Lung lobe and liver segment segmentation method based on artificial intelligence network

Abstract

The invention is applied to the technical field of clinical image recognition, and discloses a lung lobe and liver segment segmentation method based on an artificial intelligence network, and the method comprises the following steps: S1, training a UNet-A to segment lung fissure or sickle ligament in the liver, S2, calculating a distance field of prediction result lung fissure or ligament, S3, training a second UNet-B, taking the distance field and mask as input, and carrying out the segmentation of the lung lobe and liver segment. And obtaining a segmentation result of the lung lobe or the liver segment. According to the lung lobe and liver segment segmentation method based on the artificial intelligence network, a mode of preprocessing and inputting an image is adopted, the size of the input image is reduced to adapt to video memory limitation during UNet training, meanwhile, a mode of carrying out resolution reduction processing on the separated left lung or right lung is adopted, so that the whole image of the left lung or the right lung can be used as network input, and the image segmentation efficiency is improved. A large image does not need to be divided into small patches, and it is guaranteed that the UNet network not only can learn gray value feature information in the image, but also can learn geometric shape and position information of a target object.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical image recognition, and in particular to a method for segmenting lung lobes and liver segments based on an artificial intelligence network. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, the medical field has also been greatly improved and enhanced. Medical image segmentation and recognition, as a key link in medical image processing, provide doctors with more accurate and reliable medical diagnosis methods. Medical image segmentation is the process of separating the region of interest in a medical image from the background. This step is crucial for extracting the lesion area, reconstructing the anatomical structure, and diagnosing the disease. By segmenting the medical image, the lesion can be accurately located, assisting doctors in making judgments on the condition, and improving the accuracy and efficiency of diagnosis. Common medical image segmentation methods In the study of medical image segmentation, there are many common methods that are widely used. Among them, the threshold-based segmentation method is the simplest and most commonly used method. This method divides pixels in the image with grayscale values ​​higher or lower than the threshold into different regions by setting a value. However, due to the noise and complexity of medical images, a single threshold value often cannot meet the accuracy requirements. In order to meet the requirements of accurate segmentation and overcome the limitations of threshold-based segmentation methods, many new segmentation algorithms have been proposed, such as edge-based segmentation methods and region-based segmentation methods. Edge-based segmentation methods use edge detection operators to detect and extract edges in medical images to achieve the purpose of segmentation. Region-based segmentation methods divide images into multiple different regions and segment them according to the statistical characteristics and texture information of the regions. The segmentation of medical images is only the first step in medical image processing. In order to better understand and use this information, the recognition and application of medical images cannot be ignored. In clinical diagnosis and treatment, image recognition and segmentation play an important role. Based on the results of segmentation, doctors can find and diagnose lesions faster and understand pathological anatomical structures. Automatic segmentation of liver and lung segments based on CT has always been a challenging topic. In order to better segment and identify lung and liver segments, the upper and lower lobes and left and right liver segments are usually segmented first.

[0003] The current common method is to train an artificial intelligence network to directly segment the lung lobes or the left and right livers. The disadvantage of this is that it is difficult to avoid false positives in the results, and the large area of ​​false positive pixels will greatly increase the difficulty of post-processing. Summary of the invention

[0004] The purpose of the present invention is to provide a lung lobe and liver segment segmentation method based on an artificial intelligence network to solve the problem of poor segmentation methods proposed in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network, the segmentation method comprising the following steps:

[0006] S1. Train Unet-A to segment the pulmonary fissure or the falciform ligament in the liver;

[0007] S2. Calculate the distance field of the predicted pulmonary fissure or ligament;

[0008] S3. Train the second UNet-B to take the distance field and mask as input to obtain the segmentation results of the lung lobe or liver segment.

[0009] Preferably, the training method in S1 includes preliminary detection of anatomical features of the input image, wherein the anatomical features of the input image are specifically manifested as pulmonary fissures in lung images, and the anatomical features of the input image are specifically manifested as falciform ligaments in liver images.

[0010] Preferably, 3D Unet is used as the segmentation network in S1, wherein the network depth parameter of the 3D Unet segmentation network is 2 layers, the filter number parameter of the 3D Unet segmentation network is 8, the Loss function parameter of the 3D Unet segmentation network is DICE, and the input image dimension parameter of the 3D Unet segmentation network is 3 dimensions.

[0011] Preferably, the input image in S1 is a preprocessed lung or liver CT image, and the preprocessing process includes image segmentation and adjustment and enhancement of the image grayscale value, wherein the image segmentation result is used as ROI and the image information of other tissues and organs is filtered out, and the image grayscale value is enhanced using a histogram equalization algorithm.

[0012] Preferably, after the lung CT preprocessing in S1 is completed, the left and right lung images are separated.

[0013] Preferably, 3D Unet is used as the segmentation network in S3, wherein the network depth parameter of the 3D Unet segmentation network is 2 layers, the filter number parameter of the 3D Unet segmentation network is 8, the Loss function parameter of the 3D Unet segmentation network is DICE, and the input image dimension parameter of the 3D Unet segmentation network is 3 dimensions.

[0014] Preferably, the 3D Unet segmentation network in S1 and S3 consists of an encoder and a decoder, and the encoder and decoder of the 3D Unet segmentation network are connected via a jump connection, and the encoder and decoder consist of multiple convolutional layers, ReLU activation functions and pooling layers.

[0015] Preferably, the 3D Unet segmentation network training in S1 and S3 includes preparing 3D segmentation data and defining a loss function and an optimizer.

[0016] Preferably, the optimizer for the 3D Unet segmentation network training in S1 and S3 adopts Dice Loss and Adam optimizer, and the 3D segmentation data preparation for the 3D Unet segmentation network training in S1 and S3 includes data collection, preprocessing, data enhancement and definition of data loader.

[0017] Preferably, the 3D Unet segmentation network input in S3 is a dual-channel structure, wherein the 3D Unet segmentation network input channel 1 in S3 is the calculation result volume_X of S2, and the 3D Unet segmentation network input channel 2 in S3 is the preprocessed image result volume_Y in S1. After the input is completed, the images of the two groups of channels are merged through the code and then input to train the 3DUnet segmentation network, wherein the code is Volume_multi_channel=np.concatenate([volume_X, volume_Y], 4).

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: the lung lobe and liver segment segmentation method based on artificial intelligence network:

[0019] 1. The image is pre-processed to reduce the size of the input image to adapt to the memory limitation when training UNet. At the same time, the separated left lung or right lung is processed by reducing the resolution so that the entire image of the left lung or right lung can be used as the network input without dividing the large image into small patches. This ensures that the UNet network can not only learn the gray value feature information in the image, but also learn the geometric shape and position information of the target object;

[0020] 2. Furthermore, by segmenting the image in the preprocessing process and calculating the distance field of the segmentation results, the information of the lung fissure or ligament can be distributed throughout the image, which is beneficial for UNet to perform classification judgment;

[0021] 3. Furthermore, by preparing 3D segmentation data for 3D Unet segmentation network training, the model can be effectively learned and generalized. At the same time, by setting up training and test groups through large-scale data collection and randomly extracting images from the training group, the network parameters can be optimized for learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a schematic diagram of the network separation process of the present invention;

[0023] Figure 2 is the CT input image preprocessed by the present invention;

[0024] Figure 3 This is the gold standard image of lung lobe separation of the present invention;

[0025] Figure 4 This is a graph of prediction results of the artificial intelligence network of the present invention;

[0026] Figure 5 Schematic diagram of calculating pixel grayscale value of distance field according to the present invention. DETAILED DESCRIPTION

[0027] 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.

[0028] See also Figure 1-5 The present invention provides a technical solution: a method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network, the segmentation method comprising the following steps:

[0029] Step 1. Train Unet-A to segment the falciform ligament in the lung fissure or liver;

[0030] The training method includes preliminary detection of anatomical features of the input image, the anatomical features of the input image are specifically manifested as pulmonary fissures in the lung image, and the anatomical features of the input image are specifically manifested as falciform ligaments in the liver image;

[0031] The input image is a preprocessed lung or liver CT image. The preprocessing process includes image segmentation and adjustment and enhancement of the image gray value. The image segmentation result is used as the ROI and the image information of other tissues and organs is filtered out. The histogram equalization algorithm is used to enhance the image gray value.

[0032] Image adjustment includes image simplification, removing small noise interference and details that are not important to perception, and smoothing the image. A reconstruction filter can be selected for simplification. The simplification of medical images is different from that of ordinary images, because in addition to considering the simplification effect, medical images also need to consider the retention of small areas of interest. Generally, the larger the structure element, the better the filtering effect. However, if the structure element is too large, the small area of ​​interest will be filtered out. When simplifying a standard image, the structure element is generally about 8x8. In order to meet the requirements of medical image segmentation, a 3x3 structure element is selected;

[0033] After lung CT preprocessing is completed, the left and right lung images are separated

[0034] 3D Unet is used as the segmentation network, where the network depth parameter of the 3D Unet segmentation network is 2 layers, the number of filters of the 3D Unet segmentation network is 8, the Loss function parameter of the 3D Unet segmentation network is DICE, and the input image dimension parameter of the 3D Unet segmentation network is 3 dimensions;

[0035] Step 2. Calculate the distance field of the predicted lung fissure or ligament;

[0036] Step 3. Train the second UNet-B to take the distance field and mask as input to obtain the segmentation results of the lung lobe or liver segment;

[0037] 3D Unet is used as the segmentation network, where the network depth parameter of the 3D Unet segmentation network is 2 layers, the number of filters of the 3D Unet segmentation network is 8, the Loss function parameter of the 3D Unet segmentation network is DICE, and the input image dimension parameter of the 3D Unet segmentation network is 3 dimensions;

[0038] The 3D Unet segmentation network consists of an encoder and a decoder, and the encoder and decoder of the 3D Unet segmentation network are connected through a skip connection. The encoder and decoder consist of multiple convolutional layers, ReLU activation functions, and pooling layers;

[0039] The skip connection improves the semantic similarity of feature maps of different scales to reduce the difficulty of the model learning multi-scale feature information. The characteristic of the skip connection is that it captures information that is easily lost in the encoding process by gradually superimposing features, thereby optimizing the multi-scale feature fusion mechanism and achieving increased segmentation accuracy on the dataset;

[0040] The convolution layer is used to automatically extract the features of the input image. The convolution layer consists of a series of learnable filters that slide on the image and calculate the dot product of the local area to generate feature maps. These feature maps are then sent to the next layer of the network for further processing. The convolution layer performs convolution operations on the input image through a set of trainable convolution kernels to obtain a set of feature maps. Each convolution kernel slides on the image, multiplies the pixel values ​​of the area it covers by the weight of the convolution kernel and sums them up, and finally obtains a scalar. This scalar can be regarded as the value of the corresponding pixel on the feature map, reflecting the response of the convolution kernel at the current position. The function of the convolution layer is mainly to extract the features of the image;

[0041] 3D Unet segmentation network training includes preparing 3D segmentation data and defining the loss function and optimizer;

[0042] The optimizer for 3D Unet segmentation network training uses Dice Loss and Adam optimizer. The 3D segmentation data preparation for 3D Unet segmentation network training includes data collection, preprocessing, data enhancement, and data loader definition;

[0043] Dice Loss is a commonly used loss function for evaluating model performance in segmentation tasks. It is particularly suitable for binary or multi-classification tasks such as medical image segmentation. The core idea is to evaluate model performance by measuring the degree of overlap between the predicted results and the true labels. The Dice coefficient measures the similarity between two sets, ranging from 0 to 1. The larger the value, the more overlap. The Dice coefficient is defined as follows:

[0044] Dice(,)=2∣∩∣|∣+∣∣Dice(A,B)=∣A∣+∣B∣2∣A∩B∣

[0045] in:

[0046] A is the predicted segmentation result;

[0047] B is the true segmentation label;

[0048] ∣∩∣∣A∩B∣ represents the intersection of the predicted and true labels, i.e., the number of overlapping voxels;

[0049] |||A| and |||B| represent the number of voxels of the predicted and true labels, respectively;

[0050] The Adam (Adaptive Moment Estimation) optimizer is a widely used optimization algorithm that is particularly suitable for processing large-scale data and high-dimensional parameter spaces. It combines the advantages of AdaGrad and RMSProp, and adjusts the learning rate by calculating the first-order momentum (mean) and second-order momentum (uncentered variance) of the gradient, thereby achieving faster convergence and better robustness.

[0051] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for segmenting lung lobes and liver segments based on an artificial intelligence network, characterized in that: The segmentation method comprises the following steps: S1. Train Unet-A to segment the pulmonary fissure or the falciform ligament in the liver; S2. Calculate the distance field of the predicted pulmonary fissure or ligament; S3. Train the second UNet-B to take the distance field and mask as input to obtain the segmentation results of the lung lobe or liver segment.

2. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 1, characterized in that: The training method in S1 includes preliminary detection of the anatomical features of the input image, wherein the anatomical features of the input image are specifically manifested as pulmonary fissures in the lung image, and the anatomical features of the input image are specifically manifested as falciform ligaments in the liver image.

3. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 2, characterized in that: In S1, 3D Unet is used as the segmentation network, wherein the network depth parameter of the 3D Unet segmentation network is 2 layers, the number of filters of the 3D Unet segmentation network is 8, the Loss function parameter of the 3D Unet segmentation network is DICE, and the input image dimension parameter of the 3D Unet segmentation network is 3 dimensions.

4. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 2, characterized in that: The input image in S1 is a preprocessed lung or liver CT image, and the preprocessing process includes image segmentation and adjustment and enhancement of the image grayscale value, wherein the image segmentation result is used as ROI and the image information of other tissues and organs is filtered out, and the image grayscale value is enhanced using a histogram equalization algorithm.

5. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 4, characterized in that: After the lung CT preprocessing in S1 is completed, the left and right lung images are separated.

6. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 1, characterized in that: In S3, 3D Unet is used as the segmentation network, wherein the network depth parameter of the 3D Unet segmentation network is 2 layers, the number of filters of the 3D Unet segmentation network is 8, the Loss function parameter of the 3D Unet segmentation network is DICE, and the input image dimension parameter of the 3D Unet segmentation network is 3 dimensions.

7. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 1, characterized in that: The 3D Unet segmentation network in S1 and S3 is composed of an encoder and a decoder, and the encoder and the decoder of the 3D Unet segmentation network are connected through a jump connection, and the encoder and the decoder are composed of multiple convolutional layers, ReLU activation functions and pooling layers.

8. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 1, characterized in that: The 3D Unet segmentation network training in S1 and S3 includes preparing 3D segmentation data and defining a loss function and an optimizer.

9. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 8, characterized in that: The optimizer for the 3D Unet segmentation network training in S1 and S3 adopts Dice Loss and Adam optimizer, and the 3D segmentation data preparation for the 3DUnet segmentation network training in S1 and S3 includes data collection, preprocessing, data enhancement and definition of data loader.

10. The method for segmenting pulmonary lobes and liver segments based on an artificial intelligence network according to claim 1, characterized in that: The 3D Unet segmentation network input in S3 is a dual-channel structure, wherein the 3D Unet segmentation network input channel one in S3 is the calculation result volume_X of S2, and the 3D Unet segmentation network input channel two in S3 is the preprocessed image result volume_Y in S1. After the input is completed, the images of the two groups of channels are merged through the code and then input to train the 3D Unet segmentation network, wherein the code is Volume_multi_channel=np.concatenate([volume_X, volume_Y], 4).

Images