Real-time reconstruction of three-dimensional topography and volume measurement of bladder based on deep learning

Abstract

The application provides a kind of based on deep learning bladder EIT three-dimensional topography real-time reconstruction and capacity determination method, belong to deep learning technical field, including: according to the three-dimensional CT image generated three-dimensional abdominal simulation model, construct the bladder EIT data set containing abnormal situation;According to the three-dimensional abdominal simulation model, obtain the bladder position information and construct the loss function containing position regularization, and the EIT measurement value in bladder EIT data set is as the input of deep learning network, the mapping relationship between EIT measurement value and bladder topography is established, and the network model after training is saved;The EIT measurement value of actual patient is as the input of network model after training, and the bladder capacity obtained by patient is compared and analyzed in real time with the bladder capacity calculated by CT true value in radiotherapy process, and the bladder capacity of patient is optimized and corrected in real time.Can be accurately reconstructed the three-dimensional topography of bladder without relying on traditional imaging technology, and the determination of bladder capacity is optimized in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and particularly relates to a bladder EIT three-dimensional topography real-time reconstruction and volume measurement method based on deep learning. BACKGROUND

[0002] Bladder topography reconstruction and volume measurement play an important role in precise radiotherapy, especially in lower abdominal radiotherapy. Changes in bladder shape and volume directly affect the positioning and dose distribution of the treatment target. Changes in bladder filling status can cause adjacent normal tissues to be exposed to radiation, increasing the risk of side effects. Therefore, real-time monitoring of bladder shape and volume is crucial to ensure accurate positioning of the target area and reduce radiation damage to healthy tissues.

[0003] Traditional methods of measuring bladder shape and volume mainly rely on clinical procedures and static imaging, such as ultrasound, CT or MRI, etc. Although these methods can provide relatively accurate volume information, there are still some measurement errors due to factors such as patient position, bladder shape changes and measurement errors. Ultrasound is a common method for bladder volume assessment, with advantages such as non-invasiveness, real-time performance and lower cost. However, ultrasound imaging has low accuracy in measuring the shape and volume of the bladder, which is limited by the technical level of the operator, changes in patient position, and the visibility of the bladder wall. In addition, ultrasound measurement usually needs to be performed under static conditions, making it difficult to cope with dynamic changes in the bladder during treatment. CT and MRI imaging can provide high-resolution images of the bladder, allowing more accurate reflection of the shape and volume of the bladder. These methods usually combine with precise positioning of the tumor target area to provide detailed anatomical data for radiotherapy. However, a significant disadvantage of CT and MRI is their inability to capture dynamic changes in the bladder in real time, especially during radiotherapy, where the shape and volume of the bladder change with physiological processes such as urination and water intake.

[0004] Electrical impedance tomography (EIT) is an imaging technique based on the principle of bioelectric impedance, which reconstructs images by measuring the electrical impedance differences of different tissues. The main advantages of EIT are its non-invasiveness, real-time performance and ability to measure in dynamic conditions. Unlike CT and MRI, EIT can provide continuous and real-time updates of bladder topography and volume information, making it particularly promising for dynamic bladder assessment. Despite its great potential, EIT still faces some technical challenges in practical applications. First, due to the small difference in electrical impedance between the liquid and tissue in the bladder, the resolution of electrical impedance imaging is low, making it difficult to accurately reconstruct the shape and volume of the bladder in complex clinical scenarios. Second, EIT imaging technology requires data collection through multiple electrode arrays on the abdominal surface, but due to differences in abdominal wall thickness, tissue type and patient size, EIT signals may be disturbed, affecting imaging accuracy.

[0005] Therefore, it is necessary to design a bladder EIT three-dimensional topography reconstruction and capacity measurement method based on deep learning. SUMMARY

[0006] The application provides a bladder EIT three-dimensional topography real-time reconstruction and capacity measurement method based on deep learning, which combines EIT technology and deep learning algorithm, can accurately reconstruct the three-dimensional topography of the bladder without relying on traditional imaging technology, and can optimize the measurement of the bladder capacity in real time.

[0007] The application provides a bladder EIT three-dimensional topography real-time reconstruction and capacity measurement method based on deep learning, which combines EIT technology and deep learning algorithm, can accurately reconstruct the three-dimensional topography of the bladder without relying on traditional imaging technology, and can optimize the measurement of the bladder capacity in real time.

[0008] S1: According to the collected three-dimensional CT image, the contour point information of the lower abdomen, bladder, pelvis and subcutaneous fat is extracted, a three-dimensional abdominal simulation model is generated, and based on the pre-deployed flexible electrode, EIT measurement values under normal conditions, electrode drop and noise addition conditions are collected to construct a bladder EIT data set containing abnormal conditions;

[0009] S2: According to the three-dimensional abdominal simulation model, the loss function containing position regularization is constructed by obtaining the bladder position information, and the EIT measurement values in the bladder EIT data set are taken as the input of the deep learning network, the mapping relationship between the EIT measurement values and the bladder topography is established, the three-dimensional topography image of the bladder is output, and the network model after training is saved;

[0010] S3: The number of voxels of the reconstructed three-dimensional topography image is calculated, and the relationship between each voxel and the true physical size is combined to obtain the reconstructed bladder capacity;

[0011] S4: The EIT measurement value of the actual patient is taken as the input of the trained network model, and the obtained patient bladder capacity is compared and analyzed in real time with the bladder capacity calculated by the CT true value of the patient during radiotherapy, and the patient bladder capacity is optimized and corrected in real time.

[0012] Preferably, the collected three-dimensional CT image and the bladder EIT data set are divided into training set, validation set and test set according to the ratio of 8:1:1.

[0013] Preferably, before saving the network model after training, it further comprises:

[0014] A deep learning network structure based on V-Net is established;

[0015] On the basis of the network structure, a full connection layer and an SE attention mechanism are added, and a cross-entropy loss function combined with position prior information and a loss function containing position regularization are combined to obtain the network model;

[0016] The network model is trained by using the training set, and meanwhile, a verification error of the trained network model is monitored by using a verification set; when the verification error is in an ascending state, the training of the network model is stopped and saved.

[0017] Preferably, the reconstructed bladder volume is obtained by:

[0018]

[0019] wherein, represents the reconstructed bladder volume; is the volume of each voxel, represents the voxel value of the volume unit (i, j, k) of the voxel, and represents that the voxel is bladder tissue; represents the number of transverse, longitudinal and vertical volume units.

[0020] Preferably, the patient's bladder volume is corrected in real time by:

[0021] An absolute value difference between the patient's bladder volume and the CT true value calculated bladder volume during radiotherapy is obtained;

[0022] If the absolute value difference is greater than a preset threshold, the patient's bladder volume and the CT true value calculated bladder volume are fused by weighting to obtain a required volume, so as to realize real-time optimization and correction;

[0023]

[0024] wherein, is the required volume; represents a weighting factor; represents the patient's bladder volume; represents the CT true value calculated bladder volume.

[0025] Preferably, the method further comprises determining the weighting factor, and specifically comprises:

[0026] The absolute value differences of N1 actual patients are extracted, and a first number M1 of the absolute value differences greater than the preset threshold and a second number M2 of the absolute value differences less than or equal to the preset threshold are counted;

[0027] The absolute value differences of the first number M1 are sorted and analyzed discretely according to the historical measurement time of the actual patients, and the absolute value differences of the second number M2 are sorted and analyzed continuously according to the historical measurement time of the actual patients;

[0028] The discrete degree of each absolute value difference in the first number M1 is determined according to the discrete analysis result, the average degree of the remaining M1-1 absolute value differences except itself is calculated, and the residual error of the corresponding absolute value difference is obtained.

[0029] Sort all absolute value differences in the first quantity M1 by size according to the residuals to obtain a first sorting sequence number;

[0030] Perform a consistency analysis on the first sorting number and the second sorting number sorted by historical measurement time based on all absolute value differences in the first quantity M1 to determine the first number change amount NB, wherein the consistency analysis means that the change between the number of each absolute value difference in the first sorting number and the number of the corresponding absolute value difference in the second sorting number is considered consistent within the range of (-a1, a1), where a1 represents the change amount;

[0031] determining a stability coefficient of the absolute value difference of the second number M2 according to the continuous analysis result, and calculating a product of the stability coefficient and an average value of the absolute value difference of the second number M2;

[0032] Determine a first absolute value of a difference between each absolute value difference in the second number M2 and the product, and a second absolute value of a difference between the product and a preset threshold, and obtain a current average of the first absolute value of the difference and the second absolute value of the difference;

[0033] Calculate the variance of all current means , if the variance Not less than the preset value , then keep the second number M2 unchanged, and obtain the weighting factor based on the first sequence number change number NB, the second number M2, and the first number M1 ;

[0034] If the variance Less than the preset value , then the second quantity M2 is adjusted to obtain the third quantity M3, and based on the first sequence number change quantity NB, the third quantity M3, and the first quantity M1, the weighting factor is obtained. ;

[0035] .

[0036] Preferably, adjusting the second quantity M2 to obtain the third quantity M3 includes:

[0037]

[0038] in, Indicates the rounding symbol.

[0039] Preferably, generating a three-dimensional abdomen simulation model includes:

[0040] The contour points of different parts of each three-dimensional CT image are labeled to obtain a labeled image, and the center of each part is calibrated according to the closed contour line in the labeled image to obtain a center point of each part;

[0041] The center points in the same three-dimensional CT image are sequentially connected according to the structure sequence of the lower abdomen, bladder, pelvis and subcutaneous fat to obtain a first connection line;

[0042] Each first connection line is standardized according to the size mapping relationship with the body structure of the collected object to obtain a first standard line, and all first standard lines are placed in alignment with a center point of a random part as an alignment reference point to obtain a center deviation set of each remaining part to obtain a deviation coefficient;

[0043]

[0044] wherein, represents the deviation coefficient of the x0th remaining part; represents the point density of the e1th unit area obtained after unit area division of the center deviation set of the x0th remaining part; represents the maximum value in all ; represents the area of the unit area corresponding to ; represents the area of the region occupied by the center deviation set of the x0th remaining part; represents the center coordinate point of the e1th unit area based on the center deviation set of the x0th remaining part; represents the center coordinate point of the unit area corresponding to ; respectively represent the variance based on all , all ;

[0045] The average coefficient is obtained by averaging all deviation coefficients, and the image enhancement coefficient is determined from the coefficient-enhancement control table, and the three-dimensional CT image with an image clarity less than a preset clarity is subjected to image enhancement processing according to the image enhancement coefficient to obtain an enhanced image;

[0046] The contour point information of the lower abdomen, bladder, pelvis and subcutaneous fat of each enhanced image and the three-dimensional CT image with an image clarity not less than a preset clarity is extracted to generate a three-dimensional abdominal simulation model.

[0047] Compared with the prior art, the beneficial effects of the present application are as follows:

[0048] Combining the EIT technology and the deep learning algorithm, the three-dimensional appearance of the bladder can be accurately reconstructed without relying on traditional imaging technology, and the determination of the bladder capacity can be optimized in real time.

[0049] Other features and advantages of the present application will be set forth in the following specification, and in part will be apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof.

[0050] The technical solutions of the present application are described in further detail below by means of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0051] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and explain the principles of the present application, and do not constitute a limitation of the present application. In the drawings:

[0052] Figure 1 A flow chart of a deep learning-based bladder EIT three-dimensional appearance real-time reconstruction and capacity determination method in an embodiment of the present application;

[0053] Figure 2 A three-dimensional abdominal simulation model in an embodiment of the present application;

[0054] Figure 3 A structural schematic diagram of a network model in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not constitute a limitation of the present application.

[0056] The present application provides a deep learning-based bladder EIT three-dimensional appearance real-time reconstruction and capacity determination method, as shown in Figure 1 The method comprises the following steps:

[0057] S1: According to the collected three-dimensional CT image, the contour point information of the lower abdomen, bladder, pelvis and subcutaneous fat is extracted, a three-dimensional abdominal simulation model is generated, and based on the pre-deployed flexible electrode, EIT measurement values under normal conditions, electrode falling and noise adding conditions are collected to construct a bladder EIT data set containing abnormal conditions;

[0058] S2: According to the three-dimensional abdominal simulation model, the loss function containing position regularization is constructed by obtaining the bladder position information, and the EIT measurement values in the bladder EIT data set are taken as the input of the deep learning network, the mapping relationship between the EIT measurement values and the bladder appearance is established, the three-dimensional appearance image of the bladder is output, and the network model after training is saved.

[0059] S3: Calculate the number of voxels of the bladder in the reconstructed three-dimensional topographic image, and obtain the bladder capacity after reconstruction in combination with the relationship between each voxel and the true physical size;

[0060] S4: Take the actual patient's EIT measurement as the input of the trained network model, and compare and analyze the patient's bladder capacity obtained with the bladder capacity calculated by the CT true value of the patient in real time during radiotherapy, and optimize and correct the patient's bladder capacity in real time.

[0061] Preferably, the three-dimensional CT image and the bladder EIT data set are divided into training set, validation set and test set in the ratio of 8:1:1 respectively.

[0062] Preferably, before saving the trained network model, it further comprises:

[0063] Establishing a deep learning network structure based on V-Net;

[0064] On the basis of the network structure, a fully connected layer and an SE attention mechanism are added, and a cross-entropy loss function combined with position prior information and a loss function containing position regularization are combined to obtain the network model.

[0065] The training set is used to train the network model, and at the same time, the validation set is used to monitor the validation error of the trained network model, and when the validation error is in an upward state, the training of the network model is stopped and saved.

[0066] Preferably, obtaining the reconstructed bladder capacity comprises:

[0067]

[0068] Wherein, V represents the reconstructed bladder capacity; is the volume of each voxel, V(i,j,k) represents the voxel value of the volume unit (i,j,k) of the voxel, and represents that the voxel is bladder tissue; represents the number of transverse, longitudinal and vertical volume units.

[0069] Preferably, the real-time optimization and correction of the patient's bladder capacity comprises:

[0070] Obtaining the absolute value difference between the patient's bladder capacity and the bladder capacity calculated by the CT true value of the patient during radiotherapy;

[0071] If the absolute value difference is greater than a preset threshold, the patient's bladder capacity and the bladder capacity calculated based on the CT true value are weighted and fused to obtain the required capacity, thereby realizing real-time optimization and correction;

[0072]

[0073] wherein, is the desired capacity; denotes a weighting factor; denotes the patient's bladder capacity; denotes the bladder capacity calculated based on the CT true value.

[0074] In this embodiment, before generating the three-dimensional abdominal simulation model, each data point scanned in the three-dimensional CT image is labeled with the conductivity attribute, so as to more realistically simulate the electrical impedance characteristics of each tissue of the human body. According to the actual conductivity distribution, the conductivity of the bladder is set to 0.5 S / m, the conductivity of the pelvic bone is 0.05 S / m, the conductivity of the subcutaneous fat is 0.2 S / m, and the conductivity of other parts is uniformly set to 0.1 S / m. For the three-dimensional abdominal simulation model, as shown in Figure 2 .

[0075] In this embodiment, before constructing the bladder EIT data set containing abnormal conditions, the following steps are included: arranging 16 flexible electrodes in a semicircle at about 8 cm below the navel, collecting EIT measurement data under different conditions, covering normal conditions, electrode falling conditions, and noise added conditions. In this way, the training data set can have stronger adaptability and cover various possible abnormal conditions.

[0076] In this embodiment, a deep learning network structure based on V-Net is used. V-Net is a convolutional neural network commonly used for medical image segmentation, and its structure has strong spatial information processing capability and is suitable for bladder shape reconstruction, as shown in Figure 3 . In order to improve the network's ability to capture details, a full connection layer (FullConnection, FC) is added before the V-Net network to extract features in the original image and strengthen the global information of feature learning as follows:

[0077]

[0078] wherein, X is the input feature map, z is the global average pooling feature of each channel, W1 and W2 are the weights of the two full connection network layers, is a Sigmoid activation function, denotes an element-wise multiplication operation. By learning the weight of each channel, different importance is assigned to each channel, thereby improving the representation ability of the model.

[0079] Loss function design: the loss function uses a cross entropy loss function (CrossEntropy Loss) that fuses position prior information, which is commonly used for distance measurement between target and predicted results in classification tasks. The public disclosure is as follows:

[0080]

[0081] The cross-entropy loss can help the network optimize the classification accuracy, and has good effect on the bladder appearance reconstruction task.

[0082] Meanwhile, the bladder position prior information is introduced into the loss function, and the model predicted bladder center position is optimized by imposing constraints.

[0083]

[0084] Wherein, x pred and y pred are the bladder center coordinates predicted by the model, and λ is a regularization coefficient for controlling the influence of this item on the overall loss.

[0085] Finally, the loss function of the network can combine the standard cross-entropy loss and the loss of the position prior information to form a comprehensive loss function:

[0086]

[0087] During the training process, the training set is used to train the model, and the performance of the model is monitored in real time through the validation set.

[0088] The beneficial effects of the above technical scheme are that the EIT technology and the deep learning algorithm are combined, the three-dimensional appearance of the bladder can be accurately reconstructed without relying on traditional imaging technology, and the determination of the bladder capacity can be optimized in real time.

[0089] The application provides a bladder EIT three-dimensional appearance real-time reconstruction and capacity determination method based on deep learning, and further comprises determining a weighting factor, specifically comprising:

[0090] The absolute value differences of N1 actual patients are extracted, and the first quantity M1 greater than the preset threshold and the second quantity M2 less than or equal to the preset threshold in N1 absolute value differences are counted;

[0091] The absolute value difference values in the first quantity M1 are sorted and discretely analyzed according to the historical measurement time of the actual patient, and the absolute value difference values in the second quantity M2 are sorted and continuously analyzed according to the historical measurement time of the actual patient;

[0092] The discrete degree of each absolute value difference value in the first quantity M1 is determined according to the discrete analysis result, the average degree of the remaining M1-1 absolute value difference values except itself is calculated, and a residual of the corresponding absolute value difference value is obtained;

[0093] All the absolute value difference values in the first quantity M1 are sorted according to the residual, and a first sorting sequence number is obtained;

[0094] The first sorting sequence number and a second sorting sequence number sorted according to the historical measurement time of the absolute value difference values in the first quantity M1 are uniformly analyzed to determine a first sequence number variation quantity NB, wherein the uniform analysis means that the variation of the sequence number of each absolute value difference value in the first sorting sequence number and the sequence number of the corresponding absolute value difference value in the second sorting sequence number is within a range of (-a1, a1) and is considered uniform, wherein a1 represents the variation quantity;

[0095] The stability coefficient of the absolute value difference values in the second quantity M2 is determined according to the continuous analysis result, and the product of the stability coefficient and the average value of the absolute value difference values in the second quantity M2 is calculated;

[0096] The first difference absolute value of each absolute value difference value in the second quantity M2 and the product, and the second difference absolute value of the product and a preset threshold value are determined, and the current average value of the first difference absolute value and the second difference absolute value is obtained;

[0097] The variance of all the current average values is calculated If the variance is not less than a preset value , the second quantity M2 is kept unchanged, and a weighting factor is obtained based on the first sequence number variation quantity NB, the second quantity M2, and the first quantity M1;

[0098] If the variance is less than a preset value , the second quantity M2 is adjusted to obtain a third quantity M3, and a weighting factor is obtained based on the first sequence number variation quantity NB, the third quantity M3, and the first quantity M1;

[0099] .

[0100] Preferably, the adjustment of the second quantity M2 to obtain the third quantity M3 includes:

[0101]

[0102] wherein, represents a rounding up symbol.

[0103] In this embodiment, the absolute value difference of the actual patient refers to the absolute value difference between the bladder capacity of the patient and the CT true value calculated bladder capacity of the patient during radiotherapy, and N1 is greater than 200.

[0104] In this embodiment, the preset threshold value is pre-set and is 30ml.

[0105] In this embodiment, the historical measurement time refers to the time measured by using the model, and there is a sequential measurement order, so the sorting can be realized based on the measurement time.

[0106] In this embodiment, the discrete analysis result refers to the average processing of the discrete degree of each absolute value difference in the first number M1 according to the absolute value of the difference between the absolute value difference and the remaining each absolute value difference in the first number M1.

[0107] In this embodiment, the residual error = the discrete degree of the corresponding absolute value difference in the first number M1 - the average degree of the remaining M1-1 absolute value differences excluding itself.

[0108] In this embodiment, for example, the second sorting sequence number sorted according to the historical measurement time is: u1--1, u2--2, u4--3, u7--4, and the first sorting sequence number is: u2--1, u4--2, u1--3, u7--4.

[0109] In this embodiment, the sequence number of u7 is unchanged, the sequence number of u4 is moved left by 1 bit, the sequence number of u2 is moved left by 1 bit, and the sequence number of u1 is moved right by 2 bits, at this time, assuming that the value of a1 is 1, that is, only u1 meets the requirements, at this time, the first sequence number variation quantity NB is 1.

[0110] In this embodiment, -a1 represents moving left by a1 bits, a1 represents moving right by a1 bits, and a1 is a positive integer.

[0111] In this embodiment, the stability coefficient = 1 - (the maximum value of the absolute value difference of the second number M2 - the minimum value of the absolute value difference of the second number M2) / the fitting coefficient of the absolute value difference of the second number M2 determined by the continuous analysis result.

[0112] In this embodiment, the product = stability coefficient * average value of the absolute value difference of the second number M2.

[0113] In this embodiment, the lower average value = (first difference absolute value + second difference absolute value) / 2.

[0114] The beneficial effects of the above technical solutions are: by taking the N1 absolute value differences as the basis for data analysis, and then subsequently determining the weighting factors in different situations through continuous analysis and discrete analysis, the sequence number variation is determined based on the discrete analysis results as the main analysis standard, and the product of the stability coefficient under the continuous analysis result is combined to determine the variance of the current average value, and then through intelligent classification judgment, the reliable acquisition of the weighting factor is realized, providing a basis for subsequent capacity optimization.

[0115] The application provides a bladder EIT three-dimensional appearance real-time reconstruction and capacity measurement method based on deep learning, generates a three-dimensional abdominal simulation model, and comprises the following steps:

[0116] The contour points of different parts of each three-dimensional CT image are labeled to obtain a labeled image, and the center points of each part are calibrated according to the closed contour lines of each part in the labeled image;

[0117] The center points in the same three-dimensional CT image are sequentially connected according to the structure sequence of the lower abdomen, the bladder, the pelvis and the subcutaneous fat to obtain a first connection line;

[0118] Each first connection line is standardized according to the size mapping relationship with the body structure of the collected object to obtain a first standard line, and all first standard lines are placed in alignment with a random center point of a part as an alignment reference point, and the center deviation set of each remaining part is obtained to obtain a deviation coefficient;

[0119]

[0120] Wherein, The deviation coefficient of the x0th remaining part is represented by x0. The point density of the e1th unit area obtained after unit area division of the center deviation set of the x0th remaining part is represented by e1. The maximum value of all is represented by max. The area of the unit area corresponding to is represented by area. The area of the region occupied by the center deviation set of the x0th remaining part is represented by area. The center coordinate point of the e1th unit area based on the center deviation set of the x0th remaining part is represented by e1. The center coordinate point of the unit area corresponding to is represented by e1. The variance based on all , all is represented by var.

[0121] The average coefficient is obtained by averaging all deviation coefficients, and the image enhancement coefficient is determined from the coefficient-enhancement correspondence table, and the three-dimensional CT image with an image definition less than a preset definition is subjected to image enhancement processing according to the image enhancement coefficient to obtain an enhanced image.

[0122] The contour point information of the lower abdomen, bladder, pelvis and subcutaneous fat is extracted from each enhanced image and the three-dimensional CT image with an image definition not less than a preset definition to generate a three-dimensional abdominal simulation model.

[0123] In this embodiment, the first connection line is obtained in the order from top to bottom and from left to right.

[0124] In this embodiment, the labeling of the contour points in the three-dimensional CT image can be directly obtained by using the related network of contour extraction, which belongs to the prior art. However, during the labeling process, certain errors may occur due to the unclear image, etc. Therefore, the center point of each part in the standard image is obtained to draw the connection line, and then the alignment and placement are performed to obtain the center deviation set.

[0125] In this embodiment, the center calibration refers to the calibration based on the center point of the closed contour line of the corresponding part.

[0126] In this embodiment, since the collected three-dimensional CT images are for different patients, and the ages of the patients have certain differences, the sizes of the corresponding parts of the patients at different ages are different. Therefore, it is necessary to rely on the size mapping relationship to standardize the part size to facilitate the subsequent alignment and placement of the alignment reference points based on the parts.

[0127] In this embodiment, for example, the center point of the lower abdomen is used as the alignment reference point for alignment and placement to obtain the center deviation set based on the bladder, pelvis and subcutaneous fat structure, and the center deviation set is a position deviation set of the center point of the corresponding part. Therefore, the analysis of the set can obtain the deviation coefficient.

[0128] In this embodiment, the average coefficient is the average value of all deviation coefficients.

[0129] In this embodiment, the coefficient-enhancement correspondence table contains the image enhancement coefficient matching the average coefficient, and the image enhancement processing according to the enhancement coefficient is based on the matching obtained from the enhancement coefficient-enhancement mode table. The table contains the image enhancement mode under different enhancement coefficients, and the enhancement processing can be directly performed to obtain the enhanced image.

[0130] In this embodiment, the last extracted contour point information is fused at the same position to obtain the three-dimensional abdominal simulation model.

[0131] The beneficial effects of the above technical scheme are: the contour points of different three-dimensional CT images are labeled to obtain the center points of each part based on the closed contour line, and then the first connection line is obtained by sequentially connecting, thereby providing a basis for subsequent alignment processing; and the adjustment of the size mapping relationship ensures the alignment standardization, ensures the reliability of the subsequent deviation coefficient acquisition, provides a real basis for image enhancement, thereby realizing the rationality of image enhancement, and indirectly improving the accuracy of the determination of the optimized bladder capacity.

[0132] Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the present application claims and their equivalents, it is intended to include these modifications and changes in the present application.

Claims

1. A method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning, characterized in that: include: S1: Based on the collected 3D CT images, the contour points of the lower abdomen, bladder, pelvis, and subcutaneous fat are extracted to generate a 3D abdominal simulation model. Simultaneously, EIT measurements under normal conditions, electrode drop conditions, and noise conditions are collected using pre-deployed flexible electrodes to construct a bladder EIT dataset that includes abnormal conditions. S2: Obtain bladder position information based on the 3D abdominal simulation model and construct a loss function including position regularization. The EIT measurements in the bladder EIT dataset are used as input to the deep learning network. A mapping relationship between the EIT measurements and bladder topography is established, and a 3D topographic image of the bladder is output. The trained network model is then saved. S3: Calculate the number of voxels of the bladder in the reconstructed 3D topographic image, and combine the relationship between each voxel and the real physical size to obtain the reconstructed bladder capacity; S4: The actual patient EIT measurement value is used as the input of the trained network model, and the obtained patient bladder capacity is compared and analyzed in real time with the bladder capacity calculated from the patient's CT true value during radiotherapy, and the patient's bladder capacity is optimized and corrected in real time; Among them, before saving the trained network model, it also includes: Establish a deep learning network structure based on V-Net; A fully connected layer is added before the network structure, an SE attention mechanism is added before each deconvolution layer, and a cross entropy loss function that integrates position prior information and a loss function that includes position regularization are combined to obtain a network model; The network model is trained using the training set, and at the same time, the verification error of the trained network model using the verification set is monitored. When the verification error is in an increasing state, the training of the network model is stopped and saved.

2. The method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning according to claim 1, characterized in that: The collected three-dimensional CT images and bladder EIT datasets were divided into training set, validation set and test set in the ratio of 8:1:

1.

3. The method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning according to claim 1, characterized in that: Obtain reconstructed bladder capacity, including: in, represents the reconstructed bladder capacity; δV is the volume of each voxel, y ijk The voxel value of the volume unit (i, j, k) represents the voxel, indicating that the voxel is bladder tissue; H, W, and D represent the horizontal number, vertical number, and vertical number of the volume unit.

4. The method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning according to claim 1, characterized in that: Real-time optimization and correction of the patient's bladder capacity, including: Obtain the absolute value difference between the patient's bladder capacity and the bladder capacity calculated based on the patient's CT true value during radiotherapy; If the absolute value difference is greater than a preset threshold, the patient's bladder capacity is weightedly fused with the bladder capacity calculated based on the CT true value to obtain the required capacity, thereby achieving real-time optimization correction; in, is the required capacity; represents the weighting factor; Indicates the patient's bladder capacity; Indicates the bladder capacity calculated based on the true CT value.

5. The method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning according to claim 4, characterized in that: Also includes: Determine the weighting factors, including: Extracting absolute value differences of N1 actual patients, and counting a first number M1 of the N1 absolute value differences that are greater than a preset threshold and a second number M2 that are less than or equal to the preset threshold; The absolute value differences of the first quantity M1 are sorted according to the historical measurement time of the actual patients and a discrete analysis is performed. Meanwhile, the absolute value differences of the second quantity M2 are sorted according to the historical measurement time of the actual patients and a continuous analysis is performed. Determine the degree of dispersion of each absolute value difference in the first number M1 according to the discrete analysis result, calculate the average degree of the remaining M1-1 absolute value differences after dividing the absolute value difference, and obtain the residual of the corresponding absolute value difference; Sort all absolute value differences in the first quantity M1 by size according to the residuals to obtain a first sorting sequence number; Perform a consistency analysis on the first sorting number and the second sorting number sorted by historical measurement time based on all absolute value differences in the first quantity M1 to determine the first number change amount NB, wherein the consistency analysis means that the change between the number of each absolute value difference in the first sorting number and the number of the corresponding absolute value difference in the second sorting number is considered consistent within the range of (-a1, a1), where a1 represents the change amount; determining a stability coefficient of the absolute value difference of the second number M2 according to the continuous analysis result, and calculating a product of the stability coefficient and an average value of the absolute value difference of the second number M2; Determine a first absolute value of a difference between each absolute value difference in the second number M2 and the product, and a second absolute value of a difference between the product and a preset threshold, and obtain a current average of the first absolute value of the difference and the second absolute value of the difference; Calculate the variance of all current means , if the variance Not less than the preset value , then keep the second number M2 unchanged, and obtain the weighting factor based on the first sequence number change number NB, the second number M2, and the first number M1 ; If the variance Less than the preset value , then the second quantity M2 is adjusted to obtain the third quantity M3, and based on the first sequence number change quantity NB, the third quantity M3, and the first quantity M1, the weighting factor is obtained. ; 。 6. The method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning according to claim 5, characterized in that: Adjusting the second quantity M2 to obtain a third quantity M3 includes: in, Indicates the round-up symbol.

7. The method for real-time reconstruction of bladder EIT three-dimensional morphology and capacity measurement based on deep learning according to claim 1, characterized in that: Generate a 3D abdomen simulation model, including: Annotating the contour points of different parts of each 3D CT image to obtain an annotation map, and performing center calibration based on the closed contour line of each part in the annotation map to obtain the center point of each part; The center points in the same three-dimensional CT image are sequentially connected according to the structural order of the lower abdomen, bladder, pelvis and subcutaneous fat to obtain a first connection line; Each first connection line is normalized according to the size mapping relationship with the body structure of the collected subject to obtain a first standard line, and all first standard lines are aligned with the center point of a random part as the alignment reference point, and the center deviation set of each remaining part is obtained to obtain the deviation coefficient; in, represents the deviation coefficient of the x0th residual part; represents the point density of the e1th unit area obtained after dividing the center deviation set of the x0th residual part into unit areas; Indicates all The maximum value in ; Represents The area of ​​the corresponding unit region; represents the area occupied by the center deviation set of the x0th residual part; represents the center coordinate point of the e1th unit area based on the center deviation set of the x0th residual part; Represents The center coordinate point of the corresponding unit area; 、 Respectively based on all ,all variance; Averaging all deviation coefficients to obtain an average coefficient, determining an image enhancement coefficient from a coefficient-enhancement comparison table, and performing image enhancement processing on a three-dimensional CT image having an image clarity less than a preset clarity according to the image enhancement coefficient to obtain an enhanced image; The contour point information of the lower abdomen, bladder, pelvis and subcutaneous fat is extracted from each enhanced image and the three-dimensional CT image with an image clarity not less than a preset clarity to generate a three-dimensional abdominal simulation model.

Images