Method for segmenting adrenal gland region in plain-scan CT (Computed Tomography) image

Abstract

The invention discloses a method for segmenting an adrenal gland region in a plain-scan CT image, and belongs to the technical field of medical data processing. The method comprises the steps of sketching a region of interest, segmenting a training set, sending the training set into a convolutional neural network for training, obtaining a differentiated data set, obtaining a fine segmentation network of three perspectives, obtaining a fusion network training set, obtaining a segmentation result of an adrenal gland region in a CT image, and the like. According to the method, a new training method is provided on the basis of a common two-step method, and an improved cross entropy loss function is supplemented, so that a segmentation model pays attention to an area of which the boundary is difficult to sketch; meanwhile, a multi-view mode is adopted, and a fusion network is introduced, so that the segmentation precision of the model is further improved, and the reliability and accuracy of a model segmentation result are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of medical data processing, and in particular relates to an automatic segmentation model for adrenal gland areas in abdominal plain scan CT images. Background Art

[0002] The adrenal gland is one of the most important endocrine organs in the human body, responsible for the synthesis and secretion of a variety of hormones. It is located on the inner side of the upper pole of both kidneys, one on each side, and can be divided into two parts, the cortex and the medulla, according to the histological structure. A variety of lesions may occur. As a non-invasive examination, CT examination is easy to find smaller lesions due to its high density resolution and high spatial resolution, and it can show certain histological characteristics of lesions, such as fat, fluid, and calcification. For adrenal hyperplasia, atrophy, adrenal medullary tumors and adrenal cysts, CT or other imaging methods should be used for diagnosis. Because the adrenal glands are surrounded by other organs, biopsy is more difficult.

[0003] Existing intelligent segmentation methods for abdominal images generally use enhanced CT as the object. Generally speaking, after preliminary diagnosis, patients with suspicious parts are usually treated with enhanced CT. This method mainly injects iodine-containing contrast agents into the patient's veins. Since iodine has a higher absorption efficiency for radiation, the contrast agent flows through various organs with the blood circulation, and the grayscale value of the corresponding organ in the CT image will be significantly increased, making the boundary of the organ more obvious and easy to distinguish. Although enhanced CT is a very effective method, it requires higher money and time costs compared to plain scan CT. Although the intravenous contrast agent used has no toxic side effects on humans, a few people may still have allergic reactions to the contrast agent; at the same time, in the diagnosis process, it takes time for the contrast agent to flow through the examination site with the blood; after the scan, the patient still needs to stay and observe for dozens of minutes to see if there are any adverse reactions; the metabolism of the contrast agent will also have an impact on human health.

[0004] The application of artificial intelligence to medical images can not only help doctors quickly analyze and interpret image data, improve work efficiency, and enable doctors to provide services to more patients in a shorter time, but also enable cross-regional remote medical consultation and diagnosis through the increasingly developed mobile Internet technology, especially in remote areas or areas with scarce resources, to provide high-quality medical services to more people. Summary of the invention

[0005] The purpose of the present invention is to overcome the problem that most existing methods are aimed at the adrenal glands in enhanced CT images and have low segmentation accuracy, and propose a segmentation method for the adrenal gland area in plain scan CT.

[0006] The technical solution of the present invention is as follows:

[0007] A method for segmenting adrenal gland regions in plain scan CT images comprises the following steps:

[0008] S1: Select an abdominal CT plain scan image without adrenal lesions and outline the region of interest of the adrenal region image;

[0009] S2: The CT images in step S1 are divided into an original training set and a test set according to a ratio of 8:2. The original training set is segmented layer by layer according to the three perspectives of horizontal plane, sagittal plane and coronal plane, which are recorded as the rough segmentation data sets of horizontal plane, sagittal plane and coronal plane respectively; for the three rough segmentation data sets, images with a pixel size of 128*128 are randomly cropped to obtain the fine segmentation data sets of horizontal plane, sagittal plane and coronal plane;

[0010] S3: The three coarse segmentation training sets in step S2 are respectively sent to three U-shaped convolutional neural networks with attention modules and a depth of 5 for training to obtain coarse segmentation networks of three perspectives;

[0011] S4: Send the three coarse segmentation training sets in step S2 to the network of the corresponding viewpoints trained in step S3, obtain the areas where the network prediction results are positive samples, obtain the largest connected domain in these areas, and crop them into images with a pixel size of 128*128 with the largest connected domain as the center, and obtain the coarse segmentation network outputs under three different viewpoints, which are recorded as differential data sets;

[0012] S5: Send the differentiated data sets obtained in step S4 to three U-shaped convolutional neural networks with attention modules and a depth of 3 for training; then send the three fine segmentation data sets obtained in step S2 to the three networks for training;

[0013] Repeat step S5 to obtain the refined segmentation network of three perspectives;

[0014] S6: Divide the original training set described in step S2 into three viewing angles, and predict it by the coarse segmentation network generated in step S4, add the regions with the obtained prediction results as positive samples with equal weights, obtain the maximum connected domain in these regions, and crop it into a 3D image with a pixel size of 128*128*128 with the maximum connected domain as the center, divide the 3D image into three viewing angles, and predict it by the fine segmentation network generated in step S5, and obtain the prediction results of the fine segmentation network under three viewing angles, combine the grayscale value of a certain pixel in the image and the surrounding pixels, and the three fine segmentation network prediction results corresponding to the pixel into a series, and obtain a fusion network training set;

[0015] S7: Send the fusion network training set generated in step S6 to a fully connected network for training to obtain a three-view fusion network. For the test set generated in step S2, predict step by step according to step S6 to obtain three fine segmentation network prediction results. Pass the results into the three-view fusion network obtained in to obtain the segmentation result of the adrenal gland area in the CT image.

[0016] Further, in step S5, when training the fine segmentation network, the loss function used includes an improved cross entropy loss function, which focuses on the predicted difference area between the coarse segmentation network and the fine segmentation network, and its formula is defined as:

[0017]

[0018] in, is the model's predicted probability for the sample, defined as:

[0019]

[0020] is the label value of the adrenal gland region delineated in step S1, is the model’s prediction result for the adrenal region, is the focus factor, is the adjustment variable of the focus factor (taken as 2 in the present invention), is the weight value, and its value is: When a sample is predicted as a positive sample by the fine segmentation network but predicted as a negative sample by the coarse segmentation network, , in other cases, ;

[0021] In step S5, a differentiated training method is used: the differentiated data set obtained in step S4 is sent to three U-shaped convolutional neural networks with attention modules and a depth of 3 layers for training, and the improved cross entropy loss function FL and DSC coefficient loss function F are used. dice Combined, the formula is expressed as:

[0022]

[0023] in, Represents the prediction result of the fine segmentation network, represents the prediction result of the coarse segmentation network, This is the region of interest in the first step.

[0024] Further, in step S7, a fully connected network that fuses the multi-view output is used, which takes a certain pixel The grayscale values ​​of all pixels in the 5*5*5 range centered on the center and the prediction results of the three fine segmentation networks corresponding to all pixels in the 5*5*5 range are combined into a series as input, and the output is The prediction results .

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The present invention is based on the segmentation of the adrenal gland region in plain scan CT images using a U-shaped convolutional neural network with an attention module. A new training method is proposed based on the commonly used two-step method, and an improved cross-entropy loss function is used to make the segmentation model focus on the difficult-to-segment boundary area. At the same time, a multi-view method is used and a fusion network is introduced to further improve the segmentation accuracy of the model, thereby improving the reliability and accuracy of the model segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the adrenal gland segmentation network.

[0028] Figure 2 Three views of the human body.

[0029] Figure 3 Schematic diagram of the network structure of a U-shaped convolutional neural network with a depth of 5 layers and an attention module.

[0030] Figure 4 Schematic diagram of the network structure of a U-shaped convolutional neural network with a depth of 3 layers and an attention module.

[0031] Figure 5 This is a schematic diagram of the fusion network structure. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] Example 1 A method for segmenting the adrenal gland region in a plain scan CT image, the method steps are as follows:

[0034] Step 1: Select the abdominal plain scan CT image containing the adrenal gland area and outline the mask of the adrenal gland area image: Use the medical imaging software ITK-SNAP to circle the tumor area and fill it with a pixel grayscale value of 1. The default pixel grayscale value of other uncircled areas is 0 to generate the region of interest (ROI) for adrenal gland segmentation.

[0035] In order to improve the performance and accuracy of the model, this example collected some samples of adrenal glands with adrenal cortical hyperplasia. The adrenal glands with cortical hyperplasia show diffuse enlargement in CT images, and the basic morphology does not change significantly. The density and morphology remain normal. Therefore, it is possible to use a single network to segment the adrenal glands with hyperplasia and normal morphology at the same time.

[0036] Table 1 shows the distribution of CT data

[0037]

[0038] Step 2: Divide the data set in the first step into the original training set and the test set according to the ratio of 8:2. Figure 2 The horizontal, sagittal and coronal planes in the image are saved layer by layer to generate three coarse segmentation datasets under different perspectives. For the three coarse segmentation datasets, a random box selection method is used to randomly select images of size 128*128 from each positive sample containing the region of interest in the coarse segmentation dataset as the fine segmentation training set.

[0039] Step 3: Build three Figure 3 The 5-layer U-shaped convolutional neural network with attention module shown in the figure is used as the coarse segmentation network of three perspectives. The three coarse segmentation data sets generated in the second step are sent to the three coarse segmentation networks respectively, and Dice Loss is used as the loss for training to obtain the coarse segmentation networks under three perspectives. The Dice Loss formula is expressed as:

[0040]

[0041] Among them, DSC is the Dice Similarity Coefficient, P refers to the prediction result of the model, and B refers to the outlined label value. It refers to the intersection of the two. Refers to the union of the two.

[0042] Step 4: Send the three coarse segmentation data sets generated in the second step to the coarse segmentation network of the corresponding view that has been trained. The positive samples are recorded as positive samples. For the positive sample area, the largest connected domain is obtained, and the image is cropped into a 128*128 image with the largest connected domain as the center, and the differentiated data sets under three different perspectives are obtained.

[0043] Step 5: Build three Figure 4The U-shaped convolutional neural network with a depth of 3 layers and an attention module is used as the fine segmentation network for the three perspectives. The three min-max values ​​generated in the fourth step are sent to the network, and the improved cross entropy loss function ( ) and Dice loss function ( ), the min-max training method is used for training, and its formula is expressed as:

[0044]

[0045] in, Represents the prediction result of the fine segmentation network, represents the prediction result of the coarse segmentation network, That is, the region of interest in the first step. The improved cross entropy loss function is defined as:

[0046]

[0047] in, is the weight value, and its value is: When a sample is predicted as a positive sample by the fine segmentation network but predicted as a negative sample by the coarse segmentation network, , in other cases, .

[0048] Using a differentiated training method, the prediction results of the fine segmentation network are made as close as possible to the region of interest, while being as different as possible from the prediction results of the coarse segmentation network under this perspective; at the same time, The loss function enables the fine segmentation network to pay more attention to difficult-to-segment or controversial areas, further improving the effect of network segmentation.

[0049] After training the fine segmentation network for one round using the differentiation method, the fine segmentation training set generated in the second step is sent to the fine segmentation network and trained for one round using Dice Loss as the loss. Two rounds of training are regarded as a complete round of the fine segmentation network to train the fine segmentation network.

[0050] Step 6: Divide the original training set described in the second step into three perspectives, and use the coarse segmentation network generated in the fourth step to predict. Add the three areas with positive prediction results with equal weights, obtain the largest connected domain in these areas, and crop it into a 3D image with a pixel size of 128*128*128 with the largest connected domain as the center. Divide the above 3D image into three perspectives and use the fine segmentation network generated in the fourth step to predict, and obtain the fine segmentation network prediction results under three perspectives. Combine the grayscale value of a pixel in the image and the surrounding pixels, and the three fine segmentation network prediction results corresponding to the pixel into a series to obtain a fusion network training set.

[0051] Step 7: Send the fusion network training set generated in S6 to a Figure 5 The fully connected network shown in the figure is trained to obtain a three-view fusion network. For the test set generated in the second step, the network is first predicted by the coarse segmentation generated in the fourth step, and the three regions with positive prediction results are added with equal weights to obtain the largest connected domain in these regions. The 3D image with a pixel size of 128*128*128 is cropped with the largest connected domain as the center. The above 3D image is divided into three perspectives and predicted by the fine segmentation network generated in the fourth step to obtain the prediction results of the fine segmentation network under three perspectives. The network takes a certain pixel as the center. The grayscale values ​​of all pixels in the 5*5*5 range centered on the center and the prediction results of the three fine segmentation networks corresponding to all pixels in the 5*5*5 range are combined into a series as input, and the output is The prediction results .

[0052] The complete segmentation process is as follows Figure 1 shown.

[0053] Table 2 shows the segmentation results (expressed in DSC coefficients) at different stages and perspectives of the segmentation model.

[0054]

[0055] It can be seen from Table 2 that although the input size of the fine segmentation network is smaller, due to the use of the differentiation method, the output results of the fine segmentation network contain more non-ROIs with similar ROI features. Therefore, the evaluation index of the fine segmentation network under the same viewing angle is inferior to that of the coarse segmentation network, but the fusion network can distinguish the non-ROI well and finally achieve excellent segmentation effect.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for segmenting adrenal gland regions in plain scan CT images, comprising the following steps: S1: Select an abdominal CT plain scan image without adrenal lesions and outline the region of interest of the adrenal region image; S2: The CT images in step S1 are divided into an original training set and a test set according to a ratio of 8:

2. The original training set is segmented layer by layer according to the three perspectives of horizontal plane, sagittal plane and coronal plane, which are recorded as the rough segmentation data sets of horizontal plane, sagittal plane and coronal plane respectively; for the three rough segmentation data sets, images with a pixel size of 128*128 are randomly cropped to obtain the fine segmentation data sets of horizontal plane, sagittal plane and coronal plane; S3: The three coarse segmentation training sets in step S2 are respectively sent to three U-shaped convolutional neural networks with attention modules and a depth of 5 for training to obtain coarse segmentation networks of three perspectives; S4: Send the three coarse segmentation training sets in step S2 to the network of the corresponding viewpoints trained in step S3, obtain the areas where the network prediction results are positive samples, obtain the largest connected domain in these areas, and crop them into images with a pixel size of 128*128 with the largest connected domain as the center, and obtain the coarse segmentation network outputs under three different viewpoints, which are recorded as differential data sets; S5: Send the differentiated data sets obtained in step S4 to three U-shaped convolutional neural networks with attention modules and a depth of 3 for training; Then the three fine segmentation data sets obtained in step S2 are respectively sent to the three networks for training; Repeat step S5 to obtain the refined segmentation network of three perspectives; S6: Divide the original training set described in step S2 into three viewing angles, and predict it by the coarse segmentation network generated in step S4, add the regions with the obtained prediction results as positive samples with equal weights, obtain the maximum connected domain in these regions, and crop it into a 3D image with a pixel size of 128*128*128 with the maximum connected domain as the center, divide the 3D image into three viewing angles, and predict it by the fine segmentation network generated in step S5, and obtain the prediction results of the fine segmentation network under three viewing angles, combine the grayscale value of a certain pixel in the image and the surrounding pixels, and the three fine segmentation network prediction results corresponding to the pixel into a series, and obtain a fusion network training set; S7: Send the fusion network training set generated in step S6 to a fully connected network for training to obtain a three-view fusion network. For the test set generated in step S2, predict step by step according to step S6 to obtain three fine segmentation network prediction results. Pass the results into the three-view fusion network obtained in to obtain the segmentation result of the adrenal gland area in the CT image.

2. The method for segmenting the adrenal gland region in a plain scan CT image according to claim 1, characterized in that: In step S5, when training the fine segmentation network, the loss function used includes an improved cross entropy loss function, which focuses on the predicted difference area between the coarse segmentation network and the fine segmentation network, and its formula is defined as: ； in, is the model's predicted probability for the sample, defined as: ； is the label value of the adrenal gland region delineated in step S1, is the model’s prediction result for the adrenal region, is the focus factor, is the moderating variable of the focal factor, is the weight value, and its value is: When a sample is predicted as a positive sample by the fine segmentation network but predicted as a negative sample by the coarse segmentation network, , in other cases, ; In step S5, a differentiated training method is used: the differentiated data set obtained in step S4 is sent to three U-shaped convolutional neural networks with attention modules and a depth of 3 layers for training, and the improved cross entropy loss function FL and DSC coefficient loss function F are used. dice Combined, the formula is expressed as: ； in, Represents the prediction result of the fine segmentation network, represents the prediction result of the coarse segmentation network, This is the region of interest in the first step.

3. The method for segmenting the adrenal gland region in a plain scan CT image according to claim 1, characterized in that: In step S7, a fully connected network that fuses the multi-view output is used. The grayscale values ​​of all pixels in the 5*5*5 range centered on the center and the prediction results of the three fine segmentation networks corresponding to all pixels in the 5*5*5 range are combined into a series as input, and the output is The prediction results .

Images