CN116563224A - Radiomics placenta accreta prediction method and device based on deep semantic features - Google Patents

Abstract

The present application discloses a radiomics placenta accreta prediction method and device based on deep semantic features. The method includes acquiring placental image data; extracting deep semantic features and radiomics features of the placental image data; The deep semantic feature and the radiomics feature are screened to obtain the target feature; based on the target feature, the implantation category corresponding to the placenta image data is predicted. In this application, by combining radiomics features and deep semantic features, feature information with rich dimensions and levels can be obtained, and then placenta accreta prediction can be performed based on feature information, which can improve the accuracy of placenta accreta prediction.

Description

Radiomics placenta accreta prediction method and device based on deep semantic features

technical field

The present application relates to the technical field of biomedical engineering, in particular to a radiomics placenta accreta prediction method and device based on deep semantic features.

Background technique

Placenta Accreta Spectrum (PAS) refers to a group of diseases in which the decidua between the placenta and the uterus is underdeveloped, and trophoblast cells abnormally invade the myometrium. Pregnant women with PAS face serious risks after delivery, such as retained placenta, life-threatening bleeding, and even death. The disease is also related to premature birth and low birth weight infants.

Ultrasound imaging is widely used for PAS identification due to its advantages of flexibility, low cost, and harmlessness to mothers and infants. However, when ultrasound images are used for PAS identification, the detection rate of PAS is low due to the high dependence of ultrasound imaging on the operator's technical ability, poor repeatability, and susceptibility to maternal obesity, intestinal gas artifacts in pregnant women, and fetal skull. reduce.

Thereby prior art still needs to improve and improve.

Contents of the invention

The technical problem to be solved in the present application is to provide a radiomics placenta accreta prediction method and device based on deep semantic features in view of the deficiencies in the prior art.

In order to solve the above technical problems, the first aspect of the embodiment of the present application provides a radiomics placenta accreta prediction method based on deep semantic features, the method comprising:

Extracting deep semantic features and radiomics features of the placenta image data;

Screening the deep semantic features and the radiomics features to obtain target features;

The implantation type corresponding to the placenta image data is predicted based on the target feature.

In the radiomics placenta accreta prediction method based on deep semantic features, the placenta image data is placental MRI data.

The radiomics placenta accreta prediction method based on deep semantic features, wherein the screening of the deep semantic features and the radiomics features to obtain target features specifically includes:

Carrying out homogeneity of variance verification on the deep semantic features and the radiomics features, to obtain the verification value of each deep semantic feature in the deep semantic features and the calibration value of each radiomics feature in the radiomics features. check value;

Screening the deep semantic features based on the verification value of the deep semantic features to obtain the target deep semantic features, and screening the radiomics features based on the verification value of each radiomics feature to obtain the target radiomics features;

The target deep semantic feature and the target radiomics feature are spliced to obtain the target feature.

The radiomics placenta accreta prediction method based on deep semantic features, wherein the extracting the deep semantic features of the placenta image data specifically includes:

Input the placenta image data into a preset semantic feature extraction module, and determine the depth semantic features of the placenta image data through the feature extraction module;

Wherein, the semantic feature extraction module includes an encoder and an adaptive average pooling layer, the encoder includes a first feature extraction unit and several cascaded second feature extraction units, the first feature extraction unit and the first feature extraction unit are located at the front Two feature extraction units are connected, and the last second feature extraction unit is connected with the adaptive average pooling layer; the second feature extraction unit includes a maximum pooling layer and a first feature extraction unit; the first feature extraction unit The feature extraction unit includes two cascaded convolutional blocks, and the convolutional block includes sequentially cascaded convolutional layers, batch normalization layers, and activation function layers.

The radiomics placenta accreta prediction method based on deep semantic features, wherein the determination process of the semantic feature extraction module specifically includes:

training the first preset network model based on a preset segmentation training set to obtain a segmentation network model, wherein the segmentation network model includes the encoder;

The encoder of the segmentation network model is extracted, and the encoder is connected with an adaptive average pooling layer to form the semantic feature extraction module.

The radiomics placenta accreta prediction method based on deep semantic features, wherein the segmentation network model also includes a decoder; the decoder includes a first upsampling unit, several cascaded second upsampling units, and volume product unit; the first upsampling unit is connected with the last second feature extraction unit; the first upsampling unit is connected with the first second upsampling unit, and the last second upsampling unit is connected with The convolution units are connected; in several second feature extraction units, each second feature extraction unit except the last second feature extraction unit is in one-to-one correspondence with each second upsampling unit and is skip-connected; the first feature extraction unit and The convolution unit skip connection; wherein, the second upsampling unit includes a first feature extraction unit and an upsampling layer; the convolution unit includes a convolution layer and a first feature extraction unit.

The radiomics placenta accreta prediction method based on deep semantic features, wherein the prediction of the accreta category corresponding to the placenta image data based on the target features specifically includes:

Inputting the target features into a trained classifier, and predicting the implantation category corresponding to the placenta image data through the classifier;

Wherein, the training process of the classifier specifically includes:

For each classification training sample in the preset classification training set, extract the deep semantic feature corresponding to the classification training sample based on the semantic feature extraction module, and extract the radiomics features of the classification training sample;

Screening the deep semantic features and the radiomics features to obtain target features;

Input the second preset network model based on the target characteristics, determine the predicted implant category through the second preset network model, and classify the second preset network model based on the predicted implant category and the labeled implant category corresponding to the classified training samples. Two preset network models are trained to obtain the classifier.

The second aspect of the embodiment of the present application provides a radiomics placenta accreta prediction device based on deep semantic features, the device comprising:

A feature extraction module, configured to extract deep semantic features and radiomics features of the placenta image data;

A screening module, configured to screen the deep semantic features and the radiomics features to obtain target features;

A classification module, configured to predict the implantation category corresponding to the placenta image data based on the target features.

The third aspect of the embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to Realize the steps in the radiomics placenta accreta prediction method based on deep semantic features as described above.

The fourth aspect of the embodiment of the present application provides a terminal device, which includes: a processor, a memory, and a communication bus; a computer-readable program executable by the processor is stored in the memory;

The communication bus realizes connection and communication between the processor and the memory;

When the processor executes the computer-readable program, the steps in any one of the methods for predicting placenta accreta based on radiomics based on deep semantic features are realized.

Beneficial effects: Compared with the prior art, the present application provides a radiomics placenta accreta prediction method and device based on depth semantic features, the method includes acquiring placenta image data; extracting the depth semantics of the placenta image data features and radiomics features; screening the deep semantic features and the radiomics features to obtain target features; predicting the implantation category corresponding to the placenta image data based on the target features. In this application, by combining radiomics features and deep semantic features, feature information with rich dimensions and levels can be obtained, and then placenta accreta prediction can be performed based on feature information, which can improve the accuracy of placenta accreta prediction.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can also be obtained based on these drawings under the premise of not conforming to creative work.

Fig. 1 is a flow chart of the radiomics placenta accreta prediction method based on deep semantic features provided by the present application.

Figure 2 is a schematic diagram of the structure of the deep semantic feature extraction module.

Fig. 3 is a schematic structural diagram of the first feature extraction unit.

Fig. 4 is a schematic diagram of the training process of the classification network model.

Figure 5 is a schematic diagram of the structure of the segmentation network model.

Fig. 6 is a structural principle diagram of the radiomics placenta accreta prediction device based on deep semantic features provided by the present application.

FIG. 7 is a schematic structural diagram of a terminal device provided by the present application.

Detailed ways

The present application provides a radiomics placenta accreta prediction method and device based on deep semantic features. In order to make the purpose, technical solution and effect of the present application clearer and clearer, the present application will be described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described here are only used to explain the present application, not to limit the present application.

Those skilled in the art will understand that unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the word "comprising" used in the specification of the present application refers to the presence of said features, integers, steps, operations, elements and/or components, but does not exclude the presence or addition of one or more other features, Integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Additionally, "connected" or "coupled" as used herein may include wireless connection or wireless coupling. The expression "and/or" used herein includes all or any elements and all combinations of one or more associated listed items.

Those skilled in the art can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meanings as commonly understood by those of ordinary skill in the art to which this application belongs. It should also be understood that terms, such as those defined in commonly used dictionaries, should be understood to have meanings consistent with their meaning in the context of the prior art, and unless specifically defined as herein, are not intended to be idealized or overly Formal meaning to explain.

It should be understood that the sequence numbers and sizes of the steps in this embodiment do not imply the order of execution, and the execution order of each process is determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiment of the present application.

After research, it was found that placenta accreta Spectrum (PAS) refers to a group of diseases in which the decidua between the placenta and the uterus is underdeveloped, and the trophoblast cells abnormally invade the myometrium. Pregnant women with PAS face serious risks after delivery, such as retained placenta, life-threatening bleeding, and even death. According to recent reports and statistics, the overall incidence of placenta accreta has reached 0.78%. will increase significantly.

Placenta accreta can be divided into placenta accreta, placenta accreta, and placenta percreta according to the depth of placenta accreta. Among them, placenta percreta may cause maternal hemorrhage and related complications, such as multi-organ Failure or functional failure can lead to severe maternal and neonatal death. Therefore, if the patients with PAS can be predicted in a timely and accurate manner, a suitable surgical plan can be formulated before the operation, the operation can be performed more safely, the risk of childbirth can be reduced, more effective care can be provided for the parturient, and the clinical outcome can be improved.

At present, imaging examinations such as ultrasound imaging and magnetic resonance imaging (Magnetic Resonance Imaging, MRI) are widely used as methods to identify PAS. Among them, ultrasound imaging has the advantages of flexibility, low cost, and harmlessness to mothers and infants. Therefore, ultrasound imaging is the preferred imaging method for the diagnosis of PAS. However, ultrasound imaging is highly dependent on the technical ability of the operator, has poor repeatability, and is also susceptible to interference from factors such as maternal obesity, intestinal gas artifacts in pregnant women, and fetal head and skull, which reduce the detection rate of PAS. MRI, on the other hand, has the advantage of providing panoramic images across different acquisition planes, without interference from maternal size, intestinal gas, or placental position, and can better assess the location of peripheral atherosclerotic areas involved and the damage of adjacent organs , has a unique advantage in identifying PAS.

However, at present, when using MRI to predict PAS, it is generally based on the subjective prediction of doctors, which will make the prediction results affected by various subjective factors (such as clinician experience, status, etc.), resulting in poor objectivity of PAS prediction and low accuracy.

In order to solve the above problems, in the embodiment of the present application, the placenta image data is obtained; the depth semantic features and radiomics features of the placenta image data are extracted; the depth semantic features and the radiomics features are screened to obtain Obtaining target features; predicting the implantation category corresponding to the placenta image data based on the target features. In this application, by combining radiomics features and deep semantic features, feature information with rich dimensions and levels can be obtained, and then placenta accreta prediction can be performed based on feature information, which can improve the accuracy of placenta accreta prediction.

The content of the application will be further explained by describing the embodiments below in conjunction with the accompanying drawings.

This embodiment provides a radiomics placenta accreta prediction method based on deep semantic features, as shown in Figure 1, the radiomics placenta accreta prediction method based on deep semantic features specifically includes:

S10. Extracting deep semantic features and radiomics features of the placenta image data.

Specifically, the placenta image data may be placenta image data with placenta accreta, or placenta image data without placenta accreta. Wherein, the placenta image data may be a real-time image collected by an image collection device, or a non-real-time image obtained by reading a storage device, or a non-real-time image sent by an external device. In an implementation manner, the placental image data is placental MRI (Magnetic Resonance Imaging, nuclear magnetic resonance imaging) image data, that is, the placental image data is a nuclear magnetic resonance image collected by a nuclear magnetic resonance imaging device.

The deep semantic features are high-level features extracted through deep learning, and the radiomics features are image features extracted from medical images, which are used to complete the transformation of image data into clinical data information. It is understandable that after the placental image data is obtained, two feature extractions are performed on the placental image data, one of which is to extract deep semantic features through deep learning, and the other is to extract image groups based on the placenta region in the placental image data. academic features.

The radiomics feature includes one or more of the size, volume, shape, texture feature, histogram data distribution and information entropy of the placental region. The radiomics features can be extracted through the radiomics feature extraction module, wherein the radiomics feature extraction module can be determined through parameter configuration. In addition, when extracting radiomics features, the region of interest of the placenta image data can be pre-determined, wherein the region of interest can be delineated by a doctor or segmented by a trained segmentation network model.

The deep semantic feature is used to reflect the detailed information carried by the placenta image data, wherein the deep semantic feature can be obtained by a trained feature extraction module. Correspondingly, the extracting the depth semantic features of the placenta image data specifically includes:

The placenta image data is input into a preset semantic feature extraction module, and the deep semantic feature of the placenta image data is determined through the feature extraction module.

Specifically, the feature extraction module is pre-trained and used to extract deep semantic features of placenta image data, that is, the input item of the semantic feature extraction module is placenta image data, and the output is deep semantic features. Wherein, the feature extraction module may include an encoder, or may include an encoder and an adaptive average pooling layer, the encoder is used to extract semantic features, and the adaptive average pooling layer is used to extract the semantic features extracted by the encoder Dimensionality reduction to reduce the inability to feature and redundant features in the deep semantic feature to reduce the parameter calculation of the semantic feature extraction module, and also integrate the global information to improve the robustness of the classification network model.

In one implementation, the semantic feature extraction module includes an encoder and an adaptive average pooling layer, the encoder includes a first feature extraction unit and several cascaded second feature extraction units, the first feature extraction unit and The second feature extraction unit located at the front is connected, and the second feature extraction unit located at the end is connected with the adaptive average pooling layer; the second feature extraction unit includes a maximum pooling layer and a first feature extraction unit; The first feature extraction unit includes two cascaded convolutional blocks, and the convolutional block includes sequentially cascaded convolutional layers, batch normalization layers, and activation function layers. Among them, the batch normalization layer can use Batch Normalization, and the activation function layer can use the Relu activation function layer. The feature extraction module in this implementation uses multiple convolutional layers with small convolution kernels (3×3) instead of convolutional layers with large convolution kernels, which reduces network parameters and is equivalent to performing more nonlinear mappings , improving the fitting and expressive capabilities of the network.

For example, as shown in Figure 2 and Figure 3, the encoder includes a first feature extraction unit and four cascaded second feature extraction units, wherein the image size of the placenta image data is 128*128*96, the first feature The image scale of the output item of the extraction unit is 128*128*96, the image scale of the output item of the first second feature extraction unit is 64*64*48, and the output item of the second feature extraction unit located at the second The image scale is 32*32*24; the image scale of the output item of the third feature extraction unit is 16*16*12; the image scale of the output item of the fourth feature extraction unit is 8 *8*6, to get 8×8×6×512-dimensional feature vectors. In addition, the adaptive average pooling layer reduces the 8×8×6×512-dimensional feature vector to 1×512 dimension, so that the average value of the feature map of each channel will be output through the adaptive average pooling layer, It reduces the amount of parameter calculation and integrates global information, which is more robust.

In an implementation manner, the determination process of the semantic feature extraction module specifically includes:

training the first preset network model based on a preset segmentation training set to obtain a segmentation network model, wherein the segmentation network model includes the encoder;

The encoder of the segmentation network model is extracted, and the encoder is connected with an adaptive average pooling layer to form the semantic feature extraction module.

Specifically, the segmented training set includes several segmented training samples, each segmented training sample is placental image data and carries a placenta label, and the segmented network model is obtained by training the first preset network model through the segmented training set, wherein, The model structure of the first preset network model is the same as the model structure of the segmented network model, and the difference between the two is that the model parameters are different, wherein the first preset network model adopts initial model parameters, and the model parameters of the segmented network model adopt trained network model parameters. In addition, in one implementation, the segmentation network model includes an encoder and a decoder, wherein the network structure of the encoder is the same as that of the above-mentioned encoder; the decoder includes a first upsampling unit, several cascaded The second upsampling unit and the convolution unit; the first upsampling unit is connected to the last second feature extraction unit; the first upsampling unit is connected to the first second upsampling unit, located at The last second upsampling unit is connected to the convolution unit; each second feature extraction unit in several second feature extraction units except the last second feature extraction unit is in one-to-one correspondence with each second upsampling unit and jumps Connection; the first feature extraction unit is skipped connected with the convolution unit; wherein, the second upsampling unit includes a first feature extraction unit and an upsampling layer; the convolution unit includes a convolution layer and a first feature extraction unit. The segmentation network model in this embodiment adopts the U-Net segmentation network based on the encoder-decoder structure, and the segmentation network model replaces the volume of the large convolution kernel by using the convolution layer of multiple small convolution kernels (3×3). Multilayer reduces the network parameters, and at the same time, it is equivalent to performing more nonlinear mappings, which improves the fitting and expression capabilities of the network. Among them, the encoder consists of 10 convolutional layers and 4 pooling layers, as well as several BatchNorm layers and Relu activation layers. The encoder gradually reduces the size of the feature map; the decoder uses upsampling and skip connections to encode The feature map output by the filter is restored to a size close to the original image.

Further, each network layer in the encoder and decoder randomly sets the weight initial value, so that the weight initially satisfies the normal distribution, which is conducive to speeding up the convergence of the network, in which the convolutional layer is directly initialized randomly, while the BN normalization layer is Set the weight to 1 and the bias to 0; in the error backpropagation training process, DICE loss is used as the loss function to ensure that the deep learning model can extract more image features.

For example: as shown in Figure 3 and Figure 5, the decoder includes three cascaded second upsampling units, the image scale of the output item of the first upsampling unit is 16*16*12; The image scale of the output item of the second upsampling unit is 32*32*24; the image scale of the output item of the second upsampling unit located in the second position is 64*64*48; the second upsampling unit located in the third position The image scale of the output item of the convolution unit is 128*128*96, and the image scale of the output item of the convolution unit is 128*128*96, where the second upsampling unit located in the first position and the second feature located in the third position The extraction unit is skipped and connected, the second upsampling unit at the second position is skipped with the second feature extraction unit at the second position; the second upsampling unit at the third position is connected with the second feature extraction unit at the first position skip connection; the convolution unit is skip connected with the first feature extraction unit.

S20. Filter the deep semantic features and the radiomics features to obtain target features.

Specifically, after the deep semantic features and radiomics features are obtained, since the deep semantic features and radiomics features will carry redundant information, the deep semantic features and radiomics features can be screened and spliced to obtain Generating target features, wherein the target features include part of deep semantic features and part of radiomics features.

Based on this, the screening of the deep semantic feature and the radiomics feature to obtain the target feature specifically includes:

Carrying out homogeneity of variance verification on the deep semantic features and the radiomics features, to obtain the verification value of each deep semantic feature in the deep semantic features and the calibration value of each radiomics feature in the radiomics features. check value;

Screening the deep semantic features based on the verification value of the deep semantic features to obtain the target deep semantic features, and screening the radiomics features based on the verification value of each radiomics feature to obtain the target radiomics features;

The target deep semantic feature and the target radiomics feature are spliced to obtain the target feature.

Specifically, the target depth semantic feature is included in the depth semantic feature, and the target radiomics feature includes the radiomics feature. It can be understood that both the depth semantic feature and the radiomics feature include multiple features, and the target depth semantic feature includes depth Some of the semantic features, the target radiomics features include some of the radiomics features. For example, the target deep semantic feature includes 2 deep semantic features in the deep semantic feature, and the target radiomics feature includes 5 radiomics features in the radiomics feature. In this way, on the one hand, the highly significant features in the deep semantic features and radiomics features can be preserved, and at the same time, the amount of features can be reduced, thereby reducing the amount of model parameters.

The variance homogeneity check is an F test, that is, after obtaining the ratio of the variance between groups and the variance within groups, the features are sorted, and the target radiomics features and target depth semantic features are selected according to the sorting order. Among them, the screening process of target deep semantic features and target radiomics features is the same, and here we take target deep semantic features as an example for illustration.

Bring the extracted deep semantic feature and the implanted gold standard into the formula for calculating the standard deviation, and obtain the square value S _x of the standard deviation of the deep semantic feature and the implanted gold standard S _label , where, used to calculate the standard deviation The formula for is:

Among them, x is the sample, x is the sample mean, and n is the number of samples.

Substitute the square value S _x of the standard deviation and the implanted gold standard S _label into the F value calculation formula to obtain the F value corresponding to the deep semantic feature, where the F value calculation formula is:

Based on the above process, the F value corresponding to each deep semantic feature is obtained, and the smaller the F value, the more significant the significant difference between the deep semantic feature and the implanted gold standard is. Based on this, the F values are sorted from small to large, and a certain number (for example, 150) of candidate deep semantic features are selected from the front to the back. Finally, the candidate deep semantic features are put into the Lasso algorithm for screening to obtain the target deep semantic features (for example, 150 deep learning semantic features are screened to obtain 2).

Further, after the target depth semantic features and target radiomics features are obtained, splicing can be performed in the order of target depth semantic features-target radiomics features, or in the order of target radiomics features-target depth semantic features to splice. In this embodiment, stitching is performed in the order of target deep semantic features-target radiomics features.

S30. Predict an implantation category corresponding to the placenta image data based on the target feature.

Specifically, the accreta types are divided into placenta accreta type, placenta accreta type and placenta percreta type according to the depth of placenta accreta. Thus, the implantation types corresponding to the placenta image data can be placenta accreta type, placenta accreta type, and placenta accreta type. One of intrusion and placenta penetration. In addition, the placenta image data may have no placenta accreta, so the implantation category corresponding to the placenta image data may also be non-placenta accreta.

In an implementation manner, the implantation category may be determined by a classifier, and correspondingly, predicting the implantation category corresponding to the placenta image data based on the target features specifically includes:

The target feature is input into a trained classifier, and the implantation category corresponding to the placenta image data is predicted by the classifier.

Specifically, the classifier is trained, as shown in Figure 4, the training process of the classifier specifically includes:

For each classification training sample in the preset classification training set, extract the deep semantic feature corresponding to the classification training sample based on the semantic feature extraction module, and extract the radiomics features of the classification training sample;

Screening the deep semantic features and the radiomics features to obtain target features;

Input the second preset network model based on the target characteristics, determine the predicted implant category through the second preset network model, and classify the second preset network model based on the predicted implant category and the labeled implant category corresponding to the classified training samples. Two preset network models are trained to obtain the classifier.

Specifically, the model structure of the second preset network model is the same as that of the classifier, and the difference between the two lies in the model parameters. The second preset network model adopts the initial network model, and the classifier adopts the network model for training. In this embodiment, the classifier may use a support vector classifier.

The preset classification training set includes a number of classification training sample images, each classification training sample image in the number of classification training sample images is placenta image data, and part of the training sample images in the number of training sample images are placenta accreta. Placenta image data, some training sample images are placenta image data with placenta accreta. In this embodiment, the preset segmentation training set and the preset classification training set can be collected synchronously, wherein, the acquisition process of the preset segmentation training set and the preset classification training set can be: first collect the placental MRI image Then divide the collected placental MRI image data into segmentation data set and classification data set, finally divide the segmentation data set into segmentation training set and segmentation test set, and divide the classification data set into classification training set and classification test set.

In an implementation manner, the determination process of dividing the data set and classifying the data set may be as follows:

The first step, data collection

Collect a first preset number (for example, 241 cases) of placental MRI image data after intraoperative pathological examination as an internal modeling data set, wherein each placental MRI image data in the internal modeling data set carries an implantation category , and the ratio of positive and negative samples in the internal modeling data set is 169:72, the positive samples are placenta MRI image data with placenta accreta, and the negative samples are placenta MRI image data without placenta accreta. Simultaneously, collect the second preset quantity (for example, 122) as the external independent test set, wherein, each placenta MRI image data in the external independent test set all carries implantation category, and the positive and negative sample ratio in the external independent test set is 106:16.

The patients corresponding to the placental MRI image data of the internal modeling data set and the external independent test set are all patients who have undergone clinical obstetric examination, prenatal MRI examination, delivery, and have a history of cesarean section or uterine cavity operation, and the corrected gestational age is greater than 18 weeks , single pregnancy and no gestational diabetes, no gestational hypertension, no blood system abnormalities and other imaging placental diseases, the patient's clinical and imaging data are complete. At the same time, for pregnant women with underlying diseases (such as diabetes, thalassemia, etc.); cases of fetal growth and development restriction or fetal diseases and placenta related images; pregnant women with contraindications for MRI examination; blurred imaging images with motion artifacts Or pregnant women who were unable to cooperate with relevant examinations were not included in the internal modeling data set and the external independent test set.

The second step is to outline the interest

After obtaining the training sample set, the region of interest (ROI) of the placental image data will be delineated and divided. Among them, the region of interest (Region of Interest, ROI) of the placental MRI data is used by radiologists with experience in MRI imaging diagnostics using open source software ITK-SNAP (version 3.4.0; https://www.itksnap.org) localizes and delineates the fine outline of the placental region on the patient's MRI images. However, randomly select a specified number (for example, 40 cases, etc.) in the internal modeling data set as the preset segmentation data set, and use the remaining placental MRI image data in the internal modeling data set and the placental MRI image data in the external independent test set as Preset classification datasets.

The third step, data preprocessing

1) Uniform image size

Since the size of each patient's MRI data on the z-axis is not uniform, the segmentation model training effect is likely to be unsatisfactory. All placenta image data were interpolated and resampled and the image size was unified to 128×128×96.

2), resampling

In order to extract radiomics features in the ROI region of the classification data set, the MRI data was resampled, wherein the training samples of the classification data set contained in the internal modeling data set in the classification training set were statistically known , the average resolution of the training samples of the classification dataset included in the internal modeling dataset in the classification training set is 0.87×0.87×5.8 mm ³ . In order to avoid resampling operations on too much data, the data in the classification training set were uniformly resampled to 0.87×0.87×5.8mm ³ . Of course, the same resampling operation can also be applied to the ROI data drawn by the doctor.

In summary, this embodiment provides a radiomics placenta accreta prediction method based on deep semantic features, the method includes acquiring placental image data; extracting deep semantic features and radiomics features of the placental image data ; Screening the deep semantic features and the radiomics features to obtain target features; predicting the implantation category corresponding to the placenta image data based on the target features. In this application, by combining radiomics features and deep semantic features, feature information with rich dimensions and levels can be obtained, and then placenta accreta prediction can be performed based on feature information, which can improve the accuracy of placenta accreta prediction.

Based on the radiomics placenta accreta prediction method based on deep semantic features, this embodiment provides a radiomics placenta accreta prediction device based on deep semantic features, as shown in FIG. 6 , the device includes:

A feature extraction module 100, configured to extract deep semantic features and radiomics features of the placenta image data;

A screening module 200, configured to screen the deep semantic features and the radiomics features to obtain target features;

The classification module 300 is configured to predict the implantation category corresponding to the placenta image data based on the target features.

Based on the radiomics placenta accreta prediction method based on deep semantic features, this embodiment provides a computer-readable storage medium, the computer-readable storage medium stores one or more programs, and the one or more The program can be executed by one or more processors, so as to realize the steps in the radiomics placenta accreta prediction method based on deep semantic features as described in the above-mentioned embodiments.

Based on the above radiomics placenta accreta prediction method based on deep semantic features, the present application also provides a terminal device, as shown in FIG. 7 , which includes at least one processor (processor) 20; display screen 21; and memory ( memory) 22, may also include a communication interface (Communications Interface) 23 and a bus 24. Wherein, the processor 20 , the display screen 21 , the memory 22 and the communication interface 23 can communicate with each other through the bus 24 . The display screen 21 is configured to display the preset user guidance interface in the initial setting mode. The communication interface 23 can transmit information. The processor 20 can invoke logic instructions in the memory 22 to execute the methods in the above-mentioned embodiments.

In addition, the above-mentioned logic instructions in the memory 22 may be implemented in the form of software functional units and when sold or used as an independent product, may be stored in a computer-readable storage medium.

As a computer-readable storage medium, the memory 22 can be configured to store software programs and computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of the present disclosure. The processor 20 runs software programs, instructions or modules stored in the memory 22 to execute functional applications and data processing, ie to implement the methods in the above-mentioned embodiments.

The memory 22 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required by a function; the data storage area may store data created according to the use of the terminal device, and the like. In addition, the memory 22 may include a high-speed random access memory, and may also include a non-volatile memory. For example, U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes can also be temporary state storage medium.

In addition, the specific process of loading and executing multiple instruction processors in the storage medium and the terminal device has been described in detail in the above method, and will not be described here one by one.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limiting them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims (10)

1. An image histology placenta implantation prediction method based on depth semantic features, which is characterized by comprising the following steps:

extracting depth semantic features and image histology features of the placenta image data;

screening the depth semantic features and the image histology features to obtain target features;

and predicting an implantation category corresponding to the placenta image data based on the target feature.

2. The depth semantic feature based imaging histology placenta implantation prediction method of claim 1, wherein the placenta image data is placenta MRI data.

3. The method of claim 1, wherein the filtering the deep semantic features and the image histology features to obtain target features specifically comprises:

performing variance alignment verification on the depth semantic features and the image group chemical features respectively to obtain verification values of all the depth semantic features in the depth semantic features and verification values of all the image group chemical features in the image group chemical features;

screening the depth semantic features based on the check values of the depth semantic features to obtain target depth semantic features, and screening the image histology features based on the check values of the image histology features to obtain target image histology features;

and splicing the target depth semantic features and the target image histology features to obtain target features.

4. The method for predicting placenta implantation in image group based on deep semantic features of claim 1, wherein the extracting the deep semantic features of the placenta image data specifically comprises:

inputting the placenta image data into a preset semantic feature extraction module, and determining the deep semantic features of the placenta image data through the feature extraction module;

the semantic feature extraction module comprises an encoder and an adaptive average pooling layer, wherein the encoder comprises a first feature extraction unit and a plurality of cascaded second feature extraction units, the first feature extraction unit is connected with the second feature extraction unit positioned at the forefront, and the second feature extraction unit positioned at the last is connected with the adaptive average pooling layer; the second feature extraction unit comprises a maximum pooling layer and a first feature extraction unit; the first feature extraction unit comprises two cascaded convolution blocks, wherein each convolution block comprises a convolution layer, a batch normalization layer and an activation function layer which are cascaded in sequence.

5. The method for predicting placenta implantation in image group based on deep semantic features of claim 4, wherein the determining process of the semantic feature extraction module specifically comprises:

training a first preset network model based on a preset segmentation training set to obtain a segmentation network model, wherein the segmentation network model comprises the encoder;

and extracting an encoder of the segmentation network model, and connecting the encoder with an adaptive average pooling layer to form the semantic feature extraction module.

6. The depth semantic feature based imaging placenta implantation prediction method of claim 5, wherein the segmentation network model further comprises a decoder; the decoder comprises a first up-sampling unit, a plurality of cascaded second up-sampling units and a convolution unit; the first up-sampling unit is connected with the second feature extraction unit positioned at the last; the first up-sampling unit is connected with a second up-sampling unit positioned at the forefront, and a second up-sampling unit positioned at the last is connected with the convolution unit; each second feature extraction unit except the last second feature extraction unit in the plurality of second feature extraction units corresponds to each second up-sampling unit one by one and is connected in a jumping manner; the first feature extraction unit is connected with the convolution unit in a jumping manner; the second upsampling unit comprises a first feature extraction unit and an upsampling layer; the convolution unit includes a convolution layer and a first feature extraction unit.

7. The depth semantic feature-based image histology placenta implantation prediction method according to any one of claims 4-6, wherein predicting the implantation category corresponding to the placenta image data based on the target feature specifically comprises:

inputting the target characteristics into a trained classifier, and predicting implantation categories corresponding to the placenta image data through the classifier;

the training process of the classifier specifically comprises the following steps:

for each classification training sample in a preset classification training set, extracting depth semantic features corresponding to the classification training samples based on the semantic feature extraction module, and extracting image histology features of the classification training samples;

screening the depth semantic features and the image histology features to obtain target features;

and inputting a second preset network model based on the target characteristics, determining a predicted implantation category through the second preset network model, and training the second preset network model based on the predicted implantation category and a labeling implantation category corresponding to the classifying training sample to obtain the classifier.

8. An image histology placenta implantation prediction device based on depth semantic features, the device comprising:

the feature extraction module is used for extracting depth semantic features and image histology features of the placenta image data;

the screening module is used for screening the depth semantic features and the image histology features to obtain target features;

and the classification module is used for predicting the implantation category corresponding to the placenta image data based on the target characteristics.

9. A computer readable storage medium storing one or more programs executable by one or more processors to perform the steps in the depth semantic feature based method of image histology placenta implantation prediction of any one of claims 1-7.

10. A terminal device, comprising: a processor, a memory, and a communication bus; the memory has stored thereon a computer readable program executable by the processor;

the communication bus realizes connection communication between the processor and the memory;

the processor, when executing the computer readable program, implements the steps in the depth semantic feature based method for predicting placenta implantation in image groups according to any one of claims 1-7.