CT image data processing method and device, storage medium and electronic equipment

Abstract

The invention relates to a CT image data processing method and device, a storage medium and electronic equipment, and belongs to the technical field of data processing, and the method comprises the steps: generating a plurality of training sets according to the combination of the types of bronchiectasis lesions, different training sets comprising CT image data corresponding to different types of combinations of bronchiectasis lesions, the first annotation information comprises a bronchiectasis lesion included in the CT image and the type of the bronchiectasis lesion; according to the plurality of training sets in sequence, training to obtain a first processing model; acquiring CT image data of the lung; inputting the CT image data into a first processing model to obtain a bronchiectasia focus output by the first processing model and determined according to the CT image data, and the type of the bronchiectasia focus; and outputting a processing result of the CT image data according to the determined bronchiectasis focus and the type of the bronchiectasis focus.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data processing, and in particular, to a method, device, storage medium, and electronic device for processing CT image data. Background Art

[0002] Bronchiectasis (referred to as "bronchiectasis" for short) refers to chronic inflammation of the bronchus and its surrounding lung tissue that damages the bronchial wall, resulting in bronchial dilation and deformation. The assessment of bronchiectasis is mainly based on high-resolution CT (Computed Tomography) images of the lungs. In the process of processing these CT images, problems such as long processing time and poor consistency of processing results are also faced. Summary of the Invention

[0003] The purpose of the present disclosure is to provide a method, device, storage medium, and electronic device for processing CT image data to solve the above-mentioned related technical problems.

[0004] To achieve the above purpose, in a first aspect, the present disclosure provides a method for processing CT image data, including: Generating a plurality of training sets according to combinations of types of bronchiectasis lesions, where different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data, the first annotation information including the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions; Successively training a model to be trained according to the plurality of training sets to obtain a first processing model; Obtaining CT image data of the lungs; Inputting the CT image data into the first processing model to obtain bronchiectasis lesions determined according to the CT image data and the types of the bronchiectasis lesions output by the first processing model; Outputting a processing result of the CT image data according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions.

[0005] Optionally, the generating a plurality of training sets according to combinations of types of bronchiectasis lesions includes: Generating a first training set, where the first training set includes first CT image data and second CT image data, the first CT image data having at least columnar bronchiectasis lesions, and the second CT image data having at least varicose bronchiectasis lesions; Generating a second training set, where the second training set includes the second CT image data and third CT image data, the third CT image data having at least cystic bronchiectasis lesions; Generate a third training set, where the third training set includes the first CT image data, the second CT image data, and the third CT image data.

[0006] Optionally, the step of sequentially training a model to be trained according to the multiple training sets to obtain a first processing model includes: Train the model to be trained with the first training set to obtain a first model; Train the first model with the second training set to obtain a second model; Train the second model with the third training set to obtain the first processing model.

[0007] Optionally, the step of generating multiple training sets according to combinations of types of bronchiectasis lesions includes: Obtain a CT image data set, where the CT image data set includes fourth CT image data and second annotation information of the fourth CT image data, and the second annotation information includes bronchiectasis lesions included in the fourth CT image and types of the bronchiectasis lesions; Verify the fourth CT image data to obtain verified fourth CT image data; Generate multiple training sets according to the verified fourth CT image data, the second annotation information of the verified fourth CT image data, and combinations of types of the bronchiectasis lesions.

[0008] Optionally, the method includes: Input the CT image data into a second processing model to obtain a segmentation result of the CT image data output by the second processing model, where the segmentation result includes lobe information in the CT image; the second processing model is trained with sample CT image data as input and a sample segmentation result of a lobe determined according to the sample CT image data as output; The step of outputting a processing result of the CT image data according to the determined bronchiectasis lesions and types of the bronchiectasis lesions includes: Determine bronchiectasis lesions included in each lobe and types of the bronchiectasis lesions in the lobe according to the determined bronchiectasis lesions, types of the bronchiectasis lesions, and the lobe information; Output a processing result of the CT image data according to the bronchiectasis lesions included in each lobe and types of the bronchiectasis lesions in the lobe.

[0009] Optionally, outputting a processing result of the CT image data according to the bronchodilation lesions included in each lung lobe and the types of the bronchodilation lesions in the lung lobe includes: Calculating a reference value according to the bronchodilation lesions included in each lung lobe and the types of the bronchodilation lesions; Obtaining symptom description information; Semantically expanding the symptom description information to obtain target description information; Querying a knowledge base of pathogenic factors according to the target description information and the reference value to obtain prompt information of pathogenic factors.

[0010] Optionally, the first processing model includes a downsampling module, an upsampling module, and a convolution module connected in sequence, The downsampling module includes a first convolutional layer, a second convolutional layer, a feature verification layer, and a downsampling layer connected in sequence, and the feature verification layer is configured to verify the features output by the second convolutional layer; The upsampling module includes a third convolutional layer, a fourth convolutional layer, a feature verification layer, and an upsampling layer connected in sequence, and the feature verification layer in the upsampling module is configured to verify the features output by the fourth convolutional layer; The convolution module includes a fifth convolutional layer, a sixth convolutional layer, a feature verification layer, and a seventh convolutional layer connected in sequence, and the feature verification layer in the convolution module is configured to verify the features output by the sixth convolutional layer.

[0011] Optionally, the feature verification layer is configured to: Obtain activation features input to the feature verification layer; Project the activation features in the depth direction, height direction, and width direction respectively to obtain corresponding first features, second features, and third features; Perform broadcast processing on the first feature, second feature, and third feature, and sum the broadcast-processed first feature, second feature, and third feature to obtain a fourth feature; Perform channel compression processing on the fourth feature through a first convolutional kernel, and perform channel restoration on the fourth feature after channel compression processing through a second convolutional kernel to obtain a calibration weight; Perform element-wise multiplication on the calibration weight and the activation features to obtain the verified features output by the feature verification layer.

[0012] Optionally, the projecting the activation features in the depth direction, height direction, and width direction respectively includes: In the depth direction, based on the height weight and width weight in the current feature verification layer, the features of the activation feature in height and width are weighted and summed to obtain the first feature; In the height direction, based on the depth weight and width weight in the current feature verification layer, the features of the activation feature in depth and width are weighted and summed to obtain the second feature; In the width direction, based on the height weight and depth weight in the current feature verification layer, the features of the activation feature in height and depth are weighted and summed to obtain the third feature.

[0013] Optionally, the downsampling module includes a first downsampling module, a second downsampling module, a third downsampling module, and a fourth downsampling module, and the upsampling module includes a first upsampling module, a second upsampling module, a third upsampling module, and a fourth upsampling module, where The outputs of the first downsampling module and the fourth upsampling module are used as the inputs of the convolutional module; The outputs of the second downsampling module and the third upsampling module are used as the inputs of the fourth upsampling module; The outputs of the third downsampling module and the second upsampling module are used as the inputs of the third upsampling module; The outputs of the fourth downsampling module and the first upsampling module are used as the inputs of the second upsampling module; The output of the fourth downsampling module is used as the input of the first upsampling module.

[0014] In a second aspect, a processing device for CT image data is provided, including: A generation module, configured to generate a plurality of training sets according to combinations of types of bronchiectasis lesions. Different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data, where the first annotation information includes the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions; A training module, configured to sequentially train a model to be trained according to the plurality of training sets to obtain a first processing model; A first module, configured to obtain CT image data of the lungs; A second module, configured to input the CT image data into the first processing model to obtain the bronchiectasis lesions determined according to the CT image data output by the first processing model, and the types of the bronchiectasis lesions; A third module, configured to output a processing result of the CT image data according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions.

[0015] Optionally, the generating module includes: A first sub-module, configured to generate a first training set, where the first training set includes first CT image data and second CT image data, the first CT image data has at least columnar bronchiectasis lesions, and the second CT image data has at least varicose bronchiectasis lesions; A second sub-module, configured to generate a second training set, where the second training set includes the second CT image data and third CT image data, and the third CT image data has at least cystic bronchiectasis lesions; A third sub-module, configured to generate a third training set, where the third training set includes the first CT image data, the second CT image data, and the third CT image data.

[0016] Optionally, the training module includes: A fourth sub-module, configured to train the model to be trained through the first training set to obtain a first model; A fifth sub-module, configured to train the first model through the second training set to obtain a second model; A sixth sub-module, configured to train the second model through the third training set to obtain the first processing model.

[0017] Optionally, the generating module includes: A seventh sub-module, configured to obtain a CT image data set, where the CT image data set includes fourth CT image data and second annotation information of the fourth CT image data, and the second annotation information includes the bronchiectasis lesions included in the fourth CT image and the types of the bronchiectasis lesions; An eighth sub-module, configured to verify the fourth CT image data to obtain verified fourth CT image data; A ninth sub-module, configured to generate a plurality of training sets according to a combination of the verified fourth CT image data, the second annotation information of the verified fourth CT image data, and the types of the bronchiectasis lesions.

[0018] Optionally, the apparatus includes: The fourth module is configured to input the CT image data into a second processing model to obtain a segmentation result of the CT image data output by the second processing model, where the segmentation result includes lobe information in the CT image; the second processing model is trained by using sample CT image data as input and a sample segmentation result of a lobe determined according to the sample CT image data as output; The third module includes: The tenth sub-module is configured to determine bronchiectatic lesions included in each lobe and types of the bronchiectatic lesions in the lobe according to the determined bronchiectatic lesions, types of the bronchiectatic lesions, and the lobe information; The eleventh sub-module is configured to output a processing result of the CT image data according to the bronchiectatic lesions included in each lobe and types of the bronchiectatic lesions in the lobe.

[0019] Optionally, the eleventh sub-module is configured to: Calculate a reference value according to the bronchiectatic lesions included in each lobe and types of the bronchiectatic lesions; Obtain symptom description information; Perform semantic expansion on the symptom description information to obtain target description information; Query a knowledge base of pathogenic factors according to the target description information and the reference value to obtain prompt information of pathogenic factors.

[0020] Optionally, the first processing model includes a downsampling module, an upsampling module, and a convolution module that are connected in sequence, The downsampling module includes a first convolutional layer, a second convolutional layer, a feature verification layer, and a downsampling layer that are connected in sequence, and the feature verification layer is configured to verify features output by the second convolutional layer; The upsampling module includes a third convolutional layer, a fourth convolutional layer, a feature verification layer, and an upsampling layer that are connected in sequence, and the feature verification layer in the upsampling module is configured to verify features output by the fourth convolutional layer; The convolution module includes a fifth convolutional layer, a sixth convolutional layer, a feature verification layer, and a seventh convolutional layer that are connected in sequence, and the feature verification layer in the convolution module is configured to verify features output by the sixth convolutional layer.

[0021] Optionally, the feature verification layer is configured to: Obtain activation features input to the feature verification layer; Project the activation features in the depth direction, height direction, and width direction respectively to obtain corresponding first features, second features, and third features; Broadcast process the first feature, the second feature, and the third feature, and sum the broadcast-processed first feature, second feature, and third feature to obtain a fourth feature; Perform channel compression processing on the fourth feature through a first convolutional kernel, and perform channel restoration on the fourth feature after channel compression processing through a second convolutional kernel to obtain a calibration weight; Perform element-wise multiplication on the calibration weight and the activation feature to obtain the verified feature output by the feature verification layer.

[0022] Optionally, the feature verification layer projects the activation feature in the depth direction, height direction, and width direction respectively, including: In the depth direction, based on the height weight and width weight in the current feature verification layer, perform weighted summation on the features in the height and width of the activation feature to obtain the first feature; In the height direction, based on the depth weight and width weight in the current feature verification layer, perform weighted summation on the features in the depth and width of the activation feature to obtain the second feature; In the width direction, based on the height weight and depth weight in the current feature verification layer, perform weighted summation on the features in the height and depth of the activation feature to obtain the third feature.

[0023] Optionally, the downsampling module includes a first downsampling module, a second downsampling module, a third downsampling module, and a fourth downsampling module, and the upsampling module includes a first upsampling module, a second upsampling module, a third upsampling module, and a fourth upsampling module, where The outputs of the first downsampling module and the fourth upsampling module are used as the inputs of the convolutional module; The outputs of the second downsampling module and the third upsampling module are used as the inputs of the fourth upsampling module; The outputs of the third downsampling module and the second upsampling module are used as the inputs of the third upsampling module; The outputs of the fourth downsampling module and the first upsampling module are used as the inputs of the second upsampling module; The output of the fourth downsampling module is used as the input of the first upsampling module.

[0024] In a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in any one of the first aspects are implemented.

[0025] In a fourth aspect, an electronic device is provided, including: A memory storing a computer program thereon; A processor configured to execute the computer program in the memory to implement the steps of the method according to any one of the first aspect.

[0026] In the above technical solution, multiple training sets can be generated according to the combination of the types of bronchiectasis lesions. Different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data, where the first annotation information includes the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions. In this way, the model to be trained can be trained successively according to the multiple training sets to obtain a first processing model. In this way, the CT image data of the lungs can be input into the first processing model to obtain the bronchiectasis lesions determined according to the CT image data output by the first processing model, and the types of the bronchiectasis lesions. In this way, according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions, the processing result of the CT image data can be output.

[0027] By adopting the above solution, multiple training sets can be generated based on the combination of the types of bronchiectasis lesions, and the first processing model can be obtained by training successively through the multiple training sets. In this way, through multi-level training, it helps to improve the model training effect. In addition, the CT image data can be processed by the first processing model obtained through training, so as to determine the bronchiectasis lesions and the types of the bronchiectasis lesions. In this way, it helps to improve the processing efficiency of the CT image data and ensure the consistency of the processing results.

[0028] Other features and advantages of the present disclosure will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The drawings are used to provide a further understanding of the present disclosure, and constitute a part of the specification, and are used to explain the present disclosure together with the following specific implementation, but do not constitute a limitation to the present disclosure. In the drawings: Figure 1 is a flowchart of a method for processing CT image data shown in an exemplary embodiment of the present disclosure.

[0030] Figure 2 is a flowchart of a data processing shown in an exemplary embodiment of the present disclosure.

[0031] Figure 3 is a flowchart of a data processing of a feature verification layer shown in an exemplary embodiment of the present disclosure.

[0032] Figure 4It is a flowchart of data processing for a feature verification layer shown in an exemplary embodiment of the present disclosure.

[0033] Figure 5 It is a schematic structural diagram of a model shown in an exemplary embodiment of the present disclosure.

[0034] Figure 6 It is a processing flowchart of CT image data shown in an exemplary embodiment of the present disclosure.

[0035] Figure 7 It is a processing flowchart of CT image data shown in an exemplary embodiment of the present disclosure.

[0036] Figure 8 It is a block diagram of a processing device for CT image data shown in an exemplary embodiment of the present disclosure.

[0037] Figure 9 It is a block diagram of an electronic device shown in an exemplary embodiment of the present disclosure. Detailed Embodiments

[0038] The following will describe the detailed embodiments of the present disclosure with reference to the accompanying drawings. It should be understood that the detailed embodiments described herein are only for explaining and illustrating the present disclosure, and are not used to limit the present disclosure.

[0039] Before introducing the CT image data processing method, device, storage medium and electronic device of the present disclosure, the relevant scenarios of the embodiments of the present disclosure will be introduced first.

[0040] Bronchiectasis refers to the chronic inflammation of the bronchus and its surrounding lung tissue that damages the bronchial wall, resulting in bronchiectasis and deformation. Currently, the clinical evaluation of bronchiectasis can be based on high-resolution CT of the lungs, and it can be determined by the visual evaluation of radiologists whether bronchiectasis, partially affected lung lobes or segments are visible. In some cases, the severity of bronchiectasis can also be evaluated. For example, in the scenario of the modified Reiff imaging score, different scores can be assigned to different types of bronchiectasis lesions according to the severity. Among them, the columnar type (also known as the "double-track sign") is 1 point, the varicose type (also known as the "cystic columnar") is 2 points, and the cystic type is 3 points. In addition, the bronchiectasis lesions and their types appearing in 6 lung lobes (the lingular segment of the left lower lobe is regarded as a separate lung lobe, which can be called the "lingular lobe") can be observed, and the score of the most severe lesion type that appears is used as the score of this lung lobe, with a maximum of 3 points (presence of cystic bronchiectasis lesions) and a minimum of 0 points (no bronchiectasis lesions). In this way, the total score of 6 lung lobes can be counted, with a maximum of 18 points and a minimum of 0 points. The higher the total score, the more severe the bronchiectasis.

[0041] However, the above method relies on manual processing of CT images. Since manual reading of images takes a long time, perhaps 10 minutes to complete the processing, there is a problem of low processing efficiency. In addition, different processors may have deviations in the judgment of CT images, so affected by subjectivity, the consistency of the processing results is low.

[0042] Moreover, the modified Reiff score emphasizes the score of the most severe lesion in the lung lobe while ignoring the number of lesions in the lung lobe. For example: the highest score of the bronchiectasis lesion in a certain lung lobe is 2 points, and the number of lesions in this lung lobe is 5, while the highest score of the bronchiectasis lesion in another lung lobe is 1 point, and the number of lesions in this lung lobe is 60. From the above example, it can be seen that evaluating the severity of lung lobe involvement due to bronchiectasis solely based on the score of bronchiectasis lesions is not comprehensive and accurate enough.

[0043] In addition, bronchiectasis can be diffusely distributed in both lungs or can be a localized lesion, and its occurrence site is related to the etiology. Therefore, the treatment strategies for patients with bronchiectasis of different etiologies may be completely different. However, the CT image reports of relevant scenarios only report whether the CT manifestations of the patient suggest bronchiectasis, the partially (but not comprehensively) involved lung lobes or segments, lacking tips on the pathogenic factors.

[0044] For this reason, the present disclosure provides a method for processing CT image data. Figure 1 is a flowchart of a method for processing CT image data shown in an exemplary embodiment of the present disclosure. Referring to Figure 1 the method includes: In step S11, multiple training sets are generated according to the combination of the types of bronchiectasis lesions.

[0045] In step S12, the model to be trained is trained sequentially according to multiple training sets to obtain a first processing model.

[0046] In step S13, CT image data of the lungs is obtained.

[0047] In step S14, the CT image data is input into the first processing model to obtain the bronchiectasis lesions determined according to the CT image data and the types of the bronchiectasis lesions output by the first processing model.

[0048] In step S15, according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions, the processing result of the CT image data is output.

[0049] The following is an exemplary description of the implementation manners of the above steps.

[0050] In step S11, multiple training sets are generated according to the combination of the types of bronchiectasis lesions.

[0051] Among them, different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data, where the first annotation information includes the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions.

[0052] Bronchiectasis generally occurs in bronchi of grades 3 to 6. According to the morphology of the dilated bronchi, bronchiectasis can be classified into columnar bronchiectasis, cystic bronchiectasis, and varicose bronchiectasis. The lesions of bronchiectasis are divided into grade 1 (columnar bronchiectasis), grade 2 (varicose bronchiectasis, also known as cystic-columnar bronchiectasis), and grade 3 (cystic bronchiectasis) according to the severity. Grade 1 is the mildest, grade 3 is the most severe, and grade 2 is between the two.

[0053] When generating the first annotation information, classification annotation can be performed according to the lesion grading, and different categories represent lesions of different grades. For example: the first type of annotated lesions is grade 1 bronchiectasis, the second type is grade 2 bronchiectasis, and the third type is grade 3 bronchiectasis. The occurrence site of the bronchiectasis lesion is annotated, and the annotation mode is semantic segmentation annotation, that is, the lesion boundary can be accurate to pixel points. The CT image data in the training set can be greater than or equal to 300 pieces.

[0054] In a possible implementation manner, generating a plurality of training sets according to the type combination of bronchiectasis lesions includes: Generating a first training set, where the first training set includes first CT image data and second CT image data, the first CT image data has at least columnar bronchiectasis lesions, and the second CT image data has at least varicose bronchiectasis lesions. Correspondingly, the annotation information of the first CT image data is at least annotated with columnar bronchiectasis lesions, and the second CT image data is at least annotated with varicose bronchiectasis lesions.

[0055] In addition, a second training set can be generated, where the second training set includes the second CT image data and third CT image data, and the third CT image data has at least cystic bronchiectasis lesions. Correspondingly, the third CT image data is at least annotated with cystic bronchiectasis lesions.

[0056] In addition, a third training set is generated, where the third training set includes the first CT image data, the second CT image data, and the third CT image data.

[0057] It should be noted that, in some embodiments, the first CT image data, the second CT image data, and the third CT image data can be generated by hiding or modifying the annotation data. For example, for the CT image data and the corresponding annotation information that simultaneously have columnar bronchiectasis lesions, varicose bronchiectasis lesions, and cystic bronchiectasis lesions, the annotation category values of the annotation pixel points can be found and modified through logical operations and multiplication operations, so as to hide the bronchiectasis annotation of a certain category, or restore the bronchiectasis annotations of all categories.

[0058] Exemplarily, the first CT image data can be obtained by hiding one or more of the varicose bronchiectasis lesions and the cystic bronchiectasis lesions. The second CT image data can be obtained by hiding one or more of the columnar bronchiectasis lesions and the cystic bronchiectasis lesions. The third CT image data can be obtained by hiding one or more of the columnar bronchiectasis lesions and the varicose bronchiectasis lesions.

[0059] In this way, it helps to increase the number of training samples and also helps to train the model in stages.

[0060] In addition, when generating the training set, the data can also be verified and preprocessed. For example, in one embodiment, generating multiple training sets according to the combination of the types of bronchiectasis lesions includes: Obtaining a CT image data set, where the CT image data set includes fourth CT image data and second annotation information of the fourth CT image data, and the second annotation information includes the bronchiectasis lesions included in the fourth CT image and the types of the bronchiectasis lesions; Verifying the fourth CT image data to obtain the fourth CT image data that passes the verification; Generating multiple training sets according to the fourth CT image data that passes the verification, the second annotation information of the fourth CT image data that passes the verification, and the combination of the types of the bronchiectasis lesions.

[0061] Figure 2 is a flowchart of a data processing shown in an exemplary embodiment of the present disclosure. Referring to Figure 2 , the fourth CT image data can be verified and processed.

[0062] Here, the fourth CT image data can be subjected to availability verification, integrity verification, repeatability verification, consistency verification, etc. The implementation manners of the availability verification, integrity verification, and repeatability verification can refer to the subsequent description of step S13.

[0063] In the consistency check, it is possible to check whether the annotation file in the second annotation information and the fourth CT image data are consistent in terms of spatial size, and whether the difference in voxel spacing is less than a set threshold. When the annotation file in the second annotation information and the fourth CT image data are consistent in spatial size and the difference in voxel spacing is less than the set threshold, it is determined that the consistency check passes; otherwise, it is determined that the consistency check fails.

[0064] In addition, data preprocessing can be performed on the fourth CT image data that has passed the check.

[0065] For example, in one implementation, a data dictionary can be established. Exemplarily, a mapping relationship between the fourth CT image data and the corresponding second annotation information can be established and encapsulated as a list of dictionaries in the format of a Python language dictionary.

[0066] In one implementation, data preprocessing includes data loading. Exemplarily, the fourth CT image data and its corresponding second annotation information (i.e., the annotation file) can be loaded, and a channel dimension can be added in front of its spatial dimension, so that the data structure changes from (depth, height, width) to (number of channels, depth, height, width).

[0067] In one implementation, data preprocessing includes processing of the annotation file (annotation information). Exemplarily, the bronchiectasis annotation of a certain category can be hidden, or the bronchiectasis annotations of all categories can be restored. For example, the annotation category value of the annotation pixel points can be found and modified through logical operations and multiplication operations.

[0068] In one implementation, data preprocessing includes data normalization. Exemplarily, the image can be standardized based on the CT value range (such as -1350 to 150 for the lung window), and the CT values (i.e., HU values) from -1350 to 150 are standardized to [0, 1]. Among them, those lower than -1350 are standardized to 0, and those higher than 150 are standardized to 1. The normalization method is as follows: 。

[0069] where the denominator is the largest HU value (150) minus the smallest HU value (-1350), and the numerator is the HU value of the pixel point minus . In this way, the standardized value of the pixel point can be calculated.

[0070] In one embodiment, data preprocessing includes foreground cropping. The foreground can refer to the manually annotated pulmonary vessel mask, and in this case, the part outside the pulmonary vessels is the background. In this way, the effective part of the image can be cropped based on the annotation to improve the efficiency and effectiveness of training and evaluation.

[0071] In one embodiment, data preprocessing includes resampling. Exemplarily, the CT image data and the annotation can be resampled to unify the anatomical coordinates and voxel spacing, so that the model can learn consistent target features of interest.

[0072] In one embodiment, data preprocessing includes random cropping. Exemplarily, one or more input blocks (crops) of a fixed size can be randomly cropped from corresponding positions in the image and the annotation. In this way, on the one hand, batch training of the input samples can be performed. On the other hand, through this random cropping, the learning target can be located at different positions in space, thereby weakening the sensitivity of the convolutional neural network to the target position and improving the generalization ability of the trained model. In one embodiment, the spatial positions of the random cropping are different in each iteration cycle during the model training process.

[0073] In one embodiment, data preprocessing includes data augmentation. Exemplarily, random affine transformations (such as rotation, scaling, etc.) can be performed on the cropped input blocks. In this way, it helps to improve the generalization ability of the trained model.

[0074] It should be noted that for the models in the embodiments of the present disclosure (such as the first processing model and the subsequent second processing model), the CT image data input in the model inference stage (application stage) can also perform the above data verification and data preprocessing operations. Since there is no annotation file for the CT image data input in the inference stage, there can be no annotation file when establishing the data dictionary for data preprocessing. Furthermore, during data preprocessing, operations such as annotation file processing, foreground cropping, random cropping, and data augmentation are not required.

[0075] Referring to Figure 1 , in step S12, the model to be trained is trained according to multiple training sets in sequence to obtain the first processing model.

[0076] Through data verification and data preprocessing, multiple training sets can be obtained. In this way, the model to be trained can be trained through the training sets. For example, in one embodiment, the step of training the model to be trained according to the multiple training sets in sequence to obtain the first processing model includes: Training the model to be trained through the first training set to obtain the first model; Training the first model through the second training set to obtain the second model; The second model is trained with the third training set to obtain the first processing model.

[0077] That is to say, the model to be trained can be trained based on the first training set, so as to train the model to be trained for three-classification (background, columnar bronchiectasis, and varicose bronchiectasis) segmentation model training of columnar bronchiectasis and varicose bronchiectasis, and obtain the first model. Based on the first model, the columnar bronchiectasis can be hidden, and the first model can be trained for three-classification (background, cystic bronchiectasis, and varicose bronchiectasis) of cystic bronchiectasis and varicose bronchiectasis to obtain the second model. Finally, the second model can be trained with the third training set to perform four-classification (background, columnar bronchiectasis, varicose bronchiectasis, and cystic bronchiectasis) on columnar bronchiectasis, varicose bronchiectasis, and cystic bronchiectasis, so as to obtain the first processing model. In some embodiments, considering that the first model and the second model have three-class output channels, when training the second model, the last convolutional layer (such as a 1×1×1 convolutional layer) of the model can also be replaced, and the output channels can be changed to four-class for training.

[0078] In the above solution, multiple training sets can be generated based on the combination of the types of bronchiectasis lesions, and the first processing model can be obtained by training with the multiple training sets in sequence. In this way, through multi-level training, it helps to improve the model training effect.

[0079] The structure and training method of the first processing model are exemplarily described below.

[0080] In a possible implementation manner, the first processing model includes a downsampling module, an upsampling module, and a convolutional module connected in sequence.

[0081] The downsampling module includes a first convolutional layer, a second convolutional layer, a feature verification layer, and a downsampling layer connected in sequence. The feature verification layer is configured to verify the features output by the second convolutional layer.

[0082] Among them, each convolutional layer in the first convolutional layer and the second convolutional layer can be connected with a BN layer (Batch Normalization) and an activation layer ReLU. The BN layer can accelerate the training of the model and improve its stability. ReLU helps the model learn complex patterns and has the advantages of simple calculation and alleviating gradient disappearance. The implementation manner of the feature verification layer will be described in subsequent embodiments. In addition, the downsampling layer can be, for example, a pooling layer, such as a max pooling layer.

[0083] In this way, every time a downsampling module passes, the size of the feature map becomes smaller and the number of output channels increases.

[0084] The upsampling module of the first processing model includes a third convolutional layer, a fourth convolutional layer, a feature verification layer, and an upsampling layer connected in sequence. The feature verification layer in the upsampling module is configured to verify the features output by the fourth convolutional layer.

[0085] Each convolutional layer in the third convolutional layer and the fourth convolutional layer can be connected with a BN layer and a ReLU activation layer. In this way, every time an upsampling module is passed through, the size of the feature map becomes larger and the number of output channels decreases.

[0086] The convolutional module includes a fifth convolutional layer, a sixth convolutional layer, a feature verification layer, and a seventh convolutional layer connected in sequence. The feature verification layer in the convolutional module is configured to verify the features output by the sixth convolutional layer. The output of the convolutional module can be a segmentation result of three channels (or four channels), which is used to identify two types of bronchiectasis categories of background and foreground (when the output is four channels, it is three types of bronchiectasis categories of the foreground). The output of the model has the same spatial resolution scale as the input.

[0087] In addition, since the features of different layers and different channels have different degrees of importance, and the conventional downsampling layer operations cannot spatially distinguish the finest-grained features (such as some small bronchiectasis lesions) and the coarsest-grained features (such as large cystic bronchiectasis). Therefore, in the embodiments of the present disclosure, a feature verification layer is provided in the downsampling module, the upsampling module, and the convolutional module, and the feature verification layer can re-verify the features output by the convolution to retain relatively important regions.

[0088] Figure 3 It is a flowchart of data processing of a feature verification layer shown in an exemplary embodiment of the present disclosure. Refer to Figure 3 , the data processing flow of the feature verification layer includes: In step S21, the activation features input to the feature verification layer are obtained.

[0089] For example, for the feature verification layer in the downsampling module, the activation features output by the second convolutional layer can be obtained.

[0090] In step S22, the activation features are projected in the depth direction, height direction, and width direction respectively to obtain corresponding first features, second features, and third features.

[0091] In one implementation, in the depth direction, based on the height weight and width weight in the current feature verification layer, the features of the activation features in the height and width are weighted and summed to obtain the first feature.

[0092] In one embodiment, in the height direction, based on the depth weight and width weight in the current feature verification layer, the features of the activation feature in the depth and width directions may be weighted and summed to obtain the second feature.

[0093] In one embodiment, in the width direction, based on the height weight and depth weight in the current feature verification layer, the features of the activation feature in the height and depth directions may be weighted and summed to obtain the third feature.

[0094] Figure 4 is a flowchart of data processing of a feature verification layer shown in an exemplary embodiment of the present disclosure. Refer to Figure 4 , the activation feature may be, for example, the activation feature output by the m-th convolutional module, denoted as . Its shape is , represents the number of channels, respectively represent the depth, height, and width of. The purpose of the feature verification layer is to find the mapping to generate a calibrated weight so as to recalibrate the activation feature . This weight can gradually bias towards important regions while ignoring uninformative regions.

[0095] Exemplarily, in the depth direction, the activation feature can be projected in the following manner to obtain the corresponding first feature : ， In addition, in the height direction, the activation feature can be projected to obtain the corresponding second feature : ， And, in the width direction, the activation feature can be projected to obtain the corresponding third feature : 。

[0096] Wherein, represents the number of channels, respectively represent the depth, height, and width of the activation feature . is the depth weight, is the height weight, is the width weight. is the value of the activation feature at the height j and width k positions across all channel dimensions and depth dimensions. is the value of the activation feature at the depth i and width k positions across all channel dimensions and height dimensions. is the value of the activation feature at the depth i and height j positions across all channel dimensions and width dimensions, where R is the real number space.

[0097] Among them, and and can be initialized to random values from a distribution with a mean of 0 and a variance of 1. After each convolutional module, these parameters can be learned and adjusted and passed to subsequent processing stages along with other model parameters.

[0098] Referring to Figure 3 , in step S23, the first feature, the second feature, and the third feature are broadcast processed, and the broadcast processed first feature, second feature, and third feature are summed to obtain a fourth feature.

[0099] Continuing with the above example, the fourth feature can be obtained through the following calculation formula : 。

[0100] Among them, B(⋅) is a broadcast operation that can broadcast projection results of different dimensions to the same dimension for addition. In step S23, by aligning and adding projection results of different dimensions, the obtained fourth feature can be a comprehensive spatial feature. In this way, it helps to merge features in different directions, thereby providing a spatial feature with a global perspective, which in turn helps with subsequent verification operations.

[0101] Referring to Figure 3 , in step S24, the fourth feature is subjected to channel compression processing through a first convolutional kernel, and the fourth feature after channel compression processing is subjected to channel restoration through a second convolutional kernel to obtain a calibration weight.

[0102] For example, in one implementation, the calibration weight can be obtained through the following calculation formula : 。

[0103] Among them, and are convolutional kernels, represents a convolution operation, represents the non-linear activation function ReLU. In combination with Figure 4 for illustration,Figure 4 The ratio r in is the compression factor, which can be used to reduce the number of output channels. In step S24, the number of channels can be reduced to by performing convolution using , and the number of channels can be restored to by performing convolution using . By reducing and increasing the number of channels in this way, multiple channels can be reorganized, thereby emphasizing the information channels and suppressing redundant channels.

[0104] Referring to Figure 3 , in step S25, the calibrated weights and activation features are multiplied element-wise to obtain the calibrated features output by the feature calibration layer.

[0105] Exemplarily, the calibrated features can be calculated by the following calculation formula : .

[0106] Where is the activation feature, is the calibrated weight, and is the element-wise multiplication operation.

[0107] In this way, by setting the feature calibration layer, the feature calibration layer can recheck the features output by the convolution to retain relatively important regions. Thus, it helps to improve the data processing effect of the model.

[0108] It should be noted that the number of downsampling modules and upsampling modules of the first processing model can be set based on requirements. Figure 5 is a schematic structural diagram of a model shown in the present disclosure. Referring to Figure 5 , in one embodiment, the downsampling module includes a first downsampling module, a second downsampling module, a third downsampling module, and a fourth downsampling module (corresponding to convolution modules 1 to 4 in Figure 5 respectively), and the upsampling module includes a first upsampling module, a second upsampling module, a third upsampling module, and a fourth upsampling module (corresponding to convolution modules 5 to 8 in Figure 5 respectively). Where out_channel is the number of output channels, Conv is convolution, and Project&Excite is the feature calibration layer.

[0109] Referring to Figure 5 , the outputs of the first downsampling module and the fourth upsampling module serve as the inputs of the convolution module; the outputs of the second downsampling module and the third upsampling module serve as the inputs of the fourth upsampling module; The outputs of the third downsampling module and the second upsampling module serve as the inputs of the third upsampling module; The outputs of the fourth downsampling module and the first upsampling module serve as the inputs of the second upsampling module; The output of the fourth downsampling module serves as the input of the first upsampling module.

[0110] In this way, through skip connections, detailed information can be retained, which helps to enhance the feature fusion effect.

[0111] The following gives an exemplary description of the training method of the first processing model.

[0112] Refer to Figure 1 , in step S13, CT image data of the lungs is obtained.

[0113] In one implementation, when obtaining the CT image data of the lungs, the CT image data can be processed. For example, the CT image data can be subjected to data verification, and when the verification passes, the CT image data is determined to be valid data, and subsequent processing can be performed based on the CT image data.

[0114] Specifically, in one embodiment, the availability of the CT image data can be verified. Exemplarily, the CT image data can be in DICOM (Digital Imaging and Communications in Medicine) format. When the number of layers of the DICOM sequence is small and / or the slice thickness is large, relatively detailed information may not be provided. Therefore, the number of layers and / or the slice thickness of the DICOM sequence can be verified. When the number of layers of the DICOM sequence is less than the layer threshold (such as 60) and / or the slice thickness (Slice Thickness) of a certain case is greater than the thickness threshold (such as 2 mm), it is determined that the CT image data is abnormal and the verification fails. Otherwise, it can be determined that the verification passes.

[0115] In one embodiment, integrity verification can be performed on CT image data. For example, it can be verified whether there are missing slices in the DICOM sequence. During the verification, the slices in the DICOM sequence can be sorted based on the unique sequence code (Instance Number) of each slice. In this way, the slice spacing (Spacing Between Slices) between adjacent slices can be checked based on the relative coordinate position (SliceLocation) of each slice, and the slice spacing can be determined through the metadata in the DICOM file. When the slice spacing between two adjacent slices is greater than the spacing threshold, it can be determined that the CT image data is incomplete, that is, the integrity verification fails. Among them, the spacing threshold can be set to 2 times or more of the slice spacing between other adjacent slices in the sorting. Additionally, when the slice spacing between each adjacent slice is less than or equal to the spacing threshold, it can be determined that the integrity verification passes.

[0116] In one embodiment, repeatability verification can be performed on CT image data. For example, it can be verified whether there are duplicate slices. During implementation, it can be checked whether there are duplicate slice sequence codes, and if there are duplicate slice codes, the redundant slices can be deleted.

[0117] Referring to Figure 1 , in step S14, the CT image data is input into the first processing model to obtain the bronchiectasis lesions determined according to the CT image data output by the first processing model, and the type of the bronchiectasis lesions.

[0118] The first processing model can be trained, for example, by taking sample CT image data as input and taking the sample bronchiectasis lesions determined according to the sample CT image data and the type of the sample bronchiectasis lesions as output.

[0119] After obtaining the bronchiectasis lesions and the type of the bronchiectasis lesions, referring to Figure 1 , in step S15, according to the determined bronchiectasis lesions and the type of the bronchiectasis lesions, the processing result of the CT image data is output.

[0120] In one possible implementation manner, the determined bronchiectasis lesions and the type of the bronchiectasis lesions can be output.

[0121] In one possible implementation manner, the processing result of the CT image data can also be output in combination with other information.

[0122] For example, in one possible implementation manner, the method further includes: Input the CT image data into a second processing model to obtain a segmentation result of the CT image data output by the second processing model, where the segmentation result includes lobe information in the CT image. The second processing model is trained by using sample CT image data as input and the sample segmentation result of the lobe determined according to the sample CT image data as output.

[0123] Here, the sample CT image data may also correspond to third annotation information. The third annotation information includes lobe annotation information for the sample CT image data. Exemplarily, 6 lobes can be annotated with 6 different types of labels, and the number of annotated sample CT image data can be greater than or equal to 200. In some scenarios, the 5 lobes and the lingula may be separately annotated. Therefore, before training, it can be detected whether the annotation class value of the lingula is different from that of the 5 lobes. If the annotation values overlap, the annotation class value of the lingula can be modified, such as by logical operations and multiplication operations to find and modify the annotation class value of the annotated pixel points.

[0124] In addition, referring to the description of the structure of the first processing model in the above embodiments, except that the final output of the second processing model is 7 channels, the remaining initial model structure can be the same as that of the first processing model. Similarly, during training, operations such as data verification and data preprocessing can also be performed on the samples for training the second processing model. For the sake of brevity of the specification, the embodiments of the present disclosure will not elaborate on this.

[0125] In this way, the CT image data can be input into the second processing model to obtain a segmentation result of the CT image data output by the second processing model, where the segmentation result includes lobe information in the CT image. In this case, the processing result of the CT image data output according to the determined bronchiectasis lesions and the type of the bronchiectasis lesions includes: Determine the bronchiectasis lesions included in each lobe and the type of the bronchiectasis lesions in the lobe according to the determined bronchiectasis lesions, the type of the bronchiectasis lesions, and the lobe information; Output the processing result of the CT image data according to the bronchiectasis lesions included in each lobe and the type of the bronchiectasis lesions in the lobe.

[0126] Figure 6 is a processing flow chart of CT image data shown in an exemplary embodiment of the present disclosure. Refer to Figure 6 In one implementation, the bronchiectasis segmentation result and the lobe segmentation result can be obtained. The bronchiectasis segmentation result includes the bronchiectasis lesions and the type of the bronchiectasis lesions.

[0127] In this way, by combining 3D connected component analysis and annotation category values, the categories, quantities, and locations of 3D lesions can be counted. On this basis, by combining the lung lobe locations and category values, the number of lesions and corresponding categories in each lung lobe can be counted, and scores, total scores, and average scores can be calculated. Finally, the modified Reiff score can be calculated based on the number of different types of lesions in each lung lobe.

[0128] In this way, the automatic calculation of the modified Reiff score can be achieved, which helps to improve the processing efficiency.

[0129] In a possible implementation manner, the processing result of the CT image data output according to the bronchiectasis lesions included in each lung lobe and the types of the bronchiectasis lesions in the lung lobe includes: According to the bronchiectasis lesions included in each lung lobe and the types of the bronchiectasis lesions, a reference value is calculated. The reference value may include, for example, the modified Reiff score. In addition, symptom description information can be obtained, and semantic expansion is performed on the symptom description information to obtain target description information. In this way, according to the target description information and the reference value, a knowledge base of pathogenic factors is queried to obtain prompt information of pathogenic factors.

[0130] Figure 7 is a processing flow chart of CT image data shown in an exemplary embodiment of the present disclosure. Refer to Figure 7 , in one implementation manner, symptom description information can be obtained, and the symptom description information may include, for example, lung lobe involvement and accompanying signs.

[0131] Exemplarily, the system can give check boxes for common pulmonary accompanying signs, and the user checks the pulmonary accompanying signs or manually enters the accompanying signs. In this way, the corresponding pulmonary accompanying signs can be obtained.

[0132] In one implementation manner, the symptom description information can be represented in XML (Extensible Markup Language) and input into the query system. In this way, the query system can perform natural language understanding on the XML-formatted input and convert it into the RDF (Resource Description Framework) format. In addition, the query system can perform semantic expansion on the input, such as synonym, near-synonym, association, hyponymy, etc. expansion, to obtain target description information.

[0133] In this way, based on the modified Reiff score and the target description information, a knowledge base of pathogenic factors is queried to obtain prompt information of pathogenic factors.

[0134] Exemplarily, a knowledge base of pathogenic factors can be established ( Figure 7 schematically shown as an expert semantic knowledge base for etiological hints in Figure 7 ). The knowledge base is described using ontology, RDF, RDF Schema, and SWRL (Semantic Web Rule Language). Among them, ontology can be a set of abstract concepts in a domain, capable of describing the common features of things within the scope and the relationships between things. In the medical field, ontology can define concepts such as "symptom", "pathogenic factor", etc., and the relationships between these concepts. RDF is a framework for representing information, which describes resources and their relationships in the form of triples (subject-predicate-object), and each triple represents a simple semantic relationship. RDF Schema is a standard for describing the structure and vocabulary of the RDF data model. It extends RDF to enable the definition of classes and properties, thereby providing richer descriptive capabilities. SWRL can write rules that can perform reasoning based on existing data. For example, possible pathogenic factors can be inferred based on specific symptom descriptions and imaging manifestations.

[0135] In this way, by establishing a knowledge base of pathogenic factors, the knowledge base of pathogenic factors can be queried based on the improved Reiff score and the target description information, and hint information of pathogenic factors can be obtained. For example, the hint information of pathogenic factors can be queried in Sparql language, and the query result can be converted into XML format for output.

[0136] Exemplarily, when bronchiectasis occurs and is accompanied by extrapulmonary manifestations such as dextrocardia or situs inversus or sinusitis or infertility, etc., the system can roughly give a hint of primary ciliary dyskinesia (Kartagener syndrome) based on the above situation. Another example is that when obvious tree-in-bud signs (peripheral airway mucus embolism) appear in the lungs and bronchiectasis occurs in more than 4 lung lobes, and columnar bronchiectasis mainly presents in each lung lobe, the system can roughly give a hint of diffuse panbronchiolitis based on the above situation.

[0137] By adopting the above solution, multiple training sets can be generated based on the combination of the types of bronchiectasis lesions, and the first processing model can be trained in sequence through the multiple training sets. In this way, through multi-level training, it helps to improve the model training effect. In addition, the CT image data can be processed by the trained first processing model to determine the bronchiectasis lesions and the types of the bronchiectasis lesions. In this way, it helps to improve the processing efficiency of the CT image data and ensure the consistency of the processing results.

[0138] Moreover, the above solution can achieve automatic recognition, segmentation, and localization of bronchiectasis lesions, so as to improve the accuracy, consistency, and efficiency of bronchiectasis recognition and classification. The above solution can also automatically calculate the modified Reiff score and count the number of different types of lesions in each lung lobe, so as to facilitate clinicians to comprehensively evaluate the severity of bronchiectasis in patients from multiple aspects. The above solution can also automatically give etiological hints and physical sputum drainage suggestions according to the category, distribution, and number of lesions. The above solution can also be integrated with the PACS system and incorporated into the daily workflows of radiologists, respiratory physicians, thoracic surgeons, and interventional physicians to assist them in improving data processing efficiency.

[0139] Based on the same inventive concept, an embodiment of the present disclosure provides a processing device for CT image data. Figure 8 is a block diagram of a processing device for CT image data shown in an exemplary embodiment of the present disclosure. Referring to Figure 8 , the device includes: A generation module 901, configured to generate a plurality of training sets according to a combination of types of bronchiectasis lesions. Different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data, where the first annotation information includes the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions; A training module 902, configured to sequentially train a model to be trained according to the plurality of training sets to obtain a first processing model; A first module 903, configured to obtain CT image data of the lungs; A second module 904, configured to input the CT image data into the first processing model to obtain bronchiectasis lesions determined according to the CT image data output by the first processing model, and the types of the bronchiectasis lesions; A third module 905, configured to output a processing result of the CT image data according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions; wherein the first processing model is trained by taking sample CT image data as input and taking sample bronchiectasis lesions determined according to the sample CT image data and the types of the sample bronchiectasis lesions as output.

[0140] By adopting the above solution, multiple training sets can be generated based on the combination of the types of bronchiectasis lesions, and the first processing model can be obtained by training successively through the multiple training sets. In this way, through multi-level training, it helps to improve the model training effect. In addition, the CT image data can be processed by the trained first processing model to determine the bronchiectasis lesions and the types of the bronchiectasis lesions. In this way, it helps to improve the processing efficiency of the CT image data and ensure the consistency of the processing results.

[0141] Optionally, the generating module 901 includes: A first sub-module, configured to generate a first training set, where the first training set includes first CT image data and second CT image data, the first CT image data has at least columnar bronchiectasis lesions, and the second CT image data has at least varicose bronchiectasis lesions; A second sub-module, configured to generate a second training set, where the second training set includes the second CT image data and third CT image data, and the third CT image data has at least cystic bronchiectasis lesions; A third sub-module, configured to generate a third training set, where the third training set includes the first CT image data, the second CT image data, and the third CT image data.

[0142] Optionally, the training module 902 includes: A fourth sub-module, configured to train the model to be trained through the first training set to obtain a first model; A fifth sub-module, configured to train the first model through the second training set to obtain a second model; A sixth sub-module, configured to train the second model through the third training set to obtain the first processing model.

[0143] Optionally, the generating module 901 includes: A seventh sub-module, configured to obtain a CT image data set, where the CT image data set includes fourth CT image data and second annotation information of the fourth CT image data, and the second annotation information includes the bronchiectasis lesions included in the fourth CT image and the types of the bronchiectasis lesions; An eighth sub-module, configured to verify the fourth CT image data to obtain verified fourth CT image data; A ninth sub-module, configured to generate multiple training sets according to the combination of the verified fourth CT image data, the second annotation information of the verified fourth CT image data, and the types of the bronchiectasis lesions.

[0144] Optionally, the device includes: A fourth module, configured to input the CT image data into a second processing model, and obtain a segmentation result of the CT image data output by the second processing model, where the segmentation result includes lobe information in the CT image; the second processing model is trained by using sample CT image data as input and using a sample segmentation result of a lobe determined according to the sample CT image data as output; The third module 905 includes: A tenth sub-module, configured to determine bronchiolar dilatation lesions included in each lobe, and types of the bronchiolar dilatation lesions in the lobe according to the determined bronchiolar dilatation lesions, types of the bronchiolar dilatation lesions, and the lobe information; An eleventh sub-module, configured to output a processing result of the CT image data according to the bronchiolar dilatation lesions included in each lobe and types of the bronchiolar dilatation lesions in the lobe.

[0145] Optionally, the eleventh sub-module is configured to: Calculate a reference value according to the bronchiolar dilatation lesions included in each lobe and types of the bronchiolar dilatation lesions; Obtain symptom description information; Perform semantic expansion on the symptom description information to obtain target description information; Query a knowledge base of pathogenic factors according to the target description information and the reference value to obtain prompt information of pathogenic factors.

[0146] Optionally, the first processing model includes a downsampling module, an upsampling module, and a convolution module that are connected in sequence, The downsampling module includes a first convolutional layer, a second convolutional layer, a feature verification layer, and a downsampling layer that are connected in sequence, and the feature verification layer is configured to verify features output by the second convolutional layer; The upsampling module includes a third convolutional layer, a fourth convolutional layer, a feature verification layer, and an upsampling layer that are connected in sequence, and the feature verification layer in the upsampling module is configured to verify features output by the fourth convolutional layer; The convolution module includes a fifth convolutional layer, a sixth convolutional layer, a feature verification layer, and a seventh convolutional layer that are connected in sequence, and the feature verification layer in the convolution module is configured to verify features output by the sixth convolutional layer.

[0147] Optionally, the feature verification layer is configured to: Obtain activation features input to the feature verification layer; Project the activation feature in the depth direction, height direction, and width direction respectively to obtain the corresponding first feature, second feature, and third feature; Perform broadcast processing on the first feature, second feature, and third feature, and sum the broadcast-processed first feature, second feature, and third feature to obtain a fourth feature; Perform channel compression processing on the fourth feature through a first convolutional kernel, and perform channel restoration on the fourth feature after channel compression processing through a second convolutional kernel to obtain a calibration weight; Perform element-wise multiplication on the calibration weight and the activation feature to obtain the verified feature output by the feature verification layer.

[0148] Optionally, the feature verification layer projects the activation feature in the depth direction, height direction, and width direction respectively, including: In the depth direction, based on the height weight and width weight in the current feature verification layer, perform weighted summation on the features in the height and width of the activation feature to obtain the first feature; In the height direction, based on the depth weight and width weight in the current feature verification layer, perform weighted summation on the features in the depth and width of the activation feature to obtain the second feature; In the width direction, based on the height weight and depth weight in the current feature verification layer, perform weighted summation on the features in the height and depth of the activation feature to obtain the third feature.

[0149] Optionally, the downsampling module includes a first downsampling module, a second downsampling module, a third downsampling module, and a fourth downsampling module, and the upsampling module includes a first upsampling module, a second upsampling module, a third upsampling module, and a fourth upsampling module, where The outputs of the first downsampling module and the fourth upsampling module serve as the inputs to the convolutional module; The outputs of the second downsampling module and the third upsampling module serve as the inputs to the fourth upsampling module; The outputs of the third downsampling module and the second upsampling module serve as the inputs to the third upsampling module; The outputs of the fourth downsampling module and the first upsampling module serve as the inputs to the second upsampling module; The output of the fourth downsampling module serves as the input to the first upsampling module.

[0150] Embodiments of the present disclosure also provide a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the steps of the method for processing CT image data described in any one of the embodiments of the present disclosure are implemented.

[0151] Embodiments of the present disclosure also provide an electronic device, including: a memory having a computer program stored thereon; a processor configured to execute the computer program in the memory to implement the steps of the method for processing CT image data described in any one of the embodiments of the present disclosure.

[0152] Regarding the device in the above embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0153] Figure 9 is a block diagram of an electronic device 700 shown according to an exemplary embodiment. As Figure 9 shown, the electronic device 700 may include: a processor 701, a memory 702. The electronic device 700 may also include one or more of a multimedia component 703, an input / output (I / O) interface 704, and a communication component 705.

[0154] Among them, the processor 701 is used to control the overall operation of the electronic device 700 to complete all or part of the steps in the above CT image data processing method. The memory 702 is used to store various types of data to support the operation of the electronic device 700. These data may include, for example, instructions for any application or method operating on the electronic device 700, as well as application-related data, such as contact data, messages sent and received, pictures, audio, video, and so on. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disc. The multimedia component 703 may include a screen and an audio component. Among them, the screen may be a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone, and the microphone is used to receive external audio signals. The received audio signals may be further stored in the memory 702 or sent through the communication component 705. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 704 provides an interface between the processor 701 and other interface modules, and the above other interface modules may be a keyboard, a mouse, buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, near field communication (NFC), 2G, 3G or 4G, or a combination of one or more of them. Therefore, the corresponding communication component 705 may include: a Wi-Fi module, a Bluetooth module, and an NFC module.

[0155] In an exemplary embodiment, the electronic device 700 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components, and is used to execute the above-described method for processing CT image data.

[0156] In another exemplary embodiment, a computer-readable storage medium including program instructions is further provided. When the program instructions are executed by a processor, the steps of the above-described method for processing CT image data are implemented. For example, the computer-readable storage medium may be the above-described memory 702 including program instructions, and the above program instructions may be executed by the processor 701 of the electronic device 700 to complete the above-described method for processing CT image data.

[0157] The preferred embodiments of the present disclosure have been described in detail above in conjunction with the accompanying drawings. However, the present disclosure is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0158] In addition, it should be noted that, in the above specific embodiments, the various specific technical features described can be combined in any suitable manner without conflict. To avoid unnecessary repetition, the present disclosure will not separately describe various possible combination manners.

[0159] Furthermore, any combination can be made between various different embodiments of the present disclosure as long as it does not violate the idea of the present disclosure, and it should also be regarded as the content disclosed by the present disclosure.

Claims

1. A method for processing CT image data, characterized in that, Including: Generating a plurality of training sets according to a combination of types of bronchiectasis lesions, where different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data, and the first annotation information includes the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions; Successively training a model to be trained according to the plurality of training sets to obtain a first processing model; Obtaining CT image data of the lungs; Inputting the CT image data into the first processing model to obtain bronchiectasis lesions determined according to the CT image data output by the first processing model and the types of the bronchiectasis lesions; Outputting a processing result of the CT image data according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions.

2. The method according to claim 1, wherein The generating a plurality of training sets according to a combination of types of bronchiectasis lesions includes: Generating a first training set, where the first training set includes first CT image data and second CT image data, the first CT image data has at least columnar bronchiectasis lesions, and the second CT image data has at least varicose bronchiectasis lesions; Generating a second training set, where the second training set includes the second CT image data and third CT image data, and the third CT image data has at least cystic bronchiectasis lesions; Generating a third training set, where the third training set includes the first CT image data, the second CT image data, and the third CT image data.

3. The method according to claim 2, wherein The successively training a model to be trained according to the plurality of training sets to obtain a first processing model includes: Training the model to be trained through the first training set to obtain a first model; Training the first model through the second training set to obtain a second model; Training the second model through the third training set to obtain the first processing model.

4. The method according to claim 1, wherein The generating a plurality of training sets according to a combination of types of bronchiectasis lesions includes: Obtaining a CT image data set, where the CT image data set includes fourth CT image data and second annotation information of the fourth CT image data, and the second annotation information includes the bronchiectasis lesions included in the fourth CT image and the types of the bronchiectasis lesions; Verifying the fourth CT image data to obtain verified fourth CT image data; Generating a plurality of training sets according to the verified fourth CT image data, the second annotation information of the verified fourth CT image data, and the combination of the types of the bronchiectasis lesions.

5. The method according to claim 1, wherein Including: Inputting the CT image data into a second processing model to obtain a segmentation result of the CT image data output by the second processing model, and the segmentation result includes lobe information in the CT image; The second processing model is trained by taking sample CT image data as input and taking a sample segmentation result of a lobe determined according to the sample CT image data as output. Output the processing result of the CT image data according to the determined bronchiectasis lesions and the type of the bronchiectasis lesions, including: Determine the bronchiectasis lesions included in each lung lobe and the type of the bronchiectasis lesions in the lung lobe according to the determined bronchiectasis lesions, the type of the bronchiectasis lesions and the lung lobe information; Output the processing result of the CT image data according to the bronchiectasis lesions included in each lung lobe and the type of the bronchiectasis lesions in the lung lobe.

6. The method according to claim 5, wherein The output of the processing result of the CT image data according to the bronchiectasis lesions included in each lung lobe and the type of the bronchiectasis lesions in the lung lobe includes: Calculate a reference value according to the bronchiectasis lesions included in each lung lobe and the type of the bronchiectasis lesions; Obtain symptom description information; Perform semantic expansion on the symptom description information to obtain target description information; Query the knowledge base of pathogenic factors according to the target description information and the reference value to obtain prompt information of pathogenic factors.

7. The method according to any one of claims 1 to 6, characterized in that The first processing model includes a downsampling module, an upsampling module and a convolution module connected in sequence, The downsampling module includes a first convolutional layer, a second convolutional layer, a feature verification layer and a downsampling layer connected in sequence, and the feature verification layer is configured to verify the features output by the second convolutional layer; The upsampling module includes a third convolutional layer, a fourth convolutional layer, a feature verification layer and an upsampling layer connected in sequence, and the feature verification layer in the upsampling module is configured to verify the features output by the fourth convolutional layer; The convolution module includes a fifth convolutional layer, a sixth convolutional layer, a feature verification layer and a seventh convolutional layer connected in sequence, and the feature verification layer in the convolution module is configured to verify the features output by the sixth convolutional layer.

8. The method according to claim 7, wherein The feature verification layer is configured to: Obtain the activation features input to the feature verification layer; Project the activation features in the depth direction, height direction and width direction respectively to obtain corresponding first features, second features and third features; Perform broadcast processing on the first feature, the second feature and the third feature, and sum the broadcast-processed first feature, second feature and third feature to obtain a fourth feature; Perform channel compression processing on the fourth feature through a first convolutional kernel, and perform channel restoration on the fourth feature after channel compression processing through a second convolutional kernel to obtain a calibration weight; Perform element-wise multiplication on the calibration weight and the activation features to obtain the verified features output by the feature verification layer.

9. The method according to claim 8, wherein The projection of the activation features in the depth direction, height direction and width direction respectively includes: In the depth direction, based on the height weight and width weight in the current feature verification layer, perform weighted summation on the features in the height and width of the activation features to obtain the first feature; In the height direction, based on the depth weight and width weight in the current feature verification layer, the features of the activation feature in the depth and width directions are weighted and summed to obtain the second feature; In the width direction, based on the height weight and depth weight in the current feature verification layer, the features of the activation feature in the height and depth directions are weighted and summed to obtain the third feature.

10. The method according to claim 7, characterized in that, The downsampling module includes a first downsampling module, a second downsampling module, a third downsampling module, and a fourth downsampling module. The upsampling module includes a first upsampling module, a second upsampling module, a third upsampling module, and a fourth upsampling module, where the outputs of the first downsampling module and the fourth upsampling module serve as the inputs to the convolutional module; the outputs of the second downsampling module and the third upsampling module serve as the inputs to the fourth upsampling module; the outputs of the third downsampling module and the second upsampling module serve as the inputs to the third upsampling module; the outputs of the fourth downsampling module and the first upsampling module serve as the inputs to the second upsampling module; the output of the fourth downsampling module serves as the input to the first upsampling module.

11. A processing device for CT image data, characterized in that, Comprising: a generation module, configured to generate a plurality of training sets according to combinations of types of bronchiectasis lesions. Different training sets include CT image data corresponding to bronchiectasis lesions of different type combinations, and first annotation information of the CT image data. The first annotation information includes the bronchiectasis lesions included in the CT image and the types of the bronchiectasis lesions; a training module, configured to sequentially train a model to be trained according to the plurality of training sets to obtain a first processing model; a first module, configured to obtain CT image data of the lungs; a second module, configured to input the CT image data into the first processing model to obtain the bronchiectasis lesions determined according to the CT image data output by the first processing model and the types of the bronchiectasis lesions; a third module, configured to output a processing result of the CT image data according to the determined bronchiectasis lesions and the types of the bronchiectasis lesions.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

13. An electronic device, characterized in that, Comprising: a memory, on which a computer program is stored; a processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1 to 10.

Images