Intelligent image processing system and method for nuclear medicine

Abstract

The invention discloses a nuclear medicine-oriented image intelligent processing system and method, and relates to the technical field of nuclear medicine image processing, and the main scheme is as follows: receiving original image data and patient medical history, age and gender data; extracting feature maps of different levels of the original image by using a convolutional neural network, calculating a mean value and a standard deviation to determine a fusion weight coefficient, and obtaining a fused feature map; analyzing and processing the feature map to obtain an input feature vector, generating a clinical information vector in combination with medical history and other data, fusing the clinical information vector to obtain a comprehensive feature vector, and calculating an output probability vector to perform preliminary focus detection; calculating a lesion judgment index according to the output probability vector, and when a preliminary judgment condition is met, calculating an uncertainty degree value and a dispersion degree value to assist in judging whether a lesion exists or not; and generating a detailed diagnosis report according to a detection result. Multi-source data and an advanced algorithm are integrated, the accuracy and reliability of nuclear medicine image focus detection are improved, and powerful support is provided for disease diagnosis and treatment.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear medicine image processing, and in particular to an image intelligent processing system and method for nuclear medicine. Background Art

[0002] In the field of modern medicine, nuclear medicine plays an irreplaceable role as an important diagnostic and therapeutic method. Nuclear medicine images can provide physiological and pathological information inside the human body, which is crucial for early detection of diseases, accurate diagnosis and formulation of treatment plans.

[0003] However, traditional nuclear medicine image analysis mainly relies on the naked eye observation and experience judgment of doctors. This method has many limitations. On the one hand, the subjective judgment of doctors is easily affected by factors such as fatigue and experience differences, which makes it difficult to ensure the accuracy and consistency of the diagnosis results. On the other hand, it is difficult to fully explore the rich information contained in the image by relying solely on naked eye observation, and some key pathological features may be missed. Summary of the invention

[0004] 1. Technical issues to be resolved In view of the deficiencies of the prior art, the present invention provides an image intelligent processing system and method for nuclear medicine, wherein the data input module receives the original image data and the patient's medical history, age and gender data, the feature extraction module uses a convolutional neural network to extract feature maps of different levels and calculates the fusion weight coefficient to obtain the fused feature map, the first data processing module fuses the clinical information with the image features to obtain a comprehensive feature vector to calculate the output probability vector for preliminary lesion detection, the second data processing module calculates the uncertainty and dispersion of the output probability vector to assist in judging the lesion, and the result output module generates a detailed diagnosis report. This improves the accuracy, objectivity and reliability of nuclear medicine image analysis, fully integrates multi-source information, reduces the risk of misjudgment, and provides doctors with a more comprehensive and accurate diagnosis basis.

[0005] (II) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: including: Receive original image data, patient's medical history data BS, age data NL and gender data XB; The convolutional neural network (CNN) is used to extract feature maps corresponding to different levels in the original image data, calculate the mean and standard deviation of all pixel values ​​in the feature maps corresponding to different levels, calculate the fusion weight coefficients of different levels based on the mean and standard deviation, and obtain the fused feature maps of different levels based on the feature maps and the fusion weight coefficients; The input feature vector x is obtained by analyzing and processing the feature graphs after fusion at different levels; the clinical information vector c is generated according to the medical history data BS, age data NL and gender data XB, and the clinical information vector c is fused with the input feature vector x to obtain the comprehensive feature vector , based on the comprehensive feature vector Calculate the output probability vector and conduct preliminary lesion detection; According to the output probability vector The probability of lesions in and normal tissue class probability Calculation of lesion determination index ; Preset judgment threshold ; The lesion determination index and judgment threshold Compare and calculate the corresponding lesion probability when the preliminary judgment conditions are met The uncertainty value and the corresponding lesion probability The dispersion value , assist in determining whether a lesion exists; Generate a detailed diagnostic report based on the test results.

[0006] In the preferred embodiment of the above-mentioned method for intelligent image processing for nuclear medicine: calculating the mean of the Lth layer and standard deviation , the formulas based on are:

[0007]

[0008] in, is the total number of pixels in the Lth layer feature map, L is the layer number of the original image data, and its value is a positive integer; Indicates that the L-th layer feature map is at coordinate The pixel value at .

[0009] In the preferred embodiment of the above-mentioned method for intelligent image processing for nuclear medicine, the method for obtaining the fused feature maps at different levels is as follows: According to the standard deviation of the pixel values ​​of the L-th level feature map And the maximum value of the standard deviation of all hierarchical feature maps Calculate the standard deviation of level L The weight adjustment coefficient , the formula based on is:

[0010] According to the mean of the pixel values ​​of the L-th level feature map Calculate the L-th level mean The weight ratio coefficient , the formula based on is:

[0011] Adjust coefficient based on standard deviation weight and weight ratio coefficient Calculate the fusion weight coefficient of the Lth level , the formula is as follows:

[0012] According to the fusion weight coefficient and feature map Calculate the L-th level fused feature map , the formula is as follows:

[0013] in, is the Lth layer feature map The weight coefficient ranges from 0 to 1, and .

[0014] In the preferred embodiment of the above-mentioned method for intelligent image processing for nuclear medicine, the specific method for obtaining the input feature vector x is: The fused feature map Sum all pixel values ​​in to get the total pixel value , the formula is as follows:

[0015] in, Represents the pixel value in the ath row and bth column of the fused feature map; It is a feature map the number of rows, It is a feature map The number of columns; According to the sum of pixel values Calculate pixel average , the formula is as follows:

[0016] Calculating pixel values With pixel average The pixel difference , the formula is as follows:

[0017] in, Represents the fused feature map The median coordinate is Pixel value and pixel average at a location The pixel difference of All pixel differences Combine them to get the input feature vector x, the formula is as follows: .

[0018] In the preferred embodiment of the above-mentioned method for intelligent image processing for nuclear medicine, the specific method for generating the clinical information vector c is: The medical history data BS is transformed into medical history keywords and normalized to obtain the medical history vector ; The age data NL is digitized and normalized to obtain the age vector ; Binary encode the gender data XB to obtain the gender vector ; Medical history vector , age vector and gender vector The clinical information vector c is obtained by sequential concatenation, based on the following formula: .

[0019] In the preferred embodiment of the above-mentioned method for intelligent processing of images for nuclear medicine: obtaining a comprehensive feature vector The formula is: .

[0020] In the preferred embodiment of the above-mentioned method for intelligent image processing in nuclear medicine: the output probability vector is calculated The method is: Using deep neural network DNN to integrate feature vector Process and calculate the linear transformation result value of the output layer , according to the linear transformation result value Calculate the output probability vector through the Softmax function , the formula based on is:

[0021] in, is the linear transformation result value corresponding to the i-th category in the linear result of the output layer, is the number of categories in the linear result of the output layer, and its value is a positive integer.

[0022] In the preferred embodiment of the above-mentioned method for intelligent image processing for nuclear medicine, the method for performing preliminary lesion detection is: Output probability vector ,in, represents the probability that the input image belongs to the i-th category, i represents the sequence number of the category, k Total number of categories; and ; Set normal tissue category number and lesion category serial number set ; ,in represents the fth lesion category, f represents the sequence number of the lesion category, Indicates the total number of lesion categories; normal tissue category number and lesion category serial number set All are part of category number i; Statistical lesion category number set The probability of lesions in different lesion categories and normal tissue class probability According to the probability of lesions and normal tissue class probability Calculation of lesion determination index , the formula based on is:

[0023] Preset judgment threshold ; when , then it is preliminarily determined that there is a corresponding lesion category in the image lesions; when < , it is preliminarily determined that there is no corresponding lesion category in the image lesions.

[0024] In the preferred embodiment of the above-mentioned method for intelligent image processing for nuclear medicine, the specific method for assisting in determining whether a lesion exists is: When the lesion determination index is met When the condition is The uncertainty value The formula used is as follows:

[0025] When the lesion determination index is met When the condition is The dispersion value The formula used is as follows:

[0026] in, is the output probability vector The mean value of is in the range of 0 to 1. represents the probability of the fth lesion category, f represents the sequence number of the lesion category, represents the total number of lesion categories; Preset first threshold and the second threshold ; The uncertainty value With the first threshold Compare and divide the dispersion value With the second threshold For comparison: When satisfied and When the condition is met, it means that there is a lesion of the corresponding lesion category in the input image; When satisfied and If the condition is met, it means that it cannot be determined whether there is a lesion of the corresponding lesion category in the input image, and the original image data needs to be re-acquired for secondary determination; When satisfied and If the condition is met, it means that it cannot be determined whether there is a lesion of the corresponding lesion category in the input image, and the original image data needs to be re-acquired for secondary determination; When satisfied and When the condition is met, it means that there is no lesion of the corresponding lesion category in the input image.

[0027] The present invention also discloses an image intelligent processing system for nuclear medicine, comprising: A data input module, used for receiving original image data, patient's medical history data BS, age data NL and gender data XB; The feature extraction module is used to extract the feature maps corresponding to different levels in the original image data through the convolutional neural network CNN, calculate the mean and standard deviation of all pixel values ​​in the feature maps corresponding to different levels, calculate the fusion weight coefficients of different levels based on the mean and standard deviation, and obtain the fused feature maps of different levels based on the feature maps and the fusion weight coefficients; The first data processing module is used to analyze and process the feature graphs after fusion at different levels to obtain the input feature vector x; generate the clinical information vector c according to the medical history data BS, age data NL and gender data XB, and fuse the clinical information vector c with the input feature vector x to obtain the comprehensive feature vector , based on the comprehensive feature vector x Calculate the output probability vector and conduct preliminary lesion detection; The second data processing module is used to output the probability vector The probability of lesions in and normal tissue class probability Calculation of lesion determination index ; Preset judgment threshold ; The lesion determination index and judgment threshold Compare and calculate the corresponding lesion probability when the preliminary judgment conditions are met The uncertainty value and the corresponding lesion probability The dispersion value , assist in determining whether a lesion exists; The result output module is used to generate a detailed diagnostic report based on the test results.

[0028] (III) Beneficial effects The present invention provides a patent name and has the following beneficial effects: (1) Receive the original image data as well as the medical history, age and gender data, and realize the organic integration of image information and clinical information. This makes the diagnosis results more comprehensive and accurate, and provides doctors with richer diagnostic basis.

[0029] (2) Use convolutional neural networks to extract feature maps at different levels, and calculate the fusion weight coefficient based on the mean and standard deviation to obtain the fused feature map. This method can make full use of the multi-scale information of the image and improve the ability to represent image features. The fused feature map combines the advantages of different levels and can more accurately reflect the essential characteristics of the image, providing a more reliable basis for subsequent analysis and processing.

[0030] (3) The clinical information vector is fused with the input feature vector to obtain a comprehensive feature vector, based on which the output probability vector is calculated and preliminary lesion detection is performed. This method combines image features and clinical information to more accurately determine the presence or absence of lesions.

[0031] (4) Calculate the uncertainty and dispersion values ​​of the output probability vector to assist in determining whether a lesion exists. This helps doctors conduct further inspections or analyses when faced with uncertain results, thereby reducing the risk of misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The figure is a schematic diagram of the working steps of an image intelligent processing method for nuclear medicine according to the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] See also Figure 1 The present invention provides an image intelligent processing method for nuclear medicine, comprising: Step 1: Receive original image data, patient's medical history data BS, age data NL and gender data XB.

[0035] The specific steps of step one are as follows: The original image data comes from nuclear medicine imaging equipment, including information data such as pixel information data, resolution data, and color depth data of the image, as well as identification information related to the patient, such as the patient's name, examination date, etc.

[0036] The patient's medical history data BS includes the patient's past medical history, symptoms, diagnosis results, treatment history and other information.

[0037] The age data NL is expressed in digital form, the unit is years.

[0038] The gender data XB is expressed in binary form.

[0039] Step 2: Extract the feature maps corresponding to different levels in the original image data through the convolutional neural network CNN, calculate the mean and standard deviation of all pixel values ​​in the feature maps corresponding to different levels, calculate the fusion weight coefficients of different levels based on the mean and standard deviation, and obtain the fused feature maps of different levels based on the feature maps and the fusion weight coefficients.

[0040] The specific steps of step 2 are as follows: Step 201: Calculate the mean of the Lth layer and standard deviation , the formulas based on are:

[0041]

[0042] in, is the total number of pixels in the Lth layer feature map, L is the layer number of the original image data, and its value is a positive integer; Indicates that the L-th layer feature map is at coordinate The pixel value at .

[0043] It should be noted that the feature map of the Lth layer Usually in the form of a two-dimensional matrix with the number of rows , the number of columns is , the total number of pixels in the L-th layer feature map Through the feature map The original image data is extracted using a convolutional neural network, and the number of layers depends on the depth of the network.

[0044] The level number L is used to distinguish the feature extraction results at different levels, and each level corresponds to features of different scales and abstraction levels. For example, the shallow feature map mainly contains low-level detail information of the image, such as edges, textures, colors, etc. These features are usually closely related to the local structure of the image, have a high sensitivity to the specific content of the image, have a high spatial resolution, and can preserve the fine structure and local details of the image. This is very important for identifying the shape and position of specific objects or regions in the image; the deep feature map contains more abstract semantic information, reflecting the high-level concepts and overall meaning of the image. These features are usually related to the category of the image, the attributes of the target object, etc., and have a higher understanding of the overall content of the image. The spatial resolution of the deep feature map is relatively low because the spatial information of the image is gradually compressed and abstracted during the step-by-step processing of the deep neural network. This makes the deep features more suitable for capturing the global structure and semantic relationships of the image.

[0045] Step 202: Based on the standard deviation of the pixel values ​​of the L-th level feature map And the maximum value of the standard deviation of all hierarchical feature maps Calculate the standard deviation of level L The weight adjustment coefficient , the formula based on is:

[0046] It should be noted that the standard deviation It measures the degree of discreteness of pixel values ​​in the feature map. If the standard deviation of a feature map is large, it means that its pixel values ​​are more dispersed, which means that the feature map contains more noise or unstable information. In the fusion process, in order to improve the reliability and smoothness of the fusion result, the weights of these unstable feature maps need to be reduced.

[0047] The principle of this formula is that when the standard deviation of a feature map is close to the maximum value, the weight adjustment coefficient will approach 0, thereby reducing the weight of the feature map; when the standard deviation is small, the weight adjustment coefficient It will be close to 1 and have little impact on the weight.

[0048] Step 203: According to the mean of the pixel values ​​of the L-th level feature map Calculate the L-th level mean The weight ratio coefficient , the formula based on is:

[0049] It should be noted that the mean It reflects the overall intensity level of the feature map. In the scenario of multi-scale feature fusion, feature maps at different levels may indeed have different intensity performances. If the mean value of the feature map at a certain level is high, it means that the pixel value of the feature map at that level is relatively large as a whole, which means that the level captures more significant features or has a stronger response. For example, when performing lesion detection, if the mean value of the feature map at a certain level is high, and the features at that level have a high correlation with the features of the lesion, then the feature map at that level can be given a higher weight during the fusion process to highlight its role in lesion detection.

[0050] The principle of this formula is to use the proportion of the mean of the feature map of the Lth layer to the sum of the mean of feature maps of all levels as the weight of the layer.

[0051] Step 204: Adjust the coefficient according to the standard deviation weight and weight ratio coefficient Calculate the fusion weight coefficient of the Lth level , the formula is as follows:

[0052] It should be noted that the fusion weight coefficient It reflects the importance of the L-th layer feature map in the fusion. The larger the weight coefficient, the greater the proportion of the feature map of this layer in the fused feature map.

[0053] Fusion weight coefficient The mean and standard deviation of the feature map are considered. The mean reflects the importance of the feature, and the standard deviation reflects the stability and noise of the feature. In this way, the value of feature maps at different levels can be more comprehensively evaluated, avoiding unreasonable weight distribution caused by relying solely on a certain attribute.

[0054] Fusion weight coefficient The fused feature map can more effectively integrate information of different scales. Giving greater weights to important and stable feature maps can highlight these key features; reducing the weights of unstable or noisy feature maps can reduce the impact of noise on the fusion results and improve the reliability and smoothness of the fusion results. This will improve the performance of the image intelligent processing system and provide more accurate and reliable feature representation for subsequent lesion detection, diagnosis and other tasks.

[0055] Step 205: According to the fusion weight coefficient and feature map Calculate the L-th level fused feature map , the formula is as follows:

[0056] in, is the Lth layer feature map The weight coefficient ranges from 0 to 1, and .

[0057] It should be noted that this formula combines the feature information of each level by weighted summation of feature maps at different levels to obtain a comprehensive feature representation. Feature maps at different levels contain information of different scales and abstract levels. After fusion, they can better capture the global and local features of the image and improve the ability to understand and analyze the image.

[0058] The fused feature map can provide richer and more effective feature representation for subsequent image intelligent processing tasks, thereby improving the performance of the system. For example, in the lesion detection task, the fused feature map can better identify lesions of different sizes and shapes and improve the accuracy of detection.

[0059] Combined with the contents of step 201 to step 205: By calculating the mean and standard deviation of feature maps at different levels, the characteristics of feature maps can be reflected from different angles. The mean reflects the overall intensity level of the feature map, which helps to identify more significant features and areas of strong response; the standard deviation measures the discrete degree of pixel values, which can be used to judge the stability and noise of the feature map. The weight adjustment coefficient and weight ratio coefficient are calculated according to the standard deviation and the mean, and the fusion weight coefficient is obtained comprehensively. This method fully considers multiple attributes such as the importance, stability and noise of the feature map, avoids the unreasonable weight distribution caused by relying solely on a certain attribute, and can more comprehensively evaluate the value of feature maps at different levels. The fusion weight is used to perform weighted summation of feature maps at different levels to obtain the fused feature map, realizing multi-scale feature fusion. The fused feature map contains information of different scales and abstract levels, which can better capture the global and local features of the image, improve the understanding and analysis ability of the image, and provide more accurate, reliable, rich and effective feature representation for subsequent image intelligent processing tasks, especially lesion detection and diagnosis tasks, and improve the overall performance of the image intelligent processing system.

[0060] Step 3: Analyze and process the fused feature graphs at different levels to obtain the input feature vector x; generate the clinical information vector c based on the medical history data BS, age data NL and gender data XB, and fuse the clinical information vector c with the input feature vector x to obtain the comprehensive feature vector , based on the comprehensive feature vector Calculate the output probability vector And perform preliminary lesion detection.

[0061] The working steps of step three are: Step 301: Fused feature map Sum all pixel values ​​in to get the total pixel value , the formula is as follows:

[0062] in, Represents the pixel value in the ath row and bth column of the fused feature map; It is a feature map the number of rows, It is a feature map The number of columns.

[0063] According to the sum of pixel values Calculate pixel average , the formula is as follows:

[0064] Calculating pixel values With pixel average The pixel difference , the formula is as follows:

[0065] in, Represents the fused feature map The median coordinate is Pixel value and pixel average at a location The pixel difference.

[0066] It should be noted that the sum of all pixel values ​​is obtained by doubly summing all pixel values ​​of the fused feature map. After the sum is obtained, the sum is divided by the total number of pixels in the feature map to obtain the average pixel value. , subtract the average pixel value from the pixel value at that coordinate to get the difference between the pixel value and the average value. The purpose of this step is to measure the degree of deviation of each pixel value from the overall average level.

[0067] Step 302: All pixel differences Combine them to get the input feature vector x, the formula is as follows:

[0068] It should be noted that the input feature vector x reflects the change in pixel distribution of the fused feature map relative to the average level.

[0069] For example, some pixel values ​​are higher than the average value, and some pixel values ​​are lower than the average value. These differences can better capture the local changes and characteristics of the image.

[0070] Assume that the fused feature map is a 3×3 matrix as shown below:

[0071] Compute sums and averages:

[0072]

[0073] Calculate the difference between each pixel value and the mean:

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] Combine the differences into the input feature vector :

[0083] Input feature vector It reflects the change in pixel distribution of the fused feature map relative to the average level of 4.89.

[0084] Step 303: The specific method of generating the clinical information vector c is: Step 3031: Perform medical history keyword conversion and normalization on the medical history data BS to obtain a medical history vector ; For example, a dictionary can be established to map different medical history keywords to specific numerical vectors, and then the medical history vectors are normalized.

[0085] Step 3032: Numerize and normalize the age data NL to obtain an age vector ; Step 3033: Binary encode the gender data XB to obtain the gender vector ; Encode male as 1 and female as 0.

[0086] Step 3034: Transform the medical history vector , age vector and gender vector The clinical information vector c is obtained by sequential concatenation, based on the following formula: .

[0087] Step 304: Obtain comprehensive feature vector The formula is:

[0088] It should be noted that the principle of this formula is to concatenate the input feature vector x and the clinical information vector c together to form a longer vector. This comprehensive feature vector contains both the feature information of the image and the clinical information of the patient.

[0089] Step 305: Use the deep neural network DNN to analyze the comprehensive feature vector Process and calculate the linear transformation result value of the output layer , according to the linear transformation result value Calculate the output probability vector through the Softmax function , the formula based on is:

[0090] in, is the linear transformation result value corresponding to the i-th category in the linear result of the output layer, is the number of categories in the linear result of the output layer, and its value is a positive integer.

[0091] It should be noted that in the calculation of the output probability vector, the deep neural network DNN uses the comprehensive feature vector The first layer is linearly transformed to obtain a new vector, which is then processed by a nonlinear activation function. The subsequent hidden layers repeat the linear transformation and nonlinear activation process to gradually extract high-level features. After the output layer is linearly transformed, the linear transformation result value of the output layer is obtained. , the linear transformation result value is transformed through the Softmax function Convert to probability distribution, that is, output probability vector .

[0092] The role of the Softmax function is to convert the scores of the output layer into the probabilities of each category so that the sum of the probabilities of all categories is 1.

[0093] It represents the linear transformation result value corresponding to the i-th category in the linear result of the output layer with the natural constant as the base.

[0094] For all possible categories The corresponding result values ​​are summed.

[0095] Step 306: The method for performing preliminary lesion detection is: Step 3061: Output probability vector ,in, represents the probability that the input image belongs to the i-th category, i represents the sequence number of the category, Total number of categories; and ; It should be noted that the output probability vector is a vector of probability values ​​that represents the probability distribution of the input image belonging to different categories. These categories include normal tissue categories, lesion categories, and other categories that are not clearly defined, such as .

[0096] The value range of each probability value is 0 to 1. When a probability value is 0, it means that the input image does not belong to this category at all; when the probability value is 1, it means that the input image definitely belongs to this category.

[0097] And the sum of all probability values ​​is 1, ensuring that the sum of the probabilities of the input image in all possible categories is 1.

[0098] Step 3062: Set the normal tissue category number and lesion category serial number set ; ,in represents the fth lesion category, f represents the sequence number of the lesion category, Indicates the total number of lesion categories; normal tissue category number and lesion category serial number set All are part of category number i; It should be noted that setting the lesion category can more comprehensively detect different types of lesions. For example, for nuclear medicine images, different lesion categories can be set, such as tumor, inflammation, hemorrhage, etc.

[0099] Normal tissue categories and lesion categories are set by comprehensively considering medical knowledge, clinical experience, previous studies, image feature analysis, and patient clinical information.

[0100] Step 3063: Count the lesion category number set The probability of lesions in different lesion categories and normal tissue class probability According to the probability of lesions and normal tissue class probability Calculation of lesion determination index , the formula based on is:

[0101] Preset judgment threshold .

[0102] when , then it is preliminarily determined that there is a corresponding lesion category in the image lesions.

[0103] when < , it is preliminarily determined that there is no corresponding lesion category in the image lesions.

[0104] It should be noted that when it is preliminarily determined that a lesion of the corresponding lesion category exists in the image, this means that based on the current output probability vector and determination index, the image has a high probability of containing a specific type of lesion, and further inspection and analysis are required to confirm the existence and specific type of the lesion.

[0105] Judgment index and judgment threshold It is set based on comprehensive consideration of medical expertise and expert clinical experience, statistical methods, model training and verification, and multimodal information fusion. In actual application, it needs to be adjusted according to the situation.

[0106] Combined with the contents of step 301 to step 306: by summing up the pixel values ​​of the fused feature map and calculating the average value, and obtaining the difference between each pixel value and the average value, the degree of deviation of each pixel value relative to the overall average level can be effectively measured, and the local changes and features of the image can be better captured, laying the foundation for the subsequent generation of the input feature vector. The medical history vector is obtained by keyword conversion and normalization of the medical history data, the age vector is obtained by digitization and normalization of the age data, and the gender vector is obtained by binary encoding of the gender data, and these vectors are spliced ​​into clinical information vectors, which fully integrates the multi-source clinical information of the patient. The input feature vector and the clinical information vector are spliced ​​to obtain a comprehensive feature vector, which contains both the feature information of the image and the clinical information of the patient, providing comprehensive data support for more accurate analysis. The comprehensive feature vector is analyzed and calculated to calculate the probability that the image belongs to different categories, so that the calculated output probability can accurately reflect the relationship between the features of the image and the category. In the initial lesion detection, by setting the normal tissue category and multiple lesion categories, combined with the output probability vector, judgment index and preset judgment threshold, it is possible to preliminarily judge whether there are lesions of the corresponding lesion category in the image, providing doctors with a preliminary basis for diagnosis. At the same time, it also reminds that when it is preliminarily determined that there are lesions, further inspection and analysis are needed to confirm the existence and specific type of lesions. When it is preliminarily determined that there are no lesions, it cannot be completely ruled out and needs to be comprehensively judged in combination with other clinical information. The setting of judgment indexes and judgment thresholds comprehensively considers medical expertise and expert clinical experience, statistical methods, model training and verification, and multimodal information fusion, making the entire system more accurate and reliable in practical applications.

[0107] Step 4: Based on the output probability vector The probability of lesions in and normal tissue class probability Calculation of lesion determination index ; Preset judgment threshold ; The lesion determination index and judgment threshold Compare and calculate the corresponding lesion probability when the preliminary judgment conditions are met The uncertainty value and the corresponding lesion probability The dispersion value , to assist in determining whether a lesion exists.

[0108] The working steps of step 4 are: Step 401: When the lesion determination index is met When the condition is The uncertainty value The formula used is as follows:

[0109] It should be noted that the purpose of taking the logarithm in this formula is to convert the probability of lesions into Mapped to a numerical range that better reflects uncertainty. The range of uncertainty values ​​is from negative infinity to 0.

[0110] When the probability of lesion Near 1 o'clock, Close to 0. In this case, it means that the existence of the lesion has a higher certainty. A high probability value means that the image is likely to belong to this lesion category and the uncertainty is low.

[0111] When the probability of lesion Near 0 o'clock, Approaching infinity. The existence of this lesion has extremely high uncertainty, and a low probability value indicates that the image is unlikely to belong to this lesion category.

[0112] In the intelligent processing of nuclear medicine images, the uncertainty degree value can help the present invention understand the reliability of the model's judgment of a specific lesion. If the uncertainty degree value is high, it means that the features in the image are unclear, or the model has limited ability to judge this type of lesion.

[0113] For example, when judging a rare lesion, due to the small number of data samples, the model may have a large uncertainty in estimating its probability. By observing the uncertainty value, doctors can be more cautious about the possible lesion judgment results and conduct further examinations or analyses.

[0114] Step 402: When the lesion determination index is met When the condition is The dispersion value The formula used is as follows:

[0115] in, is the output probability vector The mean value of is in the range of 0 to 1. represents the probability of the fth lesion category, f represents the sequence number of the lesion category, represents the total number of lesion categories; It should be noted that this formula represents the difference between the probability of a specific lesion and the mean probability of all categories. When the probability of a lesion is large or small, it means that the probability of the lesion is greatly different from the mean and the distribution is relatively scattered. At this time, it cannot be determined that it is a real lesion.

[0116] Step 403: Preset a first threshold and the second threshold ; The uncertainty value With the first threshold Compare and divide the dispersion value With the second threshold For comparison: When satisfied and When the condition is met, it means that there is a lesion of the corresponding lesion category in the input image.

[0117] When satisfied and If the condition is not met, it means that it cannot be determined whether there is a lesion of the corresponding lesion category in the input image, and the original image data needs to be re-acquired for secondary judgment; for example, in a lung nuclear medicine image, if the edge of a lung nodule is found to be blurred, try to adjust the scanner's resolution or enhance the contrast settings, and then rescan to see whether the features of the nodule are more obvious.

[0118] When satisfied and When the condition is met, it means that it cannot be determined whether there is a lesion of the corresponding lesion category in the input image, and it is necessary to re-acquire the original image data for a second judgment; for example, there are problems such as artifacts and noise in the original image data. In the nuclear medicine images of patients with hip replacement, the artifacts around the metal implants may affect the lesion judgment; in cardiac nuclear medicine images, noise may mask some small myocardial lesion areas, making the model's judgment on these areas uncertain, thereby increasing the dispersion value, so it is necessary to re-acquire clearer image data for a second judgment.

[0119] When it is uncertain whether there is a lesion of the corresponding lesion category in the input image, it can also be handled by combining the patient's medical history, clinical symptoms and other examination results, through comprehensive judgment by a professional doctor, etc. The specific situation is selected according to actual needs.

[0120] When satisfied and When the condition is met, it means that there is no lesion of the corresponding lesion category in the input image.

[0121] First threshold and the second threshold It is determined based on a large amount of experimental data and clinical experience. By analyzing a large number of known cases, an appropriate range of uncertainty and dispersion is found to distinguish between cases where the lesion is confirmed and cases where the lesion is uncertain.

[0122] For example, different types of nuclear medicine images can be collected, including images with clear lesions and images without clear lesions, and then their uncertainty and dispersion values ​​can be calculated through the model. According to the distribution of these data, a threshold range that can better distinguish the presence or absence of lesions can be determined.

[0123] Combined with the contents of step 401 to step 403: by combining the lesion probability and the normal tissue category probability in the output probability vector to calculate the lesion determination index, and further calculating the uncertainty degree value and the dispersion degree value to assist in judging whether the lesion exists, a more comprehensive and accurate basis is provided for the lesion diagnosis of nuclear medicine images. When the uncertainty degree and the dispersion degree are within an acceptable range, it can be relatively certain that the corresponding lesion category exists in the input image, and the doctor can formulate further diagnosis and treatment plans accordingly. In the case of different combinations of uncertainty and dispersion degree, it can clearly prompt the doctor that it is not certain whether the lesion exists at present, prompting the doctor to take different countermeasures, such as re-evaluating the image quality, using other diagnostic methods, analyzing image features, or combining the patient's medical history for comprehensive judgment. The first threshold and the second threshold determined by a large amount of experimental data and clinical experience ensure the scientificity and reliability of the judgment, which helps to improve the accuracy and practicality of the nuclear medicine image intelligent processing system in lesion diagnosis, and provide doctors with more valuable diagnostic references, so as to better serve the treatment and rehabilitation of patients.

[0124] Step 5: Generate a detailed diagnostic report based on the test results.

[0125] On the other hand, the present invention also discloses an image intelligent processing system for nuclear medicine, comprising: The data input module is used to receive original image data, patient's medical history data BS, age data NL and gender data XB.

[0126] The feature extraction module is used to extract the feature maps corresponding to different levels in the original image data through the convolutional neural network CNN, calculate the mean and standard deviation of all pixel values ​​in the feature maps corresponding to different levels, calculate the fusion weight coefficients of different levels based on the mean and standard deviation, and obtain the fused feature maps of different levels based on the feature maps and the fusion weight coefficients.

[0127] The first data processing module is used to analyze and process the feature graphs after fusion at different levels to obtain the input feature vector x; generate the clinical information vector c according to the medical history data BS, age data NL and gender data XB, and fuse the clinical information vector c with the input feature vector x to obtain the comprehensive feature vector , based on the comprehensive feature vector x Calculate the output probability vector And perform preliminary lesion detection.

[0128] The second data processing module is used to output the probability vector The probability of lesions in and normal tissue class probability Calculation of lesion determination index ; Preset judgment threshold ; The lesion determination index and judgment threshold Compare and calculate the corresponding lesion probability when the preliminary judgment conditions are met The uncertainty value and the corresponding lesion probability The dispersion value , to assist in determining whether a lesion exists.

[0129] The result output module is used to generate a detailed diagnostic report based on the test results.

[0130] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented in electronic hardware or in combination with computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0131] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0132] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. An intelligent image processing method for nuclear medicine, characterized in that: include: Receive original image data, patient's medical history data BS, age data NL and gender data XB; The convolutional neural network (CNN) is used to extract feature maps corresponding to different levels in the original image data, calculate the mean and standard deviation of all pixel values ​​in the feature maps corresponding to different levels, calculate the fusion weight coefficients of different levels based on the mean and standard deviation, and obtain the fused feature maps of different levels based on the feature maps and the fusion weight coefficients; The input feature vector x is obtained by analyzing and processing the feature graphs after fusion at different levels; the clinical information vector c is generated according to the medical history data BS, age data NL and gender data XB, and the clinical information vector c is fused with the input feature vector x to obtain the comprehensive feature vector , based on the comprehensive feature vector Calculate the output probability vector and conduct preliminary lesion detection; According to the output probability vector The probability of lesions in and normal tissue class probability Calculation of lesion determination index ; Preset judgment threshold ; The lesion determination index and judgment threshold Compare and calculate the corresponding lesion probability when the preliminary judgment conditions are met The uncertainty value and the corresponding lesion probability The dispersion value , assist in determining whether a lesion exists; Generate a detailed diagnostic report based on the test results.

2. The method for intelligent image processing for nuclear medicine according to claim 1, characterized in that: Calculate the mean of the Lth layer and standard deviation , the formulas based on are: in, is the total number of pixels in the Lth layer feature map, L is the layer number of the original image data, and its value is a positive integer; Indicates that the L-th layer feature map is at coordinate The pixel value at .

3. The method for intelligent image processing for nuclear medicine according to claim 2, characterized in that: The method to obtain the fused feature maps at different levels is: According to the standard deviation of the pixel values ​​of the L-th level feature map And the maximum value of the standard deviation of all hierarchical feature maps Calculate the standard deviation of level L The weight adjustment coefficient , the formula based on is: According to the mean of the pixel values ​​of the L-th level feature map Calculate the L-th level mean The weight ratio coefficient , the formula based on is: Adjust coefficient based on standard deviation weight and weight ratio coefficient Calculate the fusion weight coefficient of the Lth level , the formula is as follows: According to the fusion weight coefficient and feature map Calculate the L-th level fused feature map , the formula is as follows: in, is the Lth layer feature map The weight coefficient ranges from 0 to 1, and .

4. The method for intelligent image processing for nuclear medicine according to claim 3, characterized in that: The specific method to obtain the input feature vector x is: The fused feature map Sum all pixel values ​​in to get the total pixel value , the formula is as follows: in, Represents the pixel value in the ath row and bth column of the fused feature map; It is a feature map the number of rows, It is a feature map The number of columns; According to the sum of pixel values Calculate pixel average , the formula is as follows: Calculating pixel values The pixel average The pixel difference , the formula is as follows: in, Represents the fused feature map The median coordinate is Pixel value and pixel average at a location The pixel difference of All pixel differences Combine them to get the input feature vector x, the formula is as follows: 。 5. The method for intelligent image processing for nuclear medicine according to claim 4, characterized in that: The specific method for generating the clinical information vector c is: The medical history data BS is transformed into medical history keywords and normalized to obtain the medical history vector ; The age data NL is digitized and normalized to obtain the age vector ; Binary encode the gender data XB to obtain the gender vector ; Medical history vector , age vector and gender vector The clinical information vector c is obtained by sequential concatenation, based on the following formula: 。 6. The method for intelligent image processing for nuclear medicine according to claim 5, characterized in that: Get the comprehensive feature vector The formula is: 。 7. The method for intelligent image processing for nuclear medicine according to claim 6, characterized in that: Calculate the output probability vector The method is: Using deep neural network DNN to integrate feature vector Process and calculate the linear transformation result value of the output layer , according to the linear transformation result value Calculate the output probability vector through the Softmax function , the formula based on is: in, is the linear transformation result value corresponding to the i-th category in the linear result of the output layer, is the number of categories in the linear result of the output layer, and its value is a positive integer.

8. The method for intelligent image processing for nuclear medicine according to claim 7, characterized in that: The methods for preliminary lesion detection are: Output probability vector ,in, represents the probability that the input image belongs to the i-th category, i represents the sequence number of the category, the total number of categories; and ; Set normal tissue category number and lesion category serial number set ; ,in represents the fth lesion category, f represents the sequence number of the lesion category, Indicates the total number of lesion categories; normal tissue category number and lesion category serial number set All are part of category number i; Statistical lesion category number set The probability of lesions in different lesion categories and normal tissue class probability According to the probability of lesions and normal tissue class probability Calculation of lesion determination index , the formula based on is: Preset judgment threshold ; when , then it is preliminarily determined that there is a corresponding lesion category in the image lesions; when < , it is preliminarily determined that there is no corresponding lesion category in the image lesions.

9. The method for intelligent image processing for nuclear medicine according to claim 8, characterized in that: The specific methods to assist in determining whether a lesion exists are: When the lesion determination index is met When the condition is The uncertainty value The formula is as follows: When the lesion determination index is met When the condition is The dispersion value The formula is as follows: in, is the output probability vector The mean value of is in the range of 0 to 1. represents the probability of the fth lesion category, f represents the sequence number of the lesion category, represents the total number of lesion categories; Preset first threshold and the second threshold ; The uncertainty value With the first threshold Compare and divide the dispersion value With the second threshold For comparison: When satisfied and When the condition is met, it means that there are lesions of the corresponding lesion category in the input image; When satisfied and If the condition is met, it means that it cannot be determined whether there is a lesion of the corresponding lesion category in the input image, and the original image data needs to be re-acquired for secondary determination; When satisfied and If the condition is met, it means that it cannot be determined whether there is a lesion of the corresponding lesion category in the input image, and the original image data needs to be re-acquired for secondary determination; When satisfied and When the condition is met, it means that there is no lesion of the corresponding lesion category in the input image.

10. An intelligent image processing system for nuclear medicine, characterized in that: include: A data input module, used for receiving original image data, patient's medical history data BS, age data NL and gender data XB; The feature extraction module is used to extract the feature maps corresponding to different levels in the original image data through the convolutional neural network CNN, calculate the mean and standard deviation of all pixel values ​​in the feature maps corresponding to different levels, calculate the fusion weight coefficients of different levels based on the mean and standard deviation, and obtain the fused feature maps of different levels based on the feature maps and the fusion weight coefficients; The first data processing module is used to analyze and process the feature graphs after fusion at different levels to obtain the input feature vector x; generate the clinical information vector c according to the medical history data BS, age data NL and gender data XB, and fuse the clinical information vector c with the input feature vector x to obtain the comprehensive feature vector , based on the comprehensive feature vector Calculate the output probability vector and conduct preliminary lesion detection; The second data processing module is used to output the probability vector The probability of lesions in and normal tissue class probability Calculation of lesion determination index ; Preset judgment threshold ; The lesion determination index and judgment threshold Compare and calculate the corresponding lesion probability when the preliminary judgment conditions are met The uncertainty value and the corresponding lesion probability The dispersion value , assist in determining whether a lesion exists; The result output module is used to generate a detailed diagnostic report based on the test results.

Images