Through topological structure analysis, the boundary features of the lesion in the image can be identified, and combined with the topological segmentation algorithm based on region growth, the lesion area can be effectively separated from the normal tissue. At this time, the interference of irrelevant tissues can be reduced to ensure that the extracted lesion features are more accurate, especially when the lesion boundary is blurred or the lesion morphology is more complex. This method can significantly improve the segmentation accuracy.

[0021] Then, the topological feature data of the lesion area is enhanced by the generative adversarial network. The features of the lesion area are enhanced by the generator network of the generative adversarial network (GAN), such as strengthening the detailed features of the lesion edge and texture. The discriminator network is used to verify whether the enhanced data has been reasonably enhanced while retaining the original features, to ensure that the enhanced feature data still conforms to the characteristics of the real lesion. At this time, the features of the lesion area are made more significant, improving the accuracy of subsequent classification and recognition. For example, in small lesions or lesions that are difficult to distinguish, GAN enhancement can effectively improve the sensitivity of model recognition.

[0022] The enhanced lesion feature data needs to be processed by spatial correlation analysis of the graph convolutional network. The graph convolutional network (GCN) can describe the spatial structure of the lesion area by extracting the spatial distribution of feature points and constructing the proximity relationship between feature points. By performing convolution operations on the nodes, the spatial correlation information between the feature points in the lesion area can be gradually obtained. At this time, through the spatial correlation of features, the complex morphological structure of the lesion can be more comprehensively understood, especially the distribution of the lesion in three-dimensional space, which is very important for improving the accuracy of classification.

[0023] Finally, the spatial correlation features of the lesions are processed by adaptive fuzzy logic classification to obtain the probability value of the image data matching the cholesteatoma sample image. In this process, fuzzy logic is used to calculate the membership of each feature, and fuzzy reasoning is performed based on these memberships to achieve the classification of lesions. Compared with traditional hard classification, adaptive fuzzy logic classification can more flexibly handle lesion features with unclear boundaries and reduce the possibility of misjudgment. For example, for some lesions that may be between normal and abnormal in morphology, fuzzy logic classification can provide a more continuous probability value to assist doctors in judging the risk level of the lesion, rather than a simple binary judgment. This flexible classification method can better adapt to the complexity of different lesion features, thereby improving the accuracy of the overall analysis.

[0024] In one example, the image data of the target patient is subjected to quantum feature screening processing to obtain image feature data after preliminary screening, including: enhancing the contrast and resolution of the image data of the target patient by quantum state conversion to highlight the detailed features of the lesion area to obtain quantum enhanced image data; using quantum state measurement combined with the minimum mean square error criterion to remove random noise in the quantum enhanced image data, retaining high confidence features related to the lesion area, and obtaining denoised quantum feature image data; applying a quantum Hamiltonian feature extraction algorithm to map the denoised quantum feature image data to a high-dimensional feature space, and extracting the features most sensitive to the lesion area by analyzing the quantum state evolution and assigning feature weights to obtain preliminary image feature data; using a multi-layer quantum neural network to perform layer-by-layer feature interaction analysis, identifying the nonlinear correlation between the preliminary image feature data through coherent operations of quantum states, and obtaining multidimensional feature interaction image data reflecting the complex spatial relationship of lesions; based on the feature contribution ranking, the correlation of each feature in the multidimensional feature interaction image data is quantitatively evaluated to screen out the features most relevant to the lesion classification, and obtain preliminary image feature data after preliminary screening.

[0025] In this example, the purpose of quantum feature screening of the target patient's imaging data is to maximize the accuracy of image analysis and the ability to identify the lesion area. First, the contrast and resolution of the image data are enhanced through quantum state conversion, which is intended to highlight the detailed features of the lesion area. The advantage of this is that it can significantly improve the contrast between the lesion area and the surrounding normal tissue, making the lesion area easier to identify and extract in subsequent processing. Specifically, quantum state conversion is equivalent to a special mathematical transformation of the image data, similar to image enhancement processing, but quantum technology can operate each pixel more finely, thereby obtaining higher quality quantum enhanced imaging data.

[0026] Next, quantum state measurement combined with the minimum mean square error criterion is used to remove random noise from quantum enhanced imaging data. This denoising method combined with quantum state measurement can retain high-confidence features related to the lesion area by minimizing the mean square error, while removing irrelevant random noise as much as possible. The advantage of this is that it can reduce the interference of noise on subsequent processing steps while retaining key lesion information, especially in imaging data with low signal-to-noise ratio. For example, for scattering noise or noise caused by imaging equipment that may exist in the image, quantum state measurement combined with the minimum mean square error criterion can accurately remove these noises and obtain clearer lesion features.

[0027] In order to further improve the expressive power of image features, the quantum Hamiltonian feature extraction algorithm is applied to map the denoised quantum feature image data into a high-dimensional feature space. The advantage of quantum Hamiltonian feature extraction is that it can extract the features that are most sensitive to the lesion area by analyzing the evolution of the quantum state and assign weights to these features. For example, for certain morphological features or texture features, the quantum Hamiltonian algorithm can analyze the dynamic evolution of these features in a high-dimensional space, extract the most important features for lesion identification, and make subsequent classification and diagnosis more accurate.

[0028] The use of a multi-layer quantum neural network to perform layer-by-layer feature interaction analysis on the preliminary data of image features is mainly aimed at identifying nonlinear correlations between features through coherent operations of quantum states. This can effectively capture the complex spatial relationships in the image data, especially when there are coupling or nonlinear relationships between different features, which may be important diagnostic bases for the disease. Through layer-by-layer quantum state operations, the multi-layer quantum neural network can extract multi-dimensional feature interaction image data, reflecting the complex spatial structure and changes in the lesion area. For example, if there is a significant nonlinear correlation between certain texture features and morphological features in the image, this quantum neural network analysis can extract this implicit relationship and use it to more accurately determine the nature of the lesion.

[0029] Finally, based on the feature contribution ranking, each feature in the multidimensional feature interaction image data is quantitatively evaluated to screen out the features that are most relevant to lesion classification. Through this screening process, features that are irrelevant to classification or have low relevance can be eliminated, thereby reducing data redundancy and improving the efficiency and accuracy of the model. The use of feature contribution ranking ensures that the model only uses those features that have a significant effect on lesion classification. For example, some features have a high degree of differentiation in the classification results of different lesions, while other features may be irrelevant to changes in normal tissue. Through screening, the retained features can maximize the diagnostic accuracy of the model and reduce computational complexity, thereby achieving more efficient image data processing.

[0030] In one example, topological data segmentation processing is performed on the image feature data after preliminary screening to obtain topological feature data of the lesion area, including: performing topological structure analysis on the image feature data after preliminary screening to identify boundary features of the lesion area to obtain preliminary topological data of the lesion; performing multi-scale segmentation processing on the preliminary topological data of the lesion, using a topological algorithm based on region growth to separate the lesion area from surrounding normal tissue to obtain multi-scale topological feature data of the lesion area; performing topological feature extraction on the multi-scale topological feature data of the lesion area to extract key structural features in the lesion area, including morphological features and texture features, to obtain topological feature description data of the lesion; performing topological consistency verification on the topological feature description data of the lesion, and determining the topological feature description data that passes the topological consistency verification as the final topological feature data of the lesion area.

[0031] It should be noted that the purpose of topological data segmentation is to accurately identify the boundaries and internal structures of the lesion area so that this information can be used more effectively in subsequent image analysis. First, the image feature data after preliminary screening is subjected to topological structure analysis to identify the boundary features of the lesion area. Topological structure analysis is used to clarify the geometric morphology and boundary information of the lesion, especially for lesions with blurred boundaries. Topological structure analysis can help extract its possible shape contour and obtain preliminary topological data of the lesion. These preliminary boundary features provide basic information for subsequent processing and help to accurately distinguish lesions from normal tissues.

[0032] After obtaining the preliminary topological data of the lesion, multi-scale segmentation processing is further performed. Multi-scale processing means analyzing the image at different spatial scales to ensure that the segmentation result has both a global perspective and can be finely processed in the local area. The lesion area can be separated from the surrounding normal tissue by using a topological algorithm based on region growing. The region growing algorithm is an algorithm that determines the lesion area by expanding from an initial seed point, and is particularly suitable for identifying lesion areas with strong connectivity. For example, the center of the lesion is selected as the seed point in the image, and the topological algorithm is gradually expanded until the lesion boundary, thereby effectively separating the lesion area from the normal tissue. The advantage of multi-scale segmentation is that it can adapt to the size and complexity of the lesion, so that lesions of different sizes and blurred boundaries can be segmented with high accuracy.

[0033] For the multi-scale topological feature data of the lesion area, the next step is to extract the topological features, with the aim of extracting the key structural features in the lesion area. Topological feature extraction mainly includes the extraction of morphological features and texture features. Morphological features describe the geometric shape of the lesion, such as area, perimeter, contour complexity, etc., while texture features describe the image grayscale distribution pattern in the lesion area, such as roughness, uniformity, etc. These features provide detailed information for subsequent classification and recognition, and can effectively distinguish different types of lesions. For example, the morphological features of cholesteatoma may appear as an irregular shape, while its texture features may appear as a specific texture pattern. This information can be used to improve the model's ability to distinguish lesions.

[0034] Finally, the topological feature description data of the lesion is verified for topological consistency to ensure that the extracted topological features are reasonable in space and consistent with the actual situation. The purpose of topological consistency verification is to ensure that the boundaries and internal features of the lesion are logically coherent and there are no obvious structural errors, such as breaks or overlaps. Through this verification, abnormal data caused by noise or algorithm errors during the extraction process can be filtered out, thereby obtaining more reliable topological feature description data. The data after topological consistency verification is determined as the final topological feature data of the lesion area, which provides a solid foundation for subsequent image classification and lesion evaluation, ensuring the accuracy and consistency of the analysis process.

[0035] In one example, the topological feature data of the lesion area is subjected to generative adversarial network enhancement processing to obtain enhanced lesion feature data, including: inputting the topological feature data of the lesion area into a preset generative adversarial network, and performing feature enhancement on the lesion area through the generator network to obtain enhanced lesion generated feature data; performing authenticity verification on the enhanced lesion generated feature data through a discriminator network to ensure that the enhanced feature data achieves an enhancement effect on the basis of retaining the original features of the lesion area, thereby obtaining verified lesion feature data; performing secondary enhancement processing on the edge and texture features of the lesion area on the verified lesion feature data to obtain final enhanced lesion feature data.

[0036] Generative adversarial network (GAN) enhancement processing is used to improve the expression effect of the features of the lesion area, making the features of the lesion clearer and more significant, so as to facilitate subsequent analysis and classification. First, the topological feature data of the lesion area is input into the preset generative adversarial network. The generative adversarial network consists of two parts: the generator network and the discriminator network. The function of the generator network is to enhance the features of the lesion area, aiming to generate clearer and more detailed feature data. Through the operation of the generator network, the edges, textures and other features of the lesion can be made more prominent. For example, for some fuzzy boundaries or unclear lesion areas, the generator can reasonably enhance these fuzzy parts by learning the feature patterns of other similar data, so that the features of the lesion are more obvious.

[0037] The enhanced lesion generated feature data needs to be verified by the discriminator network for authenticity. The role of the discriminator network is to determine whether the enhanced features generated by the generator retain the true characteristics of the original lesion. This is to ensure that the enhancement process does not over-modify the lesion features or introduce pseudo features, thereby affecting the subsequent diagnosis results. Through the verification of the discriminator, it can be ensured that the enhanced feature data has been improved in detail while still retaining the key features of the original lesion. For example, when enhancing the edges of the lesion, the discriminator will determine whether these edges conform to the true morphology of the lesion, thereby ensuring that the enhanced edge features are credible.

[0038] After being verified by the discriminator, these enhanced lesion feature data are subjected to secondary enhancement processing. Specifically, the edge and texture features of the lesion area are further enhanced to make these features clearer and form a clear contrast with the surrounding normal tissue. The purpose of secondary enhancement is to maximize the recognition of the lesion area, especially in the subsequent automatic analysis and classification process, the clear edges and unique texture features help to improve the classification accuracy of the model. Through this multiple enhancement, the final lesion feature data is not only richer in details, but also has high robustness, can adapt to different types of lesions, and ensure good recognition effects under complex imaging conditions. For example, the edge of the lesion may not be obvious from the boundary of normal tissue in the image, and the secondary enhancement processing can further strengthen these features, making it easier for the model to identify and classify the lesion area.

[0039] In one example, the enhanced lesion feature data is subjected to a graph convolutional network spatial association analysis process to obtain spatial association features of the lesion, including: extracting the spatial distribution of feature points from the enhanced lesion feature data to form a feature point set, and constructing the spatial proximity relationship between the feature points in the feature point set to reflect the spatial distribution structure of the lesion area to obtain a graph data structure; initializing the features of each node in the graph data structure to generate an initial feature value, and embedding the node local features extracted from the enhanced lesion feature data into each node to construct an embedded graph structure containing local feature information; applying a graph convolutional network to perform a convolution operation on each node in the embedded graph structure to obtain the spatial association features between the nodes, and gradually optimizing the convolved spatial association features through layer-by-layer feature fusion processing to construct an updated embedded graph structure containing node spatial association information; performing global feature aggregation processing on the updated embedded graph structure to obtain a global spatial association description of the lesion area, and performing feature weighting processing based on the global spatial association description of the lesion area to generate weighted spatial association features of the lesion area; and regularizing the weighted spatial association features to obtain the final spatial association features.

[0040] That is, the purpose of the spatial correlation analysis processing of the graph convolutional network (GCN) is to extract more complex spatial correlation information from the lesion feature data, to help more comprehensively understand the distribution of lesions in the image and their internal structure. First, the spatial distribution of feature points is extracted from the enhanced lesion feature data to form a feature point set. Each feature point represents an important position in the lesion area, which may correspond to a boundary, texture or other significant features. After forming the feature point set, the spatial proximity relationship between these feature points is constructed based on these feature points to reflect the overall spatial distribution structure of the lesion area, thus obtaining a graph data structure. The nodes in the graph represent feature points, and the edges represent the proximity relationship between feature points. This graph data structure can intuitively describe the geometric characteristics of the lesions.

[0041] Next, the features of each node in the graph data structure are initialized to generate the initial feature value of each node, and the local features of the nodes extracted from the enhanced lesion feature data are embedded into each node, thereby constructing an embedded graph structure containing local feature information. In this way, the embedded graph structure not only contains the spatial position of the node, but also carries rich local feature information, such as grayscale value, texture features, etc., so that each node can better represent its features in the lesion.

[0042] Then, a graph convolutional network is applied to each node in the embedded graph structure for convolution operation. Through graph convolution, the spatial correlation features between nodes can be obtained, thereby capturing the relationship between feature points, such as their distance, interaction, etc. Through layer-by-layer feature fusion processing, the convolution operation can continuously fuse the information of each node, so that each node contains not only its own information, but also the correlation information between neighboring nodes. In this way, after multiple layers of convolution, the characteristics of the node can more comprehensively reflect the spatial relationship of the entire lesion area, and finally construct an updated embedded graph structure containing the spatial correlation information of the nodes.

[0043] After obtaining the updated embedded graph structure, global feature aggregation processing is performed to integrate the features of the entire lesion area to obtain a global spatial association description of the lesion area. This step ensures that the overall features of the lesion area are described in a unified manner, so that not only the local node information is considered, but also the overall spatial characteristics of the lesion are reflected. For example, for lesions with irregular morphology, this step can help to more comprehensively describe their overall morphological characteristics and spatial distribution patterns. Based on the global spatial association description, the features are also weighted to ensure that features that contribute more to lesion classification receive higher weights in subsequent processing. The benefit of this step is that it can further enhance the influence of important features, thereby improving the accuracy of classification and recognition.

[0044] Finally, the weighted spatial correlation features are regularized to obtain the final spatial correlation features. The purpose of regularization is to normalize the eigenvalues ​​to a reasonable range to avoid calculation errors or instability due to excessively large or small values ​​in subsequent analysis. The final spatial correlation features are a high-level description of the lesion area based on the combination of overall spatial distribution and local features, which can be used for further classification and evaluation. Through the spatial correlation analysis of this graph convolutional network, the spatial distribution information of the lesion is fully extracted and integrated, which helps to improve the accuracy and robustness of the model in identifying and classifying the lesion area.

[0045] In one example, adaptive fuzzy logic classification processing is performed on the spatial correlation features of the lesion to obtain a probability value that the image data matches a cholesteatoma sample image, including: performing fuzzy membership calculation on the spatial correlation features of the lesion to obtain a fuzzy membership value of each feature, and fuzzifying the features based on the fuzzy membership values ​​to obtain a feature set containing all fuzzified features; performing fuzzy rule base matching on the fuzzified feature set to obtain a fuzzy rule set that best matches the lesion features, and performing fuzzy reasoning on the fuzzy rule set to obtain a preliminary classification membership of the lesion features; adaptively weighting the preliminary classification membership to obtain an optimized classification membership, and defuzzifying the optimized classification membership to obtain a deterministic feature value of lesion classification; and performing probability calculation processing on the deterministic feature value of the lesion classification to obtain a probability value that the image data matches a cholesteatoma sample image, as well as the category of each matched cholesteatoma sample image.

[0046] The purpose of adaptive fuzzy logic classification processing is to further process the spatial correlation features of the lesions to obtain the probability value of the similarity between the target image data and the cholesteatoma sample image, so as to help doctors better judge the nature and classification of the lesions. First, the fuzzy membership of the spatial correlation features of the lesions is calculated. The fuzzy membership indicates the possibility that the feature belongs to a certain category. It is not a simple binary judgment, but a value between 0 and 1, which is used to describe uncertainty. For example, a lesion feature may partially conform to a disease feature, but may also partially not conform. This uncertainty can be quantified through the fuzzy membership.

[0047] Based on the fuzzy membership value of each feature, the features are fuzzified to obtain a feature set containing all fuzzy features. This step transforms the features of the lesions from the original deterministic description to a description with uncertainty, making the features more representative of the actual situation, especially for lesions with irregular shapes or blurred boundaries. Fuzzy processing helps to make more robust classification in the presence of ambiguity and uncertainty.

[0048] Next, the fuzzy feature set is matched with the fuzzy rule base, with the goal of finding the rule set that best matches the lesion features. The fuzzy rule base contains classification criteria for different feature combinations, which are constructed through prior knowledge and training data. Through matching, the fuzzy rules that best match the lesion features can be found, and fuzzy reasoning can be performed through these rules to obtain the preliminary classification membership of the lesion features. Fuzzy reasoning is the logical inference of features based on fuzzy rules to determine the possible category of the lesion. For example, if the morphological features of the lesion are highly similar to the typical features of a certain type of cholesteatoma, fuzzy reasoning will result in a higher classification membership value.

[0049] For the initially obtained classification membership, adaptive weight adjustment is performed to optimize the classification results. The weight adjustment process is to make the features that have a greater impact on the classification results have a higher weight in the reasoning process, while those less important features have their influence reduced. This adaptive adjustment can make the classification results more reliable and accurate. After the weight adjustment is completed, the optimized classification membership is defuzzified to obtain the deterministic feature value of the lesion classification. Defuzzification is to convert fuzzy descriptions into clear values ​​for subsequent calculations and analysis. For example, the fuzzy probability values ​​of a lesion belonging to different categories are merged into a definite classification result.

[0050] Finally, the deterministic feature values ​​of lesion classification are probability calculated to determine the probability value of the target image data matching the cholesteatoma sample image. This probability value represents the degree of similarity between the image data and different types of cholesteatoma, helping doctors to determine the possible type of lesions. In addition, the category of each matching cholesteatoma sample image can also be obtained. Through this probability calculation, doctors can determine the category and risk level of the lesion based on different probability values, thereby assisting in decision-making. For example, the probability of a lesion image matching a certain type of cholesteatoma is 85%, while the probability of matching with other categories is lower. In this case, it can be judged that the lesion is highly suspected to be a cholesteatoma of this category, thereby improving the accuracy of diagnosis.

[0051] It should be noted that the method provided in this embodiment is essentially only an image processing method rather than a diagnostic method, and this is specially explained.

[0052] In summary, the purpose of the present invention is to achieve accurate detection and classification of cholesteatoma lesions, and by combining a variety of advanced data processing technologies, the identification of lesion areas in images is more accurate. In order to help understand and actually implement the entire solution, a more detailed operation explanation will be provided here.

[0053] First, it is necessary to obtain the medical imaging data of the target patient, which usually comes from CT (computed tomography) or MRI (magnetic resonance imaging) equipment and is saved in DICOM format or other medical imaging standard formats. The quality of data acquisition determines the reliability of subsequent analysis, so the image should be clear and the artifacts and noise that may exist in the imaging process should be minimized. After the image is acquired, it is first necessary to perform quantum feature screening. The so-called quantum feature screening is to enhance the contrast and resolution of the image so that the characteristics of the lesion area are more prominent. This step can be done with the help of the quantum state conversion algorithm to simulate the change of quantum state and optimize the contrast and resolution of the image at the pixel level. In short, it is to enhance the image details through mathematical calculation methods to make the lesion area more prominent. After that, the quantum state measurement combined with the minimum mean square error (MSE) method is used to remove the noise in the image. The purpose of this method is to reduce the influence of random noise on the characteristics of the lesion. For example, noise points that may appear in the image due to the influence of the imaging environment can be removed in this way.

[0054] After the quantum feature screening process is completed, the enhanced and denoised image data is then subjected to topological data segmentation. The core purpose of the segmentation process is to accurately separate the lesion area from the normal tissue for subsequent analysis. First, a topological structure analysis is required. This step attempts to identify the boundaries of the lesion by analyzing the image pixels. Topological analysis is similar to outlining the outline of an object in a complex image. For the lesion area in the image, this step can help determine its shape and location. After the initial identification of the lesion area boundary, a multi-scale segmentation method is used to further refine the segmentation. This means that the image is processed multiple times at different scales (i.e., at different image magnifications) to ensure that even if the lesion area boundary is blurred, the segmentation result can still achieve a high degree of accuracy.

[0055] The topological algorithm based on region growing is used here. This is an algorithm that starts from the seed point (i.e. the initial position of the lesion area) and gradually expands to separate the lesion area. The key to the region growing algorithm is to select a suitable seed point and judge whether the adjacent pixels should belong to the lesion area according to certain growth criteria. This method is very effective for lesions with good connectivity. It can clearly separate the lesions from the surrounding normal tissues and obtain multi-scale topological feature data of a lesion. In the process of multi-scale processing, boundary detection and region growth can be performed separately for different scales to ensure the integrity and accuracy of the lesion boundary and adapt to lesion areas of different sizes and shapes.

[0056] For the lesion area that has been segmented, it is necessary to further extract its internal key structural features. These structural features mainly include the morphological characteristics of the lesion (such as shape, area, perimeter, contour complexity, etc.) and texture features (such as the regularity of internal grayscale distribution, roughness, uniformity, etc.). These features provide detailed information for subsequent classification. By extracting morphological features, the geometric information of the lesion can be obtained, such as whether the area of ​​a lesion has increased, whether the shape is irregular, etc. By extracting texture features, the detailed features inside the lesion area can be analyzed, such as whether there is a certain specific texture pattern inside the cholesteatoma. This information is very critical for determining whether the lesion belongs to a cholesteatoma.

[0057] In order to further improve the feature expression effect of the lesion area in the image, the generative adversarial network (GAN) is used for feature enhancement. GAN is a neural network structure that generates and enhances data by confronting the generator and the discriminator. First, the topological feature data of the lesion area is input into the generator network, and the generator will enhance these features to make the edges and textures of the lesions more prominent. After enhancement, the authenticity of the enhanced features is verified through the discriminator network to ensure that these enhanced features achieve a reasonable enhancement effect while retaining the original lesion features, avoiding the introduction of false features or information during the enhancement process.

[0058] The working principle of the generative adversarial network is to optimize feature expression through adversarial learning between the generator and the discriminator. The goal of the generator is to generate more realistic lesion features, while the goal of the discriminator is to distinguish between the generated features and the original features. Through this adversarial training, the generator gradually learns to generate features that retain the original features and are more significant. In multiple rounds of iterative training of the generator and the discriminator, the enhanced feature data is finally more detailed and accurate, thus providing better data support for subsequent classification. After verification by the discriminator, the enhanced lesion feature data is also subjected to secondary enhancement processing of edges and textures to further refine these features, ensuring that the features of the lesion area are prominent enough in complex imaging environments to facilitate subsequent automated analysis and classification.

[0059] The enhanced lesion feature data then needs to undergo graph convolutional network (GCN) spatial correlation analysis. GCN is a deep learning model that can process graph structured data, and is used here to extract spatial correlation features of the lesion area. First, the spatial distribution of feature points is extracted from the lesion feature data, and the spatial proximity relationship between feature points is constructed to describe the spatial distribution structure of the lesion area. Each feature point represents an important position in the image, such as a feature at the boundary or inside. Through these feature points and their proximity relationships, a graph data structure can be formed. In this graph structure, nodes represent feature points, and edges represent the relationships between feature points.

[0060] After initializing each node, the enhanced local features are embedded into the node to construct an embedded graph structure containing local feature information. Then, the embedded graph structure is convolved through a graph convolutional network. The convolution operation can capture the spatial correlation information between feature points, thereby gradually fusing the information of neighboring nodes together, so that each node contains not only its own information, but also the correlation information with neighboring feature points. This gradual fusion process allows the overall spatial characteristics of the lesion area to be better expressed, which helps to more accurately describe the morphology and structure of the lesion.

[0061] After completing the spatial correlation analysis, it is also necessary to perform global feature aggregation on the spatial correlation features of the lesion area. The purpose of global feature aggregation is to integrate the overall features of the lesion area so that the information of each feature point can contribute to the description of the overall lesion area. For example, for lesions with complex morphology, global aggregation can help understand their overall morphological characteristics, not just local details. By weighting the global features, features that contribute more to classification can occupy higher weights in subsequent processing. In this way, the final weighted spatial correlation features can better reflect the overall structure of the lesion area.

[0062] After the global feature aggregation is completed, these features are regularized to ensure that all feature values ​​are within a reasonable range. The purpose of regularization is to avoid the introduction of calculation errors due to feature values ​​that are too large or too small, and to ensure the stability and reliability of subsequent processing. After regularization, the final spatial correlation features are obtained, which will be used for the classification and evaluation of lesions.

[0063] The last step is to perform adaptive fuzzy logic classification on the spatial correlation features of the lesions to obtain the probability value of the target patient's image data matching the cholesteatoma sample image. Fuzzy logic classification is a classification method that is very suitable for dealing with uncertainty and fuzzy boundaries. Especially in medical images, the morphology of lesions is often not completely regular and difficult to classify using precise mathematical descriptions. First, the fuzzy membership of the spatial correlation features is calculated, which means that the probability of each feature point belonging to a certain category is given. The membership value is between 0 and 1, indicating the attribution of the feature between different categories.

[0064] Based on these fuzzy memberships, the features are fuzzified to form a feature set containing all fuzzified features. Then, these fuzzy features are matched with a pre-built fuzzy rule base, and the preliminary classification membership of the lesion features is obtained through fuzzy reasoning. Fuzzy reasoning is to logically infer features through fuzzy rules to determine the possible category of the lesion. For example, if the morphological and texture features of the lesion meet the typical characteristics of a certain cholesteatoma, fuzzy reasoning will result in a higher classification membership value. In order to improve the accuracy of classification, the preliminary classification membership is adaptively weighted so that the most important features in the classification process occupy a larger proportion, while those relatively unimportant features have a reduced impact on the results. The process of adaptive weight adjustment is similar to assigning different importance to each feature, ensuring that the features that have a greater impact on the classification have a higher weight, thereby optimizing the classification results and making the classification results more reliable and stable.

[0065] After the weight adjustment is completed, the fuzzy membership value is converted into a deterministic eigenvalue through defuzzification processing, which facilitates the clear judgment of the category of the lesion. The defuzzification process is to convert the fuzzy description into a specific numerical value, which represents the possibility that the lesion belongs to a certain category. Finally, the probability calculation of the eigenvalue is performed to obtain the matching probability value of the image data and the cholesteatoma sample. This probability value is used to assess the possibility that the target patient has cholesteatoma, helping doctors to make a more accurate diagnosis. In addition, the calculated matching probabilities of each category can also provide doctors with more detailed information to assist doctors in judging the risk level and treatment strategy of the lesion.

[0066] Reference Figure 3 In an embodiment of the present invention, a cholesteatoma image analysis system based on deep learning is also provided, comprising: An acquisition unit 1 is used to acquire the image data of a target patient, and perform quantum feature screening processing on the image data of the target patient to obtain image feature data after preliminary screening; Segmentation unit 2, used for performing topological data segmentation processing on the image feature data after preliminary screening to obtain topological feature data of the lesion area; Enhancement unit 3, used for performing generative adversarial network enhancement processing on the topological feature data of the lesion area to obtain enhanced lesion feature data; The correlation analysis unit 4 is used to perform a graph convolution network spatial correlation analysis on the enhanced lesion feature data to obtain the spatial correlation features of the lesion; The classification unit 5 is used to perform adaptive fuzzy logic classification processing on the spatial correlation characteristics of the lesion to obtain a probability value that the image data matches the cholesteatoma sample image.

[0067] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0068] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0069] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0070] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0071] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0072] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0073] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A cholesteatoma image analysis method based on deep learning, characterized in that: The following steps are involved: Acquiring image data of a target patient, and performing quantum feature screening processing on the image data of the target patient to obtain image feature data after preliminary screening; Performing topological data segmentation processing on the image feature data after preliminary screening to obtain topological feature data of the lesion area; Performing generative adversarial network enhancement processing on the topological feature data of the lesion area to obtain enhanced lesion feature data; Performing graph convolutional network spatial correlation analysis on the enhanced lesion feature data to obtain spatial correlation features of the lesion; Adaptive fuzzy logic classification processing is performed on the spatial correlation features of the lesions to obtain a probability value that the image data matches the cholesteatoma sample image.

2. The cholesteatoma image analysis method according to claim 1, characterized in that: The image data of the target patient is subjected to quantum feature screening processing to obtain image feature data after preliminary screening, including: The contrast and resolution of the image data of the target patient are enhanced by quantum state conversion to highlight the detailed features of the lesion area and obtain quantum enhanced image data; Using quantum state measurement combined with the minimum mean square error criterion to remove random noise in the quantum enhanced image data, retaining high confidence features related to the lesion area, and obtaining denoised quantum feature image data; The denoised quantum feature image data is mapped to a high-dimensional feature space using a quantum Hamiltonian feature extraction algorithm, and the most sensitive features to the lesion area are extracted by analyzing the quantum state evolution and feature weights are assigned to obtain preliminary image feature data; A multi-layer quantum neural network is used to perform layer-by-layer feature interaction analysis, and the nonlinear correlation between the preliminary image feature data is identified through coherent operations of quantum states, so as to obtain multi-dimensional feature interaction image data reflecting the complex spatial relationship of lesions; Based on the feature contribution ranking, the relevance of each feature in the multi-dimensional feature interaction image data is quantitatively evaluated, and the features most relevant to the lesion classification are screened out to obtain the image feature data after preliminary screening.

3. The cholesteatoma image analysis method according to claim 1, characterized in that: The image feature data after preliminary screening is subjected to topological data segmentation processing to obtain topological feature data of the lesion area, including: Perform topological structure analysis on the image feature data after preliminary screening to identify the boundary features of the lesion area and obtain preliminary topological data of the lesion; Performing multi-scale segmentation processing on the preliminary topological data of the lesion, separating the lesion area from the surrounding normal tissue using a topological algorithm based on region growing, and obtaining multi-scale topological feature data of the lesion area; Performing topological feature extraction on the multi-scale topological feature data of the lesion area, extracting key structural features in the lesion area, including morphological features and texture features, to obtain topological feature description data of the lesion; The topological feature description data of the lesion is subjected to topological consistency verification, and the topological feature description data that has passed the topological consistency verification is determined as the final topological feature data of the lesion region.

4. The cholesteatoma image analysis method according to claim 1, characterized in that: The topological feature data of the lesion area is subjected to generative adversarial network enhancement processing to obtain enhanced lesion feature data, including: The topological feature data of the lesion area is input into a preset generative adversarial network, and the feature of the lesion area is enhanced by the generator network to obtain enhanced lesion generated feature data; The enhanced lesion generated feature data is verified for authenticity through the discriminator network to ensure that the enhanced feature data achieves the enhancement effect on the basis of retaining the original features of the lesion area, and obtains the verified lesion feature data; The verified lesion feature data is subjected to secondary enhancement processing of the edge and texture features of the lesion area to obtain the final enhanced lesion feature data.

5. The cholesteatoma image analysis method according to claim 1, characterized in that: The enhanced lesion feature data is subjected to graph convolutional network spatial correlation analysis to obtain spatial correlation features of the lesion, including: Extracting the spatial distribution of feature points from the enhanced lesion feature data to form a feature point set, and constructing a spatial proximity relationship between the feature points in the feature point set to reflect the spatial distribution structure of the lesion area, thereby obtaining a graph data structure; Initializing the features of each node in the graph data structure to generate initial feature values, and embedding the node local features extracted from the enhanced lesion feature data into each node to construct an embedded graph structure containing local feature information; Applying a graph convolutional network to perform a convolution operation on each node in the embedded graph structure to obtain spatial correlation features between nodes, and gradually optimizing the convolved spatial correlation features through layer-by-layer feature fusion processing to construct an updated embedded graph structure containing node spatial correlation information; Performing global feature aggregation processing on the updated embedded graph structure to obtain a global spatial association description of the lesion area, and performing feature weighting processing based on the global spatial association description of the lesion area to generate a weighted spatial association feature of the lesion area; and Regularization is performed on the weighted spatial correlation features to obtain final spatial correlation features.

6. The cholesteatoma image analysis method according to claim 1, characterized in that: The spatial correlation features of the lesions are subjected to adaptive fuzzy logic classification processing to obtain a probability value of the image data matching the cholesteatoma sample image, including: Calculating the fuzzy membership of the spatial correlation features of the lesions to obtain the fuzzy membership value of each feature, and fuzzifying the features based on the fuzzy membership value to obtain a feature set containing all fuzzy features; Performing fuzzy rule base matching on the fuzzified feature set to obtain a fuzzy rule set that best matches the lesion feature, and performing fuzzy reasoning on the fuzzy rule set to obtain a preliminary classification membership of the lesion feature; Adaptively adjusting the weight of the preliminary classification membership to obtain an optimized classification membership, and defuzzifying the optimized classification membership to obtain a deterministic characteristic value of lesion classification; The deterministic characteristic values ​​of the lesion classification are subjected to probability calculation processing to obtain the probability value of the image data matching the cholesteatoma sample image and the category of each matched cholesteatoma sample image.

7. A cholesteatoma image analysis system based on deep learning, characterized in that: include: An acquisition unit, used to acquire image data of a target patient, and perform quantum feature screening processing on the image data of the target patient to obtain image feature data after preliminary screening; A segmentation unit, used to perform topological data segmentation processing on the image feature data after preliminary screening to obtain topological feature data of the lesion area; An enhancement unit, configured to perform a generative adversarial network enhancement process on the topological feature data of the lesion area to obtain enhanced lesion feature data; A correlation analysis unit, used to perform graph convolution network spatial correlation analysis on the enhanced lesion feature data to obtain spatial correlation features of the lesion; The classification unit is used to perform adaptive fuzzy logic classification processing on the spatial correlation characteristics of the lesion to obtain a probability value of the image data matching the cholesteatoma sample image.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images