CT sequence image processing method and device and storage medium

Abstract

The invention discloses a CT sequence image processing method and device and a storage medium, and the method comprises the steps: obtaining a to-be-processed CT sequence image; performing dimension reduction processing on the CT sequence image through a bone pressing average density projection method to obtain a two-dimensional projection image; and processing the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of the CT sequence image. According to the method, the continuous CT cross-sectional images are subjected to dimensionality reduction to form the two-dimensional projection images, the data scale is reduced, the calculation complexity is reduced, the requirement for hardware resources is remarkably reduced, the calculation speed is remarkably increased, the two-dimensional projection images are processed by adopting the lightweight classification model based on the reduced data scale, and the calculation efficiency is improved. The inference speed can be improved, large-batch CT image data can be analyzed more quickly, and then the efficiency of analyzing and processing the CT sequence images is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a method, a device and a storage medium for processing CT sequence images. Background Art

[0002] With the development of medical artificial intelligence, the demand for CT image data is increasing, and the corresponding CT data processing is becoming more and more important. Plain scan and enhanced sequences in CT images have important complementarity in clinical diagnosis: plain scan images provide basic structure and density information of tissues, which can be used for preliminary detection and location of abnormal areas; while enhanced sequences enhance the contrast of blood vessels and tissues by injecting contrast agents, which can more clearly identify and evaluate tumors, infections or other abnormalities. Therefore, the distinction between plain scan and enhanced sequences is of great significance.

[0003] However, the labels of plain scan and enhanced sequences of CT image data collected from clinical practice are stored in the metadata of Digital Imaging and Communications in Medicine (DICOM) files. Most of these data rely on manual clinical input, and are often inaccurately and incompletely labeled, making them difficult to use directly. Manual verification of these data is time-consuming and labor-intensive, which increases the cost of CT data collection exponentially. Therefore, it is necessary to develop an automated data cleaning model that can accurately and efficiently identify CT plain scan and enhanced sequences.

[0004] Since deep learning has shown excellent performance in image classification tasks, Ye et al. have proposed using deep convolutional neural networks (CNN) to classify plain scan and enhanced sequences. Since CT is a tomographic image, a single sequence often contains dozens to hundreds of tomographic images. The existing technology usually treats each layer of images in the CT scan as a separate individual, inputs each layer of images independently into the CNN for prediction, generates a local prediction result, and finally aggregates it into a single label for the entire sequence. That is to say, in a CT sequence, the model actually performs a complete reasoning process on each tomographic image in the sequence, and outputs independent prediction results, and finally averages them to obtain the final single plain scan or enhanced label.

[0005] The existing technology mainly adopts the method of inputting all the tomographic images in the CT scan into the CNN network for analysis, but there are some obvious shortcomings: 1. High hardware resource requirements: The current method requires that all tomographic images of the entire CT scan be loaded into the graphics processing unit (GPU) of the computer at the same time. Unit (GPU), which leads to high GPU resource requirements and usually requires expensive hardware equipment to support, which may be impractical for some medical institutions and research teams; 2. It is difficult to meet the processing requirements of large batches of data: As the amount of CT image data increases, large quantities of CT image data usually need to be processed simultaneously in the real data cleaning process, such as large-scale cancer screening or long-term follow-up CT images for large population cohorts; the current method inputs all images of the same sequence into the CNN network at the same time, especially in large-scale data processing, which will lead to a significant increase in computing time and resource costs, which limits the scalability of the technology and makes it difficult to meet the requirements of high throughput; 3. Incoherence and introduction of irrelevant information: Independently predicting each image may lead to incoherence between the output results, especially for CT tomographic images, not every layer of tomographic images contains the corresponding features of enhanced scanning; the traditional method passes all images to the CNN network, which may introduce a lot of irrelevant information, which may not contribute to the final analysis and diagnosis, but increase the complexity of the calculation and reduce the accuracy of the analysis. These shortcomings affect its performance and efficiency in medical application scenarios.

[0006] There are many deficiencies in the way existing technologies process CT images, which limits the applicability and efficiency of this technology in practical applications. Therefore, new methods and technologies are needed to address these challenges. Summary of the invention

[0007] The present invention provides a method, device and storage medium for processing CT sequence images to solve the above technical problems and effectively improve the efficiency of analyzing and processing CT sequence images.

[0008] In order to solve the above technical problems, the present invention provides a method for processing CT sequence images, comprising:

[0009] Acquire CT sequence images to be processed;

[0010] Performing dimension reduction processing on the CT sequence images by using a bone compression average density projection method to obtain a two-dimensional projection image;

[0011] The two-dimensional projection image is processed based on a preset lightweight classification model to obtain a discrimination result of the CT sequence image.

[0012] The beneficial effects of the present invention are:

[0013] The present invention performs dimensionality reduction processing on the CT sequence images to be processed by the bone compression average density projection method to obtain a two-dimensional projection image, and processes the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of the CT sequence image. Compared with the prior art that all tomographic images in the CT scan are input into a processing network for analysis, resulting in a significant increase in computing time and resource costs, the present invention reduces the data scale by reducing the dimension of continuous CT tomographic images into two-dimensional projection images, which helps to reduce the computational complexity and significantly reduce the requirements for hardware resources, and significantly improves the computing speed. In addition, based on the reduced data scale, the present invention uses a lightweight classification model to process the two-dimensional projection image, which can improve the reasoning speed and analyze large quantities of CT image data more quickly, thereby effectively improving the efficiency of analyzing and processing CT sequence images.

[0014] As a preferred solution, the dimensionality reduction processing of the CT sequence images by the bone compression average density projection method to obtain a two-dimensional projection image includes:

[0015] Based on the CT sequence images, the pixel value of each pixel in each image layer is obtained;

[0016] In each image layer, the pixel value of each pixel is converted into the initial CT value to obtain the CT value of all pixels;

[0017] Through the bone compression average density projection method, bone compression processing is performed on each image layer based on the bone compression algorithm and CT value to obtain each bone compression image layer;

[0018] Projection calculation is performed on each bone compression image layer to obtain a two-dimensional projection image.

[0019] The present invention achieves a reduction in data dimension by converting continuous CT tomographic images into two-dimensional projection images. This reduction in data dimension not only helps to reduce computational complexity, but also can significantly reduce the requirements for hardware resources and significantly improve the computing speed. As a result, large quantities of CT image data can be analyzed more quickly, thereby effectively improving the efficiency of analyzing and processing CT sequence images.

[0020] As a preferred solution, the bone compression algorithm is expressed by the following formula:

[0021]

[0022]

[0023] In the above formula, α is a preset hyperparameter; CL is a preset window position of the CT sequence images, and WW is a preset window width of the CT sequence images, and the preset window position and the preset window width are values ​​determined based on the contrast agent density range.

[0024] The present invention uses the contrast agent density range as a reference for the window center level (CL) and window width (WW) of the CT sequence image, which can maximize the density difference of the contrast agent, thereby improving the discrimination of plain scan and enhanced images. In addition, since high-density objects such as bones, calcifications, and metals have extremely high HU values ​​(generally greater than 1000HU) on CT images, the proportion of the contrast agent will be diluted during average density projection, interfering with the discrimination between plain scan and enhanced images. Therefore, the present invention suppresses the CT value of the high-density structure, that is, for the pixel points with a CT value greater than α, it is reassigned to the minimum value β of the CT value range in the image, thereby improving the discrimination of plain scan and enhanced images, so as to improve the discrimination accuracy of CT sequence images.

[0025] As a preferred solution, projection calculation is performed on each bone compression image layer using the following formula:

[0026]

[0027] In the above formula, BSAIP(y,z) is the CT value of the pixel point with coordinates (y,z) in the two-dimensional projection image; x is the sagittal axis coordinate of the CT sequence image, y is the coronal axis coordinate of the CT sequence image, and z is the transverse axis coordinate of the CT sequence image; n is the total number of pixels along the sagittal axis.

[0028] The present invention can display the average density of tissues or organs by performing projection calculations on each bone compression image layer, and achieves a reduction in data dimension by converting continuous CT tomographic images into two-dimensional projection images. This reduction in data dimension not only helps to reduce computational complexity, but also can significantly reduce the requirements for hardware resources and significantly improve the computing speed. As a result, large quantities of CT image data can be analyzed more quickly, thereby effectively improving the efficiency of analyzing and processing CT sequence images.

[0029] As a preferred solution, after the CT sequence images are subjected to dimension reduction processing by the bone compression average density projection method to obtain a two-dimensional projection image, the method further includes:

[0030] The two-dimensional projection image is resampled to update the two-dimensional projection image.

[0031] Since the pixel intervals in various directions of the CT sequence images may be inconsistent, the present invention can maintain the original anatomical proportions of the projected two-dimensional projection images through resampling.

[0032] As a preferred solution, the method for obtaining the preset lightweight classification model includes:

[0033] Obtaining initial CT sequence images and annotation labels of the initial CT sequence images;

[0034] Performing dimensionality reduction processing on the initial CT sequence images by using a bone compression average density projection method to obtain a CT projection image;

[0035] The preset basic model is trained and verified based on the CT projection images and the annotated labels to obtain a preset lightweight classification model.

[0036] As a preferred solution, after obtaining the CT projection image, the method further includes:

[0037] The CT projection image is preprocessed to update the CT projection image.

[0038] The present invention can increase the diversity of data through preprocessing. The model requires a large amount of data to train, but the original data may have problems such as sample imbalance, noise data and missing data. This solution can remove noise data, fill missing data, and balance the data by preprocessing the CT projection images, making the data more diverse and helping to train a better model.

[0039] As a preferred solution, the two-dimensional projection image is processed based on a preset lightweight classification model to obtain a discrimination result of a CT sequence image, including:

[0040] Inputting the two-dimensional projection image into a preset lightweight classification model; the lightweight classification model includes a MobileNetV3-small module and an MLP module;

[0041] Extracting image features of the two-dimensional projection image through a MobileNetV3-small module;

[0042] The image features are input into the MLP module to obtain the discrimination results of the CT sequence images.

[0043] Since the present invention reduces the data scale of CT sequence images, the present invention can use a lightweight convolutional neural network to analyze and process the two-dimensional projection images, thereby improving the reasoning speed and being able to analyze large quantities of CT image data more quickly; and, the present invention inputs the fused and reduced-dimensional two-dimensional projection images into the network through the bone compression average density projection method, avoiding passing all tomographic images to the lightweight classification model, and filtering out image changes that are not related to specific tasks to a certain extent, avoiding the incoherence of the output results and the introduction of irrelevant information, thereby improving the accuracy of the analysis.

[0044] Accordingly, in order to solve the above technical problems, the present invention also provides a CT sequence image processing device, comprising: an image acquisition module, a dimension reduction processing module and an image processing module;

[0045] Wherein, the image acquisition module is used to acquire the CT sequence images to be processed;

[0046] The dimension reduction processing module is used to perform dimension reduction processing on the CT sequence images by using a bone compression average density projection method to obtain a two-dimensional projection image;

[0047] The image processing module is used to process the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of a CT sequence image.

[0048] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for processing CT sequence images as described in any one of the above items are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 : A flow chart of an embodiment of a method for processing CT sequence images provided by the present invention;

[0050] Figure 2 : A spatial coordinate diagram of a CT sequence image provided by an embodiment of the present invention;

[0051] Figure 3 : A schematic diagram of a plain scan sequence and an enhanced sequence in a CT sequence image provided by an embodiment of the present invention;

[0052] Figure 4 : A schematic diagram of the structure of a preset basic model provided in an embodiment of the present invention;

[0053] Figure 5 : A structural schematic diagram of an embodiment of a CT sequence image processing device provided by the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The existing technology mainly adopts the method of inputting all the tomographic images in the CT scan into the CNN network for analysis, but there are some obvious shortcomings: 1. High hardware resource requirements: The current method requires that all tomographic images of the entire CT scan be loaded into the graphics processing unit (GPU) of the computer at the same time. Unit (GPU), which leads to high GPU resource requirements and usually requires expensive hardware equipment to support, which may be impractical for some medical institutions and research teams; 2. It is difficult to meet the processing requirements of large batches of data: As the amount of CT image data increases, large quantities of CT image data usually need to be processed simultaneously in the real data cleaning process, such as large-scale cancer screening or long-term follow-up CT images for large population cohorts; the current method inputs all images of the same sequence into the CNN network at the same time, especially in large-scale data processing, which will lead to a significant increase in computing time and resource costs, which limits the scalability of the technology and makes it difficult to meet the requirements of high throughput; 3. Incoherence and introduction of irrelevant information: Independently predicting each image may lead to incoherence between the output results, especially for CT tomographic images, not every layer of tomographic images contains the corresponding features of enhanced scanning; the traditional method passes all images to the CNN network, which may introduce a lot of irrelevant information, which may not contribute to the final analysis and diagnosis, but increase the complexity of the calculation and reduce the accuracy of the analysis. These shortcomings affect its performance and efficiency in medical application scenarios.

[0056] The existing technology has many deficiencies in the way it processes CT images, which limits the applicability and efficiency of the technology in practical applications. Therefore, new methods and technologies are needed to address these challenges. Improve the efficiency and accuracy of automated data cleaning of CT plain scan and enhanced sequence images.

[0057] Embodiment 1

[0058] In order to solve the above technical problems, the present invention provides a method, device and storage medium for processing CT sequence images, which effectively improve the efficiency of analyzing and processing CT sequence images.

[0059] Please refer to Figure 1 , is a flow chart of an embodiment of a method for processing CT sequence images provided in an embodiment of the present invention.

[0060] like Figure 1 As shown, the method for processing CT sequence images includes steps 101 to 103, specifically:

[0061] Step 101: Acquire a CT sequence image to be processed;

[0062] Step 102: performing dimension reduction processing on the CT sequence images by using a bone compression average density projection method to obtain a two-dimensional projection image;

[0063] Step 103: Processing the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of the CT sequence image.

[0064] The present invention performs dimensionality reduction processing on the CT sequence images to be processed by the bone compression average density projection method to obtain a two-dimensional projection image, and processes the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of the CT sequence image. Compared with the prior art that all tomographic images in the CT scan are input into a processing network for analysis, resulting in a significant increase in computing time and resource costs, the present invention reduces the data scale by reducing the dimension of continuous CT tomographic images into two-dimensional projection images, which helps to reduce the computational complexity and significantly reduce the requirements for hardware resources, and significantly improves the computing speed. In addition, based on the reduced data scale, the present invention uses a lightweight classification model to process the two-dimensional projection image, which can improve the reasoning speed and analyze large quantities of CT image data more quickly, thereby effectively improving the efficiency of analyzing and processing CT sequence images.

[0065] Further, in step 102, the dimensionality reduction processing is performed on the CT sequence images by using the bone compression average density projection method to obtain a two-dimensional projection image, including:

[0066] Based on the CT sequence images, the pixel value of each pixel in each image layer is obtained;

[0067] In each image layer, the pixel value of each pixel is converted into the initial CT value to obtain the CT value of all pixels;

[0068] Through the bone compression average density projection method, bone compression processing is performed on each image layer based on the bone compression algorithm and CT value to obtain each bone compression image layer; projection calculation is performed on each bone compression image layer to obtain a two-dimensional projection image.

[0069] In this embodiment, based on the CT sequence images, the pixel value of each pixel in each image layer is obtained; in each image layer, the pixel value of each pixel is converted into an initial CT value to obtain the CT value of all pixels, specifically:

[0070] Get the DICOM file corresponding to the CT sequence image;

[0071] Extract the metadata of the pixel value of the DICOM file to obtain the pixel value of each pixel in each image layer;

[0072] The pixel value of each pixel is converted to the initial CT value expressed in Hounsfield Units (HU) by the following formula:

[0073] Initial CT value (HU) = RS × SV + RIA;

[0074] In the above formula, SV (stored values) is the pixel value; RS (Rescale Slope) is the adjustment slope; RIA is the Rescale Intercept Attribute.

[0075] The present invention achieves a reduction in data dimension by converting continuous CT tomographic images into two-dimensional projection images. This reduction in data dimension not only helps to reduce computational complexity, but also can significantly reduce the requirements for hardware resources and significantly improve the computing speed. As a result, large quantities of CT image data can be analyzed more quickly, thereby effectively improving the efficiency of analyzing and processing CT sequence images.

[0076] Furthermore, the bone compression algorithm is expressed by the following formula:

[0077]

[0078]

[0079] In the above formula, α is a preset hyperparameter; CL is a preset window position of the CT sequence images, and WW is a preset window width of the CT sequence images, and the preset window position and the preset window width are values ​​determined based on the contrast agent density range.

[0080] In this embodiment, the preset window level and the preset window width are values ​​determined based on the contrast agent density range. For example, if the contrast agent density range is 0-600HU, CL=300HU and WW=600HU are set.

[0081] In this embodiment, α is freely adjustable as a preset hyperparameter and is generally set to 600-1000 HU.

[0082] The present invention uses the contrast agent density range as a reference for the window position (center-level, CL) and window width (window-width, WW) of the CT sequence image, which can maximize the density difference of the contrast agent, thereby improving the discrimination of plain scan and enhanced images. In addition, since high-density objects such as bones, calcifications, and metals have extremely high HU values ​​(generally greater than 1000HU) on CT images, the proportion of the contrast agent will be diluted during average density projection, interfering with the discrimination between plain scan and enhanced images. Therefore, the present invention suppresses the CT value of high-density structures, that is, for pixels with CT values ​​greater than α, they are reassigned to the minimum value β of the CT value range in the image, thereby improving the discrimination of plain scan and enhanced images, thereby improving the discrimination accuracy of CT sequence images. ;

[0083] Furthermore, projection calculation is performed on each bone compression image layer using the following formula:

[0084]

[0085] In the above formula, BSAIP(y,z) is the CT value of the pixel point with coordinates (y,z) in the two-dimensional projection image; x is the sagittal axis coordinate of the CT sequence image, y is the coronal axis coordinate of the CT sequence image, and z is the transverse axis coordinate of the CT sequence image; n is the total number of pixels along the sagittal axis.

[0086] In this embodiment, since the common CT sequence image resolution is 512x512 pixels, n is 512.

[0087] Please see below Figure 2 , Figure 2 This is a spatial coordinate diagram of a CT sequence image provided by an embodiment of the present invention.

[0088] In this embodiment, projection calculation is performed on each image layer after bone compression by using an average density projection method. The average density projection method is completed by calculating the weighted average of the CT values ​​on each projection line. Specifically, That is, in the zth layer of the CT sequence image, at the position where the coronal axis coordinate is y, the CT values ​​of n pixels along the x direction are summed, and finally this embodiment divides it by the total number of pixels along the x axis, that is, the total number of pixels along the sagittal axis n, to calculate the average CT value of the bone compression pixel. Then the calculated average CT value is normalized and converted into a gray value.

[0089] This embodiment is not limited to using the average density projection method to reduce data dimension, and other dimensionality reduction methods such as the maximum density projection method may also be used.

[0090] For the most common cross-sectional tomographic images, the projection direction is perpendicular to the coronal plane. In the present embodiment, the projection direction for the projection calculation of each bone compression image layer is along the direction perpendicular to the coronal plane, but the present invention is not limited to setting the projection direction to be along the direction perpendicular to the coronal plane, and can also be projected in other directions, such as along the direction perpendicular to the cross-sectional direction, along the direction perpendicular to the sagittal plane, etc., and projection along other directions can also achieve similar effects to that of the present embodiment.

[0091] Please see below Figure 3 , Figure 3 A schematic diagram of a plain scan sequence and an enhanced sequence in a CT sequence image provided by an embodiment of the present invention.

[0092] The present invention integrates the bone suppression algorithm and the projection calculation of each bone suppression image layer to obtain a bone suppression average density projection method (Bone-Suppression Average Intensity Projection, BSAIP). The CT sequence images are converted into two-dimensional projection images through the bone suppression average density projection algorithm. Specifically, the present invention can display the average density of tissues or organs by performing projection calculation on each bone suppression image layer, which is helpful for more accurate diagnosis and analysis of lesions, such as Figure 3 As shown in the figure, the image difference between CT plain scan sequence images and CT enhanced sequence images can be clearly displayed by the bone compression average density projection algorithm; if the bone compression algorithm is not used and only the projection calculation is performed on the CT sequence images, some performance may be lost; and the continuous CT tomographic images are converted into two-dimensional average density projection images by the bone compression average density projection method, thereby achieving a reduction in data dimension. This reduction in data dimension not only helps to reduce computational complexity, but also can significantly reduce the requirements for hardware resources and significantly improve the computing speed, thereby allowing large quantities of CT image data to be analyzed more quickly, thereby effectively improving the efficiency of analyzing and processing CT sequence images.

[0093] Furthermore, after the dimensionality reduction process is performed on the CT sequence images by the bone compression average density projection method to obtain a two-dimensional projection image, the method further includes:

[0094] The two-dimensional projection image is resampled to update the two-dimensional projection image.

[0095] Since the pixel intervals in various directions of the CT sequence images may be inconsistent, the present invention can maintain the original anatomical proportions of the projected two-dimensional projection images through resampling.

[0096] Furthermore, the method for obtaining the preset lightweight classification model includes:

[0097] Obtaining initial CT sequence images and annotation labels of the initial CT sequence images;

[0098] Performing dimensionality reduction processing on the initial CT sequence images by using a bone compression average density projection method to obtain a CT projection image;

[0099] The preset basic model is trained and verified based on the CT projection images and the annotated labels to obtain a preset lightweight classification model.

[0100] In this embodiment, the annotation labels of the initial CT sequence images include plain scan labels and increase and decrease labels; the annotation labels of the initial CT sequence images are obtained by manually annotating plain scan and enhanced images.

[0101] In this embodiment, the CT projection image is read in grayscale format and scaled to a resolution of 224×224.

[0102] In this embodiment, since the compression bone average density projection method can greatly compress the dimension of the CT sequence image, for example, for a CT image sequence with a resolution of 512x512 pixels, the data scale can be compressed to 1 / 512 of the original, so that the lightweight network can achieve excellent performance in this scenario. Therefore, the present invention selects the MobileNetV3-small network as the framework of the preset basic model. The MobileNetV3-small network is a lightweight convolutional neural network with fewer parameters (less than 2 million) and lower computational complexity. In order to output the binary classification result, the present invention connects an MLP module after the last layer of the MobileNetV3-small network to reduce the image features to 2 dimensions to achieve flat scanning and enhanced binary classification. Therefore, the main structure of the preset basic model consists of a MobileNetV3-small module and a multilayer perceptron (MLP, Multilayer Perceptron) module.

[0103] See also Figure 4 , Figure 4 A schematic diagram of the structure of a preset basic model provided in an embodiment of the present invention.

[0104] like Figure 4 As shown, the MobileNetV3-small module includes the following main components:

[0105] ① Input layer: accepts image data.

[0106] ②Convolutional layer Conv2d: includes multiple convolution and h-swish activation function layers.

[0107] ③bneck unit: MobileNetV3 uses multiple bneck units, each of which includes a series of depth-separable convolution and h-swish activation function operations, as well as a Squeeze-and-Excitation attention module and an inverted residual block with linear bottleneck.

[0108] ④Global Average Pooling: used to reduce the size of the feature map.

[0109] ⑤Fully Connected Layer: Outputs 2048-dimensional image features.

[0110] The MLP module includes the following main components:

[0111] ①Fully connected layer: accepts the image features output by the last layer of the MobileNetV3-small module.

[0112] ②BatchNorm layer: implements batch normalization, accelerates model training, and improves generalization capabilities.

[0113] ③ReLU activation function: increases the nonlinear ability of the model and improves model performance.

[0114] ④Fully Connected Layer: Outputs 2D classification results.

[0115] In this embodiment, the preset basic model is trained and verified based on the CT projection image and the annotated label to obtain a preset lightweight classification model, which is specifically:

[0116] Use the pre-trained parameters of the MobileNetV3-small network in the ImageNet dataset to initialize the preset basic model;

[0117] Among them, the specific training parameters of the preset basic model are as follows:

[0118] ①Batch size: 64;

[0119] ②Optimizer: Stochastic Gradient Descent (SGD);

[0120] ③Learning rate: 0.008 for MobileNetV3-small module and 0.024 for MLP module;

[0121] ④ Momentum: 0.9;

[0122] ⑤Weight decay: 5e-4;

[0123] ⑥Learning rate scheduler: Cosine Annealing LR;

[0124] ⑦Loss function: cross entropy;

[0125] ⑧Maximum number of iterations: 100;

[0126] Train the preset basic model based on specific training parameters, specifically:

[0127] S1. Input the CT projection image into the preset basic model, and obtain the preset basic model's prediction label for the plain scan or enhancement of the CT projection image through the forward propagation of the preset basic model;

[0128] S2. Calculate the cross entropy between the predicted label and the annotated label (the actual label of the sample), and calculate the parameter gradient in the neural network through back propagation; the following formula is the cross entropy loss function:

[0129]

[0130] Where pi is the i-th output result of the classifier, and qi is the annotation label corresponding to the i-th output;

[0131] S3, minimizing the cross entropy loss function through the stochastic gradient descent algorithm to update the preset basic model parameters, and repeating the process of S1 to S3 until the preset basic model converges and reaches the maximum number of iterations;

[0132] S4. During the training process, the model with the best performance in the validation set is selected as the preset lightweight classification model.

[0133] Furthermore, after obtaining the CT projection image, the method further includes:

[0134] The CT projection image is preprocessed to update the CT projection image.

[0135] The present invention takes into account the characteristics of CT projection images and uses random vertical flipping to perform data enhancement on the images to simulate positioning situations in different directions.

[0136] The present invention can increase the diversity of data through preprocessing. The model requires a large amount of data to train, but the original data may have problems such as sample imbalance, noise data and missing data. This solution can remove noise data, fill missing data, and balance the data by preprocessing the CT projection images, making the data more diverse and helping to train a better model.

[0137] Furthermore, in step 103, the two-dimensional projection image is processed based on the preset lightweight classification model to obtain the discrimination result of the CT sequence image, including:

[0138] Inputting the two-dimensional projection image into a preset lightweight classification model; the lightweight classification model includes a MobileNetV3-small module and an MLP module;

[0139] Extracting image features of the two-dimensional projection image through a MobileNetV3-small module;

[0140] The image features are input into the MLP module to obtain the discrimination results of the CT sequence images.

[0141] Since the present invention reduces the data scale of CT sequence images, the present invention can use a lightweight convolutional neural network to analyze and process the two-dimensional projection images, thereby improving the reasoning speed and being able to analyze large quantities of CT image data more quickly; and, the present invention inputs the fused and reduced-dimensional two-dimensional projection images into the network through the bone compression average density projection method, avoiding passing all tomographic images to the lightweight classification model, and filtering out image changes that are not related to specific tasks to a certain extent, avoiding the incoherence of the output results and the introduction of irrelevant information, thereby improving the accuracy of the analysis.

[0142] In addition to performing plain scanning and enhanced identification of CT sequence images, the present invention can also analyze and process other tomographic images such as MRI.

[0143] In order to overcome the shortcomings of the prior art in identifying CT plain scan and enhanced sequence images, in particular, the prior art does not process the patient's CT sequence data, but independently inputs each tomographic image into a neural network for prediction. Each patient will form multiple prediction results, which are finally aggregated to form a final label, resulting in large data scale, slow running speed and other problems. The present invention reduces the dimension of continuous CT sequence images into two-dimensional projection images by improving the average density projection method. While maximizing the retention of the enhancement characteristics of the contrast agent, by changing the input of the classification model from the original CT sequence to the two-dimensional projection image, the data scale can be greatly reduced, the GPU resource demand can be reduced, and a lightweight CNN network is used to maximize the low-dimensional advantage of two-dimensional images, which significantly improves the efficiency of image analysis, so as to meet the needs of large-scale data processing and provide a more efficient and accurate solution for CT sequence image data cleaning.

[0144] Accordingly, in order to solve the above technical problems, the present invention also provides a CT sequence image processing device, please refer to Figure 5 , which is a structural schematic diagram of an embodiment of a CT sequence image processing device provided by the present invention.

[0145] like Figure 5 As shown, the CT sequence image processing device 20 includes: an image acquisition module 201, a dimension reduction processing module 202 and an image processing module 203;

[0146] Wherein, the image acquisition module 201 is used to acquire the CT sequence images to be processed;

[0147] The dimension reduction processing module 202 is used to perform dimension reduction processing on the CT sequence images by using a bone compression average density projection method to obtain a two-dimensional projection image;

[0148] The image processing module 203 is used to process the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of a CT sequence image.

[0149] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for processing CT sequence images as described in any one of the above items are implemented.

[0150] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0151] 1. The present invention reduces the dimension of the data by converting continuous CT sequence images into 2D average density projection images. The reduction of data dimension not only helps to reduce the computational complexity, but also significantly reduces the requirements for hardware resources and significantly improves the calculation speed. In addition, the whole process of the bone compression average density projection algorithm proposed in the present invention does not need to rely on GPU computing resources, but makes full use of the parallel processing capabilities of the computer central processing unit (CPU), which is particularly suitable for environments that do not have high-end GPU resources, and can realize large-scale parallel computing and improve operating efficiency.

[0152] 2. Since the present invention reduces the size of data to 1 / 512 of the original, the present invention can adopt a lightweight convolutional neural network; and the use of a lightweight convolutional neural network improves the reasoning speed of the present invention, and can analyze large amounts of CT image data more quickly.

[0153] 3. The present invention uses the bone compression average density projection method to input the fused dimensionality-reduced two-dimensional image into the network, avoiding passing all images to the CNN network. It can filter out image changes that are not related to specific tasks to a certain extent, achieve one-to-one correspondence between patients, images, and prediction labels, avoid the incoherence of output results and the introduction of irrelevant information, thereby improving the reasoning speed and the accuracy of analysis.

[0154] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for processing CT sequence images, characterized in that: include: Acquire CT sequence images to be processed; Performing dimension reduction processing on the CT sequence images by using a bone compression average density projection method to obtain a two-dimensional projection image; The two-dimensional projection image is processed based on a preset lightweight classification model to obtain a discrimination result of the CT sequence image.

2. A method for processing CT sequence images according to claim 1, characterized in that: The dimensionality reduction process of the CT sequence images is performed by the bone compression average density projection method to obtain a two-dimensional projection image, including: Based on the CT sequence images, the pixel value of each pixel in each image layer is obtained; In each image layer, the pixel value of each pixel is converted into the initial CT value to obtain the CT value of all pixels; Through the bone compression average density projection method, bone compression processing is performed on each image layer based on the bone compression algorithm and CT value to obtain each bone compression image layer; Projection calculation is performed on each bone compression image layer to obtain a two-dimensional projection image.

3. A method for processing CT sequence images as claimed in claim 2, characterized in that: The bone compression algorithm is expressed by the following formula: In the above formula, α is a preset hyperparameter; CL is a preset window position of the CT sequence images, and WW is a preset window width of the CT sequence images, and the preset window position and the preset window width are values ​​determined based on the contrast agent density range.

4. A method for processing CT sequence images as claimed in claim 3, characterized in that: The projection calculation of each bone compression image layer is performed using the following formula: In the above formula, BSAIP(y,z) is the CT value of the pixel point with coordinates (y,z) in the two-dimensional projection image; x is the sagittal axis coordinate of the CT sequence image, y is the coronal axis coordinate of the CT sequence image, and z is the transverse axis coordinate of the CT sequence image; n is the total number of pixels along the sagittal axis.

5. The method for processing CT sequence images according to claim 1, characterized in that: After the dimensionality reduction process is performed on the CT sequence images by the bone compression average density projection method to obtain a two-dimensional projection image, the method further includes: The two-dimensional projection image is resampled to update the two-dimensional projection image.

6. A method for processing CT sequence images according to claim 1, characterized in that: The method for obtaining the preset lightweight classification model includes: Obtaining initial CT sequence images and annotation labels of the initial CT sequence images; Performing dimensionality reduction processing on the initial CT sequence images by using a bone compression average density projection method to obtain a CT projection image; The preset basic model is trained and verified based on the CT projection images and the annotated labels to obtain a preset lightweight classification model.

7. A method for processing CT sequence images according to claim 6, characterized in that: After obtaining the CT projection image, the method further includes: The CT projection image is preprocessed to update the CT projection image.

8. The method for processing CT sequence images according to claim 1, characterized in that: The two-dimensional projection image is processed based on a preset lightweight classification model to obtain a discrimination result of a CT sequence image, including: Inputting the two-dimensional projection image into a preset lightweight classification model; the lightweight classification model includes a MobileNetV3-small module and an MLP module; Extracting image features of the two-dimensional projection image through a MobileNetV3-small module; The image features are input into the MLP module to obtain the discrimination results of the CT sequence images.

9. A CT sequence image processing device, characterized in that: include: Image acquisition module, dimension reduction processing module and image processing module; Wherein, the image acquisition module is used to acquire the CT sequence images to be processed; The dimension reduction processing module is used to perform dimension reduction processing on the CT sequence images by using a bone compression average density projection method to obtain a two-dimensional projection image; The image processing module is used to process the two-dimensional projection image based on a preset lightweight classification model to obtain a discrimination result of a CT sequence image.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for processing CT sequence images according to any one of claims 1 to 8 are implemented.

Images