A feature classification method for medical image detection

Abstract

The present invention discloses a feature classification method for medical image detection, aiming to solve the problem of low accuracy and efficiency in detecting rare disease features in the prior art. The present invention significantly improves the detection capability of rare and unknown disease features in medical images by integrating object-aware distillation mechanism and pyramid feature fusion technology. The deep learning model is used for efficient feature extraction, and an advanced knowledge distillation process is adopted to transfer the rich knowledge in known disease images to the model to assist the model in learning and understanding the features of unknown or rare diseases.

Description

Technical Field

[0001] The invention relates to a feature classification technology for medical image detection, and belongs to the field of image detection. Background Art

[0002] With the rapid development of medical imaging technology, medical images play an increasingly important role in the modern medical field. These images provide valuable information for the early detection, diagnosis and treatment of diseases. However, with the explosive growth of medical data, traditional manual image analysis methods are increasingly unable to cope with the situation due to their low efficiency and high subjectivity. Therefore, it is imperative to develop efficient and automated medical image analysis technology.

[0003] In the past few decades, medical image analysis technology has evolved from early rule-based methods to modern deep learning-based methods. Although deep learning methods have made significant progress in the automatic recognition and classification of medical images, they still face a series of challenges. First, most deep learning models require a large amount of annotated data for training, and in the medical field, obtaining a large amount of high-quality annotated data is both expensive and time-consuming. Second, deep learning models usually perform poorly when dealing with rare diseases or emerging disease types, because these cases often lack sufficient representation in the training data. In addition, existing medical image analysis technologies are also insufficient in multimodal data fusion. The diversity of medical imaging technologies (such as X-ray, CT, MRI, etc.) provides different angles of view for disease diagnosis, but how to effectively integrate information from different imaging modes to provide a more comprehensive view of the disease remains an open research problem. In addition, existing technologies often use a single method for image feature extraction, which makes it difficult to capture the rich and hierarchical information in the image, limiting the accuracy of the final detection and classification.

[0004] Knowledge distillation, as an emerging technology, shows potential in improving model performance by transferring knowledge from large pre-trained models to smaller models. However, in the field of medical image analysis, how to design effective knowledge distillation strategies to improve the model's ability to recognize rare or novel diseases remains a problem worth exploring. In addition, traditional feature fusion methods often ignore key objects and regions in the image when processing complex medical images, resulting in the model being unable to fully utilize all useful information in the image. Therefore, developing a new medical image analysis technology that can overcome the above limitations, effectively fuse multi-source information, and improve the ability to recognize rare diseases is of great significance to promoting the development of the field of medical image analysis. Summary of the invention

[0005] Purpose of the invention: This invention proposes a feature classification method for medical image detection, combining the object perception mechanism and the framework of the distillation pyramid structure to improve the recognition accuracy of unknown or rare disease features in medical images. By introducing the object perception module, this technology can more accurately identify and extract key objects and areas in the image; at the same time, by designing the distillation pyramid structure, it realizes the effective transfer of knowledge and the deep fusion of features, thereby enhancing the generalization ability and feature expression ability of the model.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A feature classification method for medical image detection comprises the following steps:

[0008] Step 1: Standardize the input medical image, including denoising, contrast enhancement and other steps, to improve the image quality and the efficiency of subsequent processing;

[0009] Step 2: Image-level feature extraction: The teacher model directly processes the entire input image. The teacher model visual encoder is used to extract the global feature representation of the image. The global path of the student model uses the feature pyramid network FPN to encode the entire image to extract global features;

[0010] Step 3: Block-level feature extraction: The input image is divided into multiple blocks of fixed size, and the visual encoder processes each block separately to extract block-level features;

[0011] Step 4: Object-level feature extraction: Transform the image and add a unique tagging module to make the model focus on the image patches covered by the proposal, thereby achieving accurate extraction of object information;

[0012] Step 5: Knowledge distillation: By optimizing the total loss, the student model not only learns visual information at different levels from global to local, but also integrates the rich semantic information provided by the teacher model;

[0013] Step 6: Rare disease detection: Use the previous object perception and knowledge distillation mechanism to improve the recognition accuracy of rare disease features;

[0014] Step 7: Multimodal data fusion: Fusion of data from different medical imaging technologies, such as X-ray, MRI, etc., using multimodal information to improve the accuracy and robustness of detection;

[0015] Step 8: Result correction: Correct the preliminary detection results through post-processing algorithms to reduce false detection and missed detection;

[0016] Step 9: Result output: The final output includes the detection results of the disease type and the corresponding confidence level.

[0017] Furthermore, the step 1 is specifically as follows:

[0018]

[0019] Where X norm Each pixel value in represents the deviation of the original pixel value from the average brightness of the image, measured in units of standard deviation, X is the original image, μ and σ are the mean and standard deviation of the image data, respectively.

[0020] Furthermore, the step 2 is specifically as follows:

[0021] d1=f(X norm )

[0022]

[0023]

[0024] Among them, f(·) is to directly extract global features, g(·) extracts features from images through the FPN pyramid network; d1 is the feature information obtained by the student model, is the feature information obtained by the teacher model, L1 is the loss under image-level feature extraction, and L is the Manhattan distance between the feature information of the teacher model and the student model.

[0025] Furthermore, step 3 is specifically as follows: the teacher model extracts features from local information representing different regions in the image. The student model processes the image through the same block division strategy and uses its feature extraction network to extract features from each block. By imitating the block features extracted by the teacher model, the student model can learn how to extract useful information from local regions.

[0026] p={X i}

[0027] d2=f(p)

[0028]

[0029]

[0030] The difference from the image level is that p is a set of blocks, X i represents the i-th block. The teacher model and the student model use the same operation as step 2 for each block, and then obtain the loss of this level; d2 is the feature information obtained by the student model. is the feature information obtained by the teacher model, and L2 is the loss under block-level feature extraction.

[0031] Furthermore, the step 4 is specifically as follows: in order to ensure the integrity of the object and include as much object information as possible, the image needs to be adaptively transformed into a square area with a side length l of:

[0032]

[0033] Among them, k is the constant scaling ratio, h and w are the horizontal and vertical pixel numbers of the image.

[0034] At this time, the image data is passed into the student model to obtain the feature information d3 obtained by the student model;

[0035] For the teacher model, a new [CHA] module is added as the feature of the disease, which is processed together with the originally used [CLS] during the forward propagation process, but it only focuses on the image blocks covered by the proposal, thereby achieving accurate extraction of object information;

[0036]

[0037] in, is the characteristic word of the image, is a [CLS] lemma, x [CHA] is the special CHA feature word added, X′ is the new image feature matrix after processing; in addition, in order to maintain attention, create a mask N x is the value of the original row of the X matrix;

[0038]

[0039] i is the i-th block after attention segmentation; the complete mask becomes:

[0040]

[0041] After the above operation, the picture is sent to the teacher model to obtain the feature information obtained by the teacher model

[0042]

[0043] L3 is the loss under object-level feature extraction.

[0044] Furthermore, the step 5 is specifically as follows:

[0045] L all =L0+w1·L1+w2·L2+w3·L3

[0046] Where w1, w2 and w3 are the corresponding loss weight coefficients, L all is the final total loss, and L0 is the loss obtained when the image is input into the R-CNN model:

[0047]

[0048] m(p,c) is the logistic score between the proposal embedding and the category embedding; e p is the embedding of proposal p, t c is the embedding of category c, ||e p || and ‖t c ‖ are the embedded modules respectively;

[0049]

[0050] P C (p,c) defines the probability of a proposal p given category c, c r is the element in category c, which is the softmax function of the logistic score so that the sum of the probabilities of all categories is 1;

[0051]

[0052] It is proposed that p belongs to its true label category c r The probability of the loss function is summed over all proposals p; C R A is the base category, while C is the total category.

[0053] Furthermore, the step 6 is specifically as follows:

[0054] P rare =DetectRare(O,K distil )

[0055] Where P rare is the probability of detecting a rare disease, K distill Represents the key information extracted from the trained teacher-student model to improve the detection ability of rare diseases.

[0056] Furthermore, the step 7 is specifically as follows:

[0057] F multi =Fuse(F X-ray ,F MRI )

[0058] where F X-ray and F MRI Represent the feature vectors of X-ray and MRI images respectively, F multi It represents a composite feature vector that fuses features from different imaging modalities, with the goal of leveraging their respective strengths to improve the comprehensiveness and accuracy of medical diagnosis.

[0059] Furthermore, the step 8 is specifically as follows:

[0060] R corrected =Correct(R initial )

[0061] Where R initial is the preliminary test result, R corrected is the result after correction.

[0062] Furthermore, the step 9 is specifically as follows:

[0063]

[0064] Where D i is the i-th detected disease type, P i is the corresponding confidence level, and M is the number of diseases detected.

[0065] In the data preprocessing stage, the present invention not only performs standardization processing on the image, but also adopts more advanced technologies, such as image enhancement, denoising, contrast adjustment, etc., to further improve the image quality. These processing technologies can improve the visibility of the image and the recognizability of the features, thereby providing a clearer and more accurate image basis for subsequent feature extraction and disease detection.

[0066] In the feature extraction stage, the present invention uses the latest deep learning architecture and training technology to significantly improve the efficiency and quality of feature extraction. Through the careful adjustment and optimization of the deep learning model, the model can more effectively learn the complex patterns and features in the image and improve the detection ability of subtle lesions.

[0067] The object perception and recognition system of the present invention can be further optimized by introducing dynamic threshold adjustment and adaptive learning mechanisms, so that the system can automatically adjust the recognition parameters according to the characteristics and complexity of different images. This dynamic adjustment mechanism can improve the system's recognition sensitivity and accuracy for objects of various types and sizes, especially when the image quality is different or the target object is relatively small.

[0068] The knowledge distillation process of the present invention can adopt more advanced methods, such as multi-task learning, transfer learning, etc., to effectively transfer the diagnostic knowledge of known diseases to the model, so as to improve the recognition ability of unknown or rare diseases.

[0069] The present invention adopts a multi-scale fusion strategy in the pyramid feature fusion process, which can capture and utilize the information in the image more comprehensively. By combining features from different depth levels, the system can focus on the global structure and local details of the image at the same time, thereby improving the recognition ability of complex disease features. In addition, the introduction of an adaptive weight allocation mechanism can further optimize the feature fusion process and ensure that key features are given more important consideration in the final disease judgment.

[0070] The beneficial effects of the present invention are: through advanced data preprocessing and optimization of deep learning architecture, the present invention can more effectively extract and utilize key features in medical images, thereby improving the accuracy of classification of various diseases, including rare disease features, and meeting the accuracy requirements of medical image detection and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of the feature classification method used for medical image detection in the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0073] like Figure 1 As shown, the feature classification method for medical image detection of the present invention comprises the following steps:

[0074] Step 1: Data Preprocessing

[0075] Medical images are normalized to eliminate differences between images.

[0076]

[0077] Where X norm Each pixel value in represents the deviation of the original pixel value from the average brightness of the image, measured in units of standard deviation, X is the original image, μ and σ are the mean and standard deviation of the image data, respectively.

[0078] Step 2: Image-level feature extraction

[0079] A pyramid distillation framework is used to extract features from images at the image level, block level, and object level. In the image-level extraction, the teacher model directly processes the entire input image. The teacher model visual encoder is used to extract the global feature representation of the image, which captures the comprehensive information of all objects in the image and their context. The global path of the student model uses the feature pyramid network FPN to encode the entire image to extract global features.

[0080] d1=f(X norm )

[0081]

[0082]

[0083] Among them, f(·) is to directly extract global features, g(·) extracts features from images through the FPN pyramid network; d1 is the feature information obtained by the student model, is the feature information obtained by the teacher model, L1 is the loss under image-level feature extraction, and L is the Manhattan distance between the feature information of the teacher model and the student model.

[0084] Step 3: Block-level feature extraction

[0085] The input image is divided into multiple fixed-size blocks, and the visual encoder processes each block separately to extract block-level features. The teacher model extracts features from these local information representing different regions in the image. The student model processes the image using the same block division strategy and uses its feature extraction network to extract features for each block. By imitating the block features extracted by the teacher model, the student model can learn how to extract useful information from local areas.

[0086] p={X i}

[0087] d2=f(p)

[0088]

[0089]

[0090] The difference from the image level is that p is a set of blocks, X i represents the i-th block. The teacher model and the student model use the same operation as step 2 for each block, and then obtain the loss of this level; d2 is the feature information obtained by the student model. is the feature information obtained by the teacher model, and L2 is the loss under block-level feature extraction.

[0091] Step 4: Object-level feature extraction

[0092] In order to ensure the integrity of the object and include as much object information as possible, the image needs to be adaptively transformed into a square area with a side length l of:

[0093]

[0094] Among them, k is the constant scaling ratio, h and w are the horizontal and vertical pixel numbers of the image.

[0095] At this time, the image data is passed into the student model to obtain the feature information d3 obtained by the student model;

[0096] For the teacher model, a new [CHA] module is added as the feature of the disease, which is processed together with the originally used [CLS] during the forward propagation process, but it only focuses on the image blocks covered by the proposal, thereby achieving accurate extraction of object information;

[0097]

[0098] in, is the characteristic word of the image, is a [CLS] lemma, x [CHA] is the special CHA feature word added, X′ is the new image feature matrix after processing; in addition, in order to maintain attention, create a mask N x is the value of the original row of the X matrix;

[0099]

[0100] i is the i-th block after attention segmentation; the complete mask becomes:

[0101]

[0102] After the above operation, the picture is sent to the teacher model to obtain the feature information obtained by the teacher model

[0103]

[0104] L3 is the loss under object-level feature extraction.

[0105] Step 5: Knowledge Distillation

[0106] By optimizing the total loss, the student model not only learns visual information at different levels from global to local, but also integrates the rich semantic information provided by the teacher model.

[0107] L all =L0+w1·L1+w2·L2+w3·L3

[0108] Where w1, w2 and w3 are the corresponding loss weight coefficients, L all is the final total loss, and L0 is the loss obtained when the image is input into the R-CNN model:

[0109]

[0110] m(p,c) is the logistic score between the proposal embedding and the category embedding; e p is the embedding of proposal p, t c is the embedding of category c, ||e p || and ‖t c ‖ are the embedded modules respectively;

[0111]

[0112] P C (p,c) defines the probability of a proposal p given category c, c ris the element in category c, which is the softmax function of the logistic score so that the sum of the probabilities of all categories is 1;

[0113]

[0114] It is proposed that p belongs to its true label category c r The probability of the loss function is summed over all proposals p; C R A is the base category, while C is the total category.

[0115] Step 6: Rare Disease Testing

[0116] Improving the detection accuracy of rare diseases using object awareness and knowledge distillation mechanisms.

[0117] P rare =DetectRare(O,K distill )

[0118] Where P rare is the probability of detecting a rare disease, K distill Represents the key information extracted from the trained teacher-student model to improve the detection ability of rare diseases.

[0119] Step 7: Multimodal data fusion

[0120] Combining different types of medical images (e.g., X-ray, MRI) to improve detection performance.

[0121] F multi =Fuse(F X-ray ,F MRI )

[0122] where F X-ray and F MRI Represent the feature vectors of X-ray and MRI images respectively, F multi It represents a composite feature vector that fuses features from different imaging modalities, with the goal of leveraging their respective strengths to improve the comprehensiveness and accuracy of medical diagnosis.

[0123] Step 8: Result Correction

[0124] The detection results are corrected through post-processing algorithms to reduce false detections.

[0125] R corrected =Correct(R initial )

[0126] Where R initial is the preliminary test result, R corrected is the result after correction.

[0127] Step 9: Output the results

[0128] Output the final test results, including disease type and confidence level.

[0129]

[0130] Where D i is the i-th detected disease type, P i is the corresponding confidence level, and M is the number of diseases detected.

[0131] The present invention adopts a pyramid feature fusion strategy to improve the model's ability to recognize complex structures and subtle lesions in medical images by combining feature information at different levels, thereby enhancing the accuracy and sensitivity of the model in comprehensive diagnosis.

[0132] The present invention adaptively optimizes the parameter settings of the object perception module through a dynamic adjustment algorithm, so that it can be automatically adjusted according to the specific characteristics and complexity of the image to improve the adaptability and recognition efficiency of various medical images.

[0133] The present invention combines data from multiple medical imaging technologies (such as X-ray and MRI) to improve the overall performance of detection, including accuracy and robustness, to adapt to complex clinical application scenarios.

[0134] The present invention introduces advanced deep learning architectures, such as attention mechanisms and residual networks, to more accurately extract and analyze key features in medical images, significantly improving the performance and reliability of the model, especially when dealing with images of varying quality or small target objects.

[0135] The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A feature classification method for medical image detection, characterized in that: The following steps are involved: Step 1: Standardize the input medical image, including denoising and contrast enhancement; Step 2: Image-level feature extraction: The teacher model directly processes the entire input image; the teacher model visual encoder is used to extract the global feature representation of the image; the global path of the student model uses the feature pyramid network FPN to encode the entire image to extract global features; Step 3: Block-level feature extraction: The input image is divided into multiple blocks of fixed size, and the visual encoder processes each block separately to extract block-level features; Step 4: Object-level feature extraction: Transform the image and add a unique tagging module to make the model focus on the image patches covered by the proposal, thereby achieving accurate extraction of object information; Step 5: Knowledge distillation: By optimizing the total loss, the student model not only learns visual information at different levels from global to local, but also integrates the rich semantic information provided by the teacher model; Step 6: Rare disease detection: Detection using previous object perception and knowledge distillation mechanisms; Step 7: Multimodal data fusion: Fusion of data from different medical imaging technologies, including X-ray and MRI; Step 8: Result correction: Correct the preliminary test results through post-processing algorithms; Step 9: Result output: The final output includes the detection results of the disease type and the corresponding confidence level; The step 1 is specifically as follows: Where X norm Each pixel value in represents the deviation of the original pixel value from the average brightness of the image, measured in units of standard deviation, X is the original image, μ and σ are the mean and standard deviation of the image data respectively; The step 2 is specifically as follows: d1=f(X norm ) Among them, f(·) is to directly extract global features, g(·) extracts features from images through the FPN pyramid network; d1 is the feature information obtained by the student model, is the feature information obtained by the teacher model, L1 is the loss under image-level feature extraction, and L is the Manhattan distance between the feature information of the teacher model and the student model; The step 3 is specifically as follows: the teacher model extracts features from local information representing different regions in the image; the student model processes the image through the same block division strategy and uses its feature extraction network to extract features from each block; by imitating the block features extracted by the teacher model, the student model can learn how to extract useful information from the local region; p={X i } d2=f(p) The difference from the image level is that p is a set of blocks, X i represents the i-th block. The teacher model and the student model use the same operation as step 2 for each block, and then obtain the loss of this level; d2 is the feature information obtained by the student model. is the feature information obtained by the teacher model, and L2 is the loss under block-level feature extraction; The step 4 is specifically as follows: In order to ensure the integrity of the object and include as much object information as possible, the image needs to be adaptively transformed into a square area with a side length l of: Where k is the constant scaling ratio, h and w are the horizontal and vertical pixel numbers of the image; At this time, the image data is passed into the student model to obtain the feature information d3 obtained by the student model; For the teacher model, a new [CHA] module is added as the feature of the disease, which is processed together with the originally used [CLS] during the forward propagation process, but it only focuses on the image blocks covered by the proposal, thereby achieving accurate extraction of object information; in, is the characteristic word of the image, is a [CLS] lemma, x [CHA] is a special CHA feature word added, X ′ is the new image feature matrix after processing; in addition, in order to maintain attention, create a mask N x is the value of the original row of the X matrix; i is the i-th block after attention segmentation; the complete mask becomes: After the above operation, the picture is sent to the teacher model to obtain the feature information obtained by the teacher model L3 is the loss under object-level feature extraction; The step 5 is specifically as follows: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> all <h2 style=";text-align:left;direction:ltr"> = L0+w1·L1+w2·L2+w3·L3 Where w1, w2 and w3 are the corresponding loss weight coefficients, L all is the final total loss, and L0 is the loss obtained when the image is input into the R-CNN model: m(p,c) is the logistic score between the proposal embedding and the category embedding; e p is the embedding of proposal p, t c is the embedding of category c, ||e p || and ‖t c ‖ are the embedded modules respectively; P C (p,c) defines the probability of a proposal p given category c, c r is the element in category c, which is the softmax function of the logistic score so that the sum of the probabilities of all categories is 1; It is proposed that p belongs to its true label category c r The probability of the loss function is summed over all proposals p; C R is the basic category, while C is the total category; The step 6 is specifically as follows: P rare =DetectRare(O,K distill ) Where P rare is the probability of detecting a rare disease, K distill Represents the key information extracted from the trained teacher-student model; The step 7 is specifically as follows: F multi =Fuse(F X-ray ,F MRI ) where F X-ray and F MRI Represent the feature vectors of X-ray and MRI images respectively, F multi represents a composite feature vector that fuses features from different imaging modalities; The step 8 is specifically as follows: R corrected =Correct(R initial ) Where R initial is the preliminary test result, R corrected is the result after correction; The step 9 is specifically as follows: Where D i is the i-th detected disease type, P i is the corresponding confidence level, and M is the number of diseases detected.

Images