An image segmentation method for intervertebral disc herniation

Abstract

The present invention provides an image segmentation method for intervertebral disc herniation, comprising a training phase and an inference phase; the training phase comprises: (1) data set acquisition; (2) data preprocessing; (3) model building; (4) model training and evaluation; the inference phase comprises: 1) data preprocessing; 2) model building; 3) model inference. The technical solution of the present invention can assist doctors in quickly determining the protruding part of the intervertebral disc by training a deep learning model. Compared with traditional segmentation methods, the model of the present invention can automatically extract image features, saving a lot of manpower and time; the model designed by the present invention increases the ability to extract context by using pooling operations of different scales, thereby improving segmentation accuracy and segmentation effect.

Description

Technical Field

[0001] The invention provides an image segmentation method for intervertebral disc herniation, belonging to the technical field of image processing. Background Art

[0002] The spine is a vital structure responsible for supporting the body, protecting internal organs and regulating balance. A herniated disc is a common spinal condition. When a protrusion compresses nerves, it can cause pain and, in severe cases, loss of mobility. Disc herniation can be categorized by the site of onset: cervical, thoracic, and lumbar.

[0003] The existing image segmentation method for intervertebral disc herniation is based on a threshold-based segmentation method, which is a very common segmentation method and is also used in medical image processing. Due to the significant difference in density between bones and intervertebral discs, bones appear white and intervertebral discs appear black on CT images. During image processing, by selecting an appropriate threshold, the bones and intervertebral discs can be segmented to obtain an image of the intervertebral disc.

[0004] Since the symptom of intervertebral disc herniation is the protrusion of the nucleus pulposus into the spinal canal after the rupture of the annulus fibrosus of the intervertebral disc, the threshold-based segmentation method can only segment the intervertebral disc but not the protruding part, it is not suitable for the segmentation task of intervertebral disc herniation.

[0005] 202010922602.8 provides a lumbar spine disease diagnosis system based on neural network. The neural network model used in this method is an improved UNet model, which is divided into two parts: encoding and decoding. The encoding part is the left half, and the decoding part is the right half. The input image is in the upper left corner. After two layers of convolution, it is pooled to half the size. Similar operations are repeated 4 times, and the image size becomes smaller and smaller. Then comes the decoding part, which uses convolution and upsampling layers to restore the original image size. In addition, a jump connection method is used in the middle of the corresponding layers to fuse shallow and deep information. During the decoding process, this method connects the output end of the third decoding unit to the input end of the first decoding unit, so that the feature map output by the third encoding unit is used as the input of the first decoding unit to improve the segmentation effect. However, this method cannot capture the contextual information in the image.

[0006] In summary, existing solutions for disc herniation struggle with threshold-based segmentation methods, while UNet models fail to capture contextual information within the image. Because the herniated portion of the disc remains part of the disc, it appears connected to the normal portion in the image. Therefore, using only the UNet architecture cannot accurately distinguish the herniated portion. Furthermore, while UNet has demonstrated excellent segmentation capabilities, there is still room for improvement, and accuracy needs to be further enhanced. Summary of the Invention

[0007] In response to the above technical problems, the present invention provides an image segmentation method for intervertebral disc herniation, which includes a training phase and an inference phase;

[0008] The training phase includes:

[0009] (1) Dataset Collection

[0010] There are two types of data that need to be obtained. One is the input image, the file format is DCM format, and the image content is a spinal CT image of intervertebral disc herniation; the other is the mask image, the file format is PNG format, and the image content is that the pixel value except the protruding part is 1, and the rest of the pixel values ​​are 0.

[0011] (2) Data preprocessing

[0012] Perform augmentation and expansion operations on each set of data in the dataset, including random rotation of the image and pixel intensity normalization.

[0013] Random rotation: A method of data augmentation that rotates the input image at random angles. The mask image and array are also rotated accordingly, keeping their relative positions unchanged. The rotated data set is then added to the dataset, doubling the amount of data in the dataset.

[0014] Pixel intensity normalization: Each pixel of the input image in a set of data is transformed as shown in formula (1), where x i is the pixel value to be processed, x min is the minimum value of all pixels of the input image, x max is the maximum value of all pixels in the input image. Pixel intensity normalization limits the data to a specified range, eliminating the negative effects of singular data.

[0015]

[0016] (3) Model construction

[0017] The model's infrastructure uses the PSPNet architecture. The input image first passes through a feature extraction network, using the VGG16 model as the feature extraction network. After feature extraction, the output feature map is divided into four regions of different sizes (1x1, 2x2, 3x3, and 6x6). Average pooling is performed within each region, followed by a convolution layer to adjust the number of channels to 1 / 4 of the original. Each region is then upsampled and concatenated in the channel dimension. Furthermore, the extracted feature map is also concatenated with the upsampled feature map as part of the concatenation. Finally, a convolution layer is used to adjust the number of channels and output the mask image.

[0018] (4) Model training and evaluation

[0019] After the model is built, the model parameters are updated through continuous iterative optimization. There are two loss functions used to evaluate the model, one corresponding to the two sets of outputs of the model. For the output mask, the binary cross entropy loss function is used as the evaluation function; for the output array, the mean square error loss function is used as the evaluation function. During the training process, the Adam optimizer is used as the optimization algorithm to iteratively update the model parameters, and the learning rate is set to 1×10 -4 .

[0020] An early stopping strategy is used to terminate model training early. The specific operations are as follows: the number of model iterations is set to 500, and the rule "if the loss value of the output mask does not decrease within 20 iterations, then terminate the training" is established.

[0021] The inference phase includes:

[0022] 1) Data preprocessing

[0023] The pixel intensity of the image to be segmented is normalized and fed into the model as the input image.

[0024] 2) Model building

[0025] Build the same model as in the training phase and import the trained model parameters.

[0026] 3) Model Inference

[0027] The input image is used as the model input, and after a series of calculations with the trained model parameters, the output mask image is obtained for the doctor's reference.

[0028] The technical solution of this invention, through training a deep learning model, can help doctors quickly identify the herniated disc. Compared with traditional segmentation methods, the model proposed in this invention can automatically extract image features, saving considerable manpower and time. The model designed in this invention enhances the ability to extract context by using pooling operations at different scales, improving segmentation accuracy and results. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a schematic cross-sectional view of a herniated intervertebral disc according to the present invention;

[0030] Figure 2 Input and output schematic diagram of the present invention;

[0031] Figure 3 is a CT image of a herniated intervertebral disc after superimposing a mask according to the present invention;

[0032] Figure 4 The technical solution of the present invention;

[0033] Figure 5 Model designed by the present invention. DETAILED DESCRIPTION

[0034] The specific technical solutions of the present invention are described with reference to the embodiments.

[0035] The present invention is mainly aimed at the image segmentation method of intervertebral disc herniation in the medical field. The surgical treatment of intervertebral disc herniation is to remove the protruding tissue and nucleus pulposus, such as Figure 1 As shown, the black circle marks the protruding part (the part to be removed). The purpose of the present invention is to build a model that can automatically segment the protruding part on the image (that is, classify the protruding part at the pixel level).

[0036] The input and output of image segmentation are as follows Figure 2 As shown, the input is a CT scan image of the intervertebral disc (left), and the output is a mask image (right). The pixel value of the normal tissue in the output image is 0, and the pixel value of the protruding part is 1 ( Figure 2 The right picture is the visualization result, the pixel value is not 1).

[0037] After visualizing the output of the present invention and superimposing it on the input image, we can obtain Figure 3 The effect makes it easier for doctors to view detailed information about herniated disc.

[0038] The technical solution of the present invention is as follows Figure 4 As shown, it consists of two stages: training stage and inference stage, which are described in detail below:

[0039] Training phase:

[0040] (1) Dataset Collection

[0041] This stage requires a lot of time and manpower. There are two types of data that need to be obtained. One is the input image, the file format is DCM format, and the image content is a spinal CT image of a herniated disc; the other is the mask image, the file format is PNG format, and the image content is that the pixel value except the protruding part is 1, and the rest of the pixel values ​​are 0.

[0042] Deep learning models require a large amount of data for training. A set of data includes the three data mentioned above. To ensure the segmentation effect, it is best to collect at least 2,000 sets of data in this stage.

[0043] (2) Data preprocessing

[0044] This stage completes the expansion and enhancement operations on each set of data in the dataset, including random rotation of the image and pixel intensity normalization.

[0045] Random rotation: A method of data augmentation that rotates the input image at random angles. The mask image and array are also rotated accordingly, keeping their relative positions unchanged. The rotated data set is then added to the dataset, doubling the amount of data in the dataset.

[0046] Pixel intensity normalization: Each pixel of the input image in a set of data is transformed as shown in formula (1), where x i is the pixel value to be processed, x min is the minimum value of all pixels of the input image, x max is the maximum value of all pixels in the input image. Pixel intensity normalization limits the data to a specified range, eliminating the negative effects of singular data.

[0047]

[0048] (3) Model construction

[0049] The schematic diagram of the segmentation model structure designed by the present invention is as follows Figure 5 As shown, the model's infrastructure uses the PSPNet architecture. The input image first passes through the feature extraction network to extract features. The VGG16 model is selected as the feature extraction network for this solution. After completing the feature extraction task, the output feature map is divided into four regions of different sizes (1x1, 2x2, 3x3, and 6x6). Average pooling is performed within each region. A convolution layer is then used to adjust the number of channels to 1 / 4 of the original number of channels. After upsampling, each region is spliced ​​in the channel dimension. In addition, the feature map after feature extraction is also used as part of the splicing operation and is spliced ​​with the upsampled feature map. Finally, a convolution layer is used to adjust the number of channels in the output and output the mask image.

[0050] (4) Model training and evaluation

[0051] After the model is built, it is necessary to update the model parameters through continuous iterative optimization. There are two loss functions used to evaluate the model, one corresponding to the two sets of outputs of the model. For the output mask, the binary cross entropy loss function is used as the evaluation function; for the output array, the mean square error loss function is used as the evaluation function. During the training process, the Adam optimizer is used as the optimization algorithm to iteratively update the model parameters, and the learning rate is set to 1×10 -4 .

[0052] To reduce the occurrence of overfitting, an early stopping strategy is used to terminate model training early. The specific operations are as follows: the number of model iterations is set to 500, and the rule "If the loss value of the output mask does not decrease within 20 iterations, terminate the training" is established.

[0053] Reasoning stage:

[0054] 1) Data preprocessing

[0055] The pixel intensity of the image to be segmented is normalized and fed into the model as the input image.

[0056] 2) Model building

[0057] Build the same model as in the training phase and import the trained model parameters.

[0058] 3) Model Inference

[0059] The input image is used as the model input, and after a series of calculations with the trained model parameters, the output mask image is obtained for the doctor's reference.

Claims

1. An image segmentation method for intervertebral disc herniation, characterized in that: Includes training phase and inference phase; The training phase includes: (1) Dataset Collection (2) Data preprocessing Perform augmentation and expansion operations on each set of data in the dataset, including random rotation of the image and pixel intensity normalization; (3) Model construction The model's infrastructure uses the PSPNet architecture. The input image first passes through a feature extraction network to extract features, using the VGG16 model as the feature extraction network. After feature extraction, the output feature map is divided into four regions of different sizes, each of which is then subjected to average pooling. A convolution layer is then used to adjust the number of channels to 1 / 4 of the original number of channels. Each region is then upsampled and concatenated in the channel dimension. Furthermore, the extracted feature map is also concatenated with the upsampled feature map as part of the concatenation. Finally, a convolution layer is used to adjust the number of channels in the output and output the mask image. (4) Model training and evaluation After the model is built, the model parameters are updated through continuous iterative optimization. There are two loss functions used to evaluate the model, corresponding to the two sets of outputs of the model. For the output mask, the binary cross entropy loss function is used as the evaluation function; for the output array, the mean square error loss function is used as the evaluation function. During the training process, the Adam optimizer is used as the optimization algorithm to iteratively update the model parameters, and the learning rate is set to 1×10 -4 ; The early stopping strategy is used to terminate model training early. The specific operations are as follows: the number of model iterations is set to 500, and the rule is established: "If the loss value of the output mask does not decrease within 20 iterations, then terminate the training." The inference phase includes: 1) Data preprocessing Perform pixel intensity normalization on the image to be segmented and feed it into the model as the input image; 2) Model building Build the same model as in the training phase and import the trained model parameters; 3) Model Inference The input image is used as the model input, and after a series of calculations with the trained model parameters, the output mask image is obtained for the doctor's reference.

2. The image segmentation method for intervertebral disc herniation according to claim 1, characterized in that: In the (1) dataset collection during the training phase, two types of data need to be obtained. One is the input image, the file format is dcm format, and the image content is a spinal CT image of a herniated disc; the other is the mask image, the file format is png format, and the image content is that the pixel value of the protruding part is 1, and the rest of the pixel values ​​are 0.

3. The image segmentation method for intervertebral disc herniation according to claim 1, characterized in that: In the (2) data preprocessing during the training phase, the random rotation method is as follows: the input image is rotated at a random angle, and the mask image and the array are also rotated accordingly, keeping the relative positions unchanged. The rotated set of data is also added to the dataset, thus doubling the number of data in the dataset. Pixel intensity normalization method: each pixel of the input image in a set of data is transformed as shown in formula (1). Where x i is the pixel value to be processed, x min is the minimum value of all pixels of the input image, x max is the maximum value of all pixels of the input image; Pixel intensity normalization limits the data to a specified range and eliminates the negative impact caused by singular data.

Images