CN113706548A - Method for automatically segmenting breast anterior mediastinum focus based on CT image - Google Patents

Abstract

The method for automatically segmenting the breast anterior mediastinum focus based on the CT image has a better segmentation result, can meet the quantitative feature analysis of subsequent deep learning, and provides a methodology basis for the automatic segmentation research of the anterior mediastinum focus. The method for automatically segmenting the breast anterior mediastinum lesion based on the CT image comprises the following steps: (1) CT image acquisition: performing chest CT contrast enhanced scanning on all patients to obtain a thin layer image reconstructed by a mediastinum window; (2) on an original CT image, a double-lung mask file is designed by utilizing the natural density difference between the lung and the surrounding tissues and adopting a method for calculating the pixel value of the lung tissue, and the region outside the mediastinum is removed; (3) performing primary segmentation on the focus by adopting a V _ Net network; (4) finely dividing the focus by adopting a Morphological Snakes algorithm; (5) the segmentation results were evaluated.

Description

Method for automatically segmenting breast anterior mediastinum focus based on CT image

Technical Field

The invention relates to the technical field of medical image processing, in particular to a method for automatically segmenting a breast anterior mediastinum focus based on a CT image.

Background

The purpose of preoperative imaging examination of the anterior mediastinal focus of the chest is to preliminarily evaluate the malignancy degree of the tumor, and further to speculate histopathological changes and carry out risk evaluation, so that preoperative treatment scheme selection and clinical prognosis judgment are assisted. The assessment of the signs in CT imaging is based on the lesion itself and its relationship to the adjacent surrounding tissue structures, and often uses empirical, observational rather than quantitative indicators. The image omics is a computer quantitative image analysis method, can extract high-throughput features of medical image images, but the traditional image omics rely on manual delineation for lesion segmentation, and complex delineation standards are often required to be designed to ensure the accuracy of manual segmentation results as much as possible. At present, most tumors are still segmented by adopting manual identification and manual delineation modes on the basis of morphological identification, and the process of manually delineating the focus is time-consuming and easily influenced by human factors due to low automation degree. Therefore, how to develop a method for automatically identifying and ideally segmenting a lesion is one of important research contents in the field of medical image segmentation at present.

The CT image-based automatic segmentation of the anterior mediastinum focus can improve the automation degree of the image omics process, improve the standardization and the accuracy of focus segmentation, and is beneficial to accurately extracting the subsequent image characteristic data and improving the classification performance of a prediction model. In addition, the segmentation of malignant tumor lesions is an important front-end process conventionally executed in radiotherapy, and the precise segmentation of tumors by adopting an automatic segmentation technology can improve the segmentation speed of the lesions before radiotherapy and improve the standardization of the process, thereby facilitating the accurate estimation of the dose in radiotherapy plans and ensuring the standard of curative effect evaluation.

The existing research shows that in the medical image automatic segmentation technology, the focus automatic segmentation based on the deep learning has obvious advantages compared with the traditional method, and the classification and identification performance of the image interested region can be obviously improved and the focus segmentation accuracy can be improved by using the deep convolution neural network architecture. Therefore, the CNN and the related derivative network architecture are adopted to become a preferred algorithm in the field of current medical image lesion segmentation. The networks derived based on CNN mainly include FCN, U-net and V-net. V _ Net is a full convolution neural network which can be used for 3D medical image segmentation and is provided based on the improvement of FCN and U-Net, a volume convolution method is used, a new objective function based on a Dice coefficient is adopted for training, a model is optimized during training, in addition, the V _ Net can also be used for data enhancement in a nonlinear transformation and histogram matching mode, and the characteristics of small medical image data quantity and strong interpretability can be better met. However, because the incidence rate of the anterior mediastinal lesion is relatively low, there is currently less research on the automatic segmentation of CT images, and therefore, how to accurately segment the anterior mediastinal lesion still faces a great challenge.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides a method for automatically segmenting the breast anterior mediastinum focus based on a CT image, which has a better result of finely segmenting the focus, can meet the quantitative feature analysis of subsequent deep learning, and provides a methodology basis for the automatic segmentation research of the anterior mediastinum focus.

The technical scheme of the invention is as follows: the method for automatically segmenting the breast anterior mediastinum focus based on the CT image comprises the following steps:

(1) CT image acquisition: all patients are subjected to CT contrast enhanced scanning to obtain a thin layer image reconstructed at a chest mediastinum window;

(2) on an original CT image, a method of calculating a lung tissue pixel value is adopted by utilizing the natural density difference between a lung and surrounding tissues, a double-lung mask file is designed, the region outside a mediastinum is removed, and a focus is reserved;

(3) performing primary segmentation on the focus by adopting a V _ Net network;

(4) finely dividing the focus by adopting a Morphological Snakes algorithm;

(5) the segmentation results were evaluated.

Firstly, based on an original CT image, a method for calculating lung tissue pixel values is adopted to design a double-lung tissue mask file to remove regions except mediastinum, and only mediastinum regions containing focuses are reserved for preliminary segmentation of the focuses in a V _ Net network, so that the complexity of inputting network images is greatly reduced, and the segmentation precision is favorably improved; after the focus is primarily segmented, the focus is finely segmented by adopting a Morphological Snakes algorithm, which is an evolution method of a contour boundary by adopting a mathematical morphology operator, so that the segmentation edge of the focus can be optimized, the boundary of the final segmentation result is smoother and more accurate, and the segmentation result is efficient and stable, therefore, the fine segmentation result of the focus is better, the quantitative feature analysis of subsequent deep learning can be met, and a methodology basis is provided for the automatic segmentation research of the front mediastinal focus.

Drawings

Fig. 1 shows a schematic diagram of a V _ Net network structure.

Figure 2 shows a Dice coefficient curve for automated lesion segmentation. (A) The graph is a Dice coefficient curve of automatic segmentation of the patient focus in a training set sample; (B) the figure is a Dice coefficient curve of automatic segmentation of the patient focus in a validation set sample.

Fig. 3 shows a flow chart of a method for automatic segmentation of anterior chest mediastinal lesions based on CT images according to the present invention.

Detailed Description

As shown in fig. 3, the method for automatically segmenting the anterior mediastinum lesion of the chest based on the CT image comprises the following steps:

(1) CT image acquisition: performing chest CT enhanced scanning on all patients to obtain a thin-layer image reconstructed on a chest mediastinum window;

(2) on an original CT image, a method of calculating a lung tissue pixel value is adopted by utilizing the natural density difference between a lung and surrounding tissues, a double-lung mask file is designed, the region outside a mediastinum is removed, and a focus is reserved;

(3) performing Initial Segmentation (Initial Segmentation) on the focus by adopting a V _ Net network;

(4) fine Segmentation (Accurate Segmentation) of the lesion using the Morphological Snakes algorithm;

(5) the segmentation results were evaluated.

Firstly, based on an original CT image, a method for calculating lung tissue pixel values is adopted to design a double-lung tissue mask file to remove regions except mediastinum, and only mediastinum regions containing focuses are reserved for preliminary segmentation of the focuses in a V _ Net network, so that the complexity of inputting network images is greatly reduced, and the segmentation precision is favorably improved; after the focus is primarily segmented, the focus is finely segmented by adopting a Morphological Snakes algorithm, which is an evolution method of a contour boundary by adopting a mathematical morphology operator, so that the segmentation edge of the focus can be optimized, the boundary of the final segmentation result is smoother and more accurate, and the segmentation result is efficient and stable, therefore, the fine segmentation result of the focus is better, the quantitative feature analysis of subsequent deep learning can be met, and a methodology basis is provided for the automatic segmentation research of the front mediastinal focus.

Preferably, in the step (1), the CT scan of the patient of the first task group adopts 16 rows (cannon aquilon RXL, Tokyo, Japan), 320 rows MDCT (cannon aquilon One, Tokyo, Japan) and 256 rows MDCT (GE revaluation, Massachusetts, USA), and the CT scan of the patient of the second task group adopts two-source CT (Siemens somehow Definition, formalheim, German), 128 rows MDCT (Siemens somehow Perspective, formalheim, German) and two-source CT (Siemens somehow Flash, formalheim, German); all patients are accommodated in the device and the device adopt automatic tube voltage and automatic tube current regulation technology, contrast agent is injected through elbow vein and adopts iopromide injection 370mgI/ml or ioversol injection 320mgI/ml, the injection speed is 3ml/s, the dosage of the contrast agent is 65-80ml, and enhanced scanning is carried out 40s after the contrast agent is injected; all patients are in supine positions, hold the head with both hands, and order the patients to perform image acquisition when holding breath after inhaling; respectively reconstructing images in the longitudinal separation windows with the window width of 400-450HU and the window position of 20-50HU, and the lung windows with the window width of 1000-1500HU and the window position of-650-450 HU; a 0.5-1.25mm thin slice image reconstructed at the mediastinal window was acquired.

Preferably, in the step (2), firstly, based on a method for calculating an individualized lung tissue threshold and searching for a maximum connected region, a mask file of a two-dimensional matrix of a double lung tissue is designed, a point multiplication operation is performed on the mask and an original image to be processed, a mediastinum 3D image is generated after layer-by-layer cyclic processing by the same algorithm, a breast peripheral tissue structure of a non-mediastinum region is removed from the image, and only a mediastinum region including a focus is reserved.

Preferably, in the step (3), the image is preprocessed before being input into the network, and the preprocessing comprises normalization of a gray level average value and a variance; in the network output up-sampling process, bicubic interpolation is adopted to make up for image dimension change generated in the pooling process; and after the segmentation result output by the V _ Net is obtained, smoothing the selected lesion area.

Preferably, in the step (3), the network structure adopts an end-to-end learning manner, the size of the input image is 256 × 256 × 32, the number of channels is 1, the network is divided into four stages during encoding, each stage includes 1 to 3 convolutional layers, the convolutional kernel used in the convolutional operation of each stage is 3 × 3 × 3, the step size is 1, a prilu nonlinear activation function is used in the whole network structure, each stage of encoding is downsampled through a convolutional kernel with the size of 2 × 2 × 2 and the step size of 2, the network adds the input and output of each stage to obtain the learning of a residual error function, and after the first stage, the dimension of the feature map becomes 16 × 256 × 256 × 32 (the number of channels × the image height × the image width × the image depth); after the second stage, the dimension of the feature map is changed into 32 × 128 × 128 × 16; through the third stage, the dimension of the feature map is changed into 64 × 64 × 64 × 8; after the fourth stage, the dimension of the feature map becomes 64 × 32 × 32 × 4; the decoding stage corresponds to the encoding stage, the up-sampling process combines the characteristics extracted by the corresponding layer in the encoding process, and the size of the convolution kernel used by the last convolution layer is 1 multiplied by 1, so that the size of the output image is kept consistent with the size of the original input image; finally, softmax is used to generate a segmentation probability map of foreground and background (fig. 1).

Preferably, in the step (3), V _ Net is trained by using a volume convolution method and a new objective function Dice loss function based on a Dice coefficient, and is optimized during training, where the Dice coefficient D is formula (1):

wherein the sum runs on N voxels, and the predicted binary segmentation body pi belongs to P and the gold standard binary body gi belongs to G; the Dice coefficient is expressed by a gradient as formula (2):

preferably, in the step (4), on the basis of performing preliminary segmentation on the lesion, a Morphological snake algorithm is adopted to perform fine segmentation, so as to further improve the accuracy of segmentation, and the pixel level set to be solved in the image to be segmented is set as a formula (3):

in the level set framework, the conventional partial differential equation represents a curve evolution equation as formula (4):

when F is ═ 1 and

the above equation is expressed as formula (5) using mathematical morphology:

wherein u represents a level set, represents a gradient operator, div represents a divergence operator, v represents a constant, and g () represents a boundary stopping function, which results in formula (6) after n times of contour boundary evolution:

the result after n times of profile evolution is formula (7):

the number of consecutive applications of the smoothing operator controls the strength of the smoothing step, this number being represented by the parameter mu as N,

the final variation after n iterations is expressed as equations (8) - (10):

wherein D is_dTo expand the operator, E_dIn order to erode the operator,

is a smoothing operator.

Preferably, in the step (5), the segmentation result is evaluated by using three indexes, namely, a Dice coefficient, a precision P, and a recall R, where the Dice coincidence is defined as formula (11):

in the formula, A represents a segmentation result, B represents a manual drawing standard corresponding to the segmentation result, A ^ B is a part belonging to a focus in a segmentation graph and is a correctly predicted part, and the closer to 1 the Dice is, the more accurate the segmentation result is;

the precision is defined as formula (12):

in the formula, TP represents a region in which a lesion is correctly divided, FP represents a region in which a non-lesion is divided into lesion regions, and P represents a percentage of the correctly divided region in the division result.

The recall is defined as equation (13):

where TP represents the region where the lesion is correctly segmented, FN represents the region where the lesion is not correctly segmented, and R represents the percentage of the region where segmentation is correct to the original lesion region.

Drawing a Dice coefficient curve (figure 2) after automatic segmentation of all focuses, wherein in a training set sample, the Dice coefficient minimum value of a patient is 0.623, the Dice coefficient maximum value of the patient is 0.999, and the Dice value of 125 samples is larger than 0.95; in the validation set samples, the patient's Dice coefficient had a minimum of 0.624 and a maximum of 0.998, with 78 samples having Dice values in the range of 0.85-0.95 (67.24%). The results of automatic segmentation of the focus show that the Dice coefficient in the training set is 0.942 +/-0.066, the precision rate is 0.915 +/-0.083 and the recall rate is 0.907 +/-0.091. The Dice coefficient in the verification set is 0.911 +/-0.051, the precision rate is 0.926 +/-0.042, and the recall rate is 0.89 +/-0.059.

In the past, methods for performing segmentation simply based on image inherent features include segmentation of information directly obtainable from an image, such as image gray scale (pixel intensity), gradient or texture, image geometric features, and the like, and the implementation modes include gray threshold segmentation, feature cluster segmentation, region growing modes, and the like. These methods typically require complementary segmentation by hand, and tend to be long and of limited effectiveness for images with low contrast or images lacking significant gradients and geometric features. Compared with the methods, the convolutional neural network based on deep learning, particularly the FCN and the derivative network thereof are used for image segmentation, end-to-end model training can be realized, the segmentation effect is good, the fault-tolerant capability is strong, the method has the characteristics of high-efficiency self-learning capability and self-adaption, and the method is suitable for application in the field of medical image segmentation.

The FCN architecture, which is more popular for medical image segmentation at present, is U-Net. Ronneberger et al creates a resultant output path based on downsampling the encoding and decoding paths corresponding to the upsampling, and combines this method with a jump connection so that combining the higher resolution features of the encoding path with the upsampling features of the decoding path can better locate and learn the feature representation of the input image. The U-Net is divided by taking an input picture as a whole classification based on the classification idea, so that the method overcomes the classification method taking pixels as units, has small calculation amount and can gradually restore the image precision, the improvement based on FCN and U-Net is V _ Net, and the U-Net is a full convolution neural network which can be used for 3D medical image division. V _ Net uses a volume convolution method and adopts a new objective function based on a Dice coefficient to train and optimize during the training, and can better meet the requirements of small image data volume and improvement of segmentation effect at the same time.

On a conventional chest CT image, a scanning range is from a lung tip to a liver top layer, multiple tissue organs are involved, the density change range is large, and the density of double lung tissues, mediastinum and peripheral tissue structures is naturally and greatly compared, so that a method for calculating a lung tissue pixel value is adopted to design a double lung tissue mask file to remove a region except the mediastinum before primary segmentation, and only a mediastinum region containing a focus is reserved for primary segmentation of the focus in a V _ Net network, so that the complexity of an input network image is greatly reduced, and the segmentation precision is favorably improved. Secondly, after the focus is initially segmented, the focus is finely segmented by adopting a Morphological Snakes algorithm, wherein the Morphological Snakes is an evolution method of a contour boundary by adopting a mathematical morphology operator, so that the focus segmentation edge can be optimized, the boundary of the final segmentation result is smoother and more accurate, and the segmentation result is efficient and stable. In the research, a Dice coefficient curve of the automatically segmented focus is displayed in a training set sample, 52.51% of sample Dice values are larger than 0.95, and the result that the Dice values are larger than 0.8 is analyzed, so that the automatically segmented focus result is highly consistent with a manually drawn focus area; analyzing the result that the Dice value is less than 0.8, it can be seen that most of the automatic segmentation results of the focus basically cover the focus area drawn manually, and a small part of the area of the edge of the automatic segmentation image exceeds the edge area of the manually drawn focus in a certain specific direction. In the validation set samples, the patient's Dice coefficient value of 67.24% was in the range of 0.85-0.95. The result shows that the research segmentation model has good effect and can meet the requirement of quantitative feature analysis of subsequent deep learning.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and all simple modifications, equivalent variations and modifications made to the above embodiment according to the technical spirit of the present invention still belong to the protection scope of the technical solution of the present invention.

Claims (8)

1. The method for automatically segmenting the breast anterior mediastinum focus based on the CT image comprises the following steps:

(1) CT image acquisition: all patients are subjected to CT contrast enhanced scanning to obtain a thin layer image reconstructed at a chest mediastinum window;

(2) on an original CT image, a method of calculating a lung tissue pixel value is adopted by utilizing the natural density difference between a lung and surrounding tissues, a double-lung mask file is designed, the region outside a mediastinum is removed, and a focus is reserved;

(3) performing primary segmentation on the focus by adopting a V _ Net network;

(4) finely dividing the focus by adopting a Morphological Snakes algorithm;

(5) the segmentation results were evaluated.

2. The method for automatic segmentation of anterior mediastinal lesion of breast based on CT image as claimed in claim 1, wherein: in the step (1), the patient CT scan of the first task group adopts 16 rows of MDCT, 320 rows of MDCT and 256 rows of MDCT, and the patient CT scan of the second task group adopts double-source CT, 128 rows of MDCT and second-generation double-source CT; all patients are accommodated in the device and the device adopt automatic tube voltage and automatic tube current regulation technology, contrast agent is injected through elbow vein and adopts iopromide injection 370mgI/ml or ioversol injection 320mgI/ml, the injection speed is 3ml/s, the dosage of the contrast agent is 65-80ml, and enhanced scanning is carried out 40s after the contrast agent is injected; all patients are in supine positions, hold the head with both hands, and order the patients to perform image acquisition when holding breath after inhaling; respectively reconstructing images in the longitudinal separation windows with the window width of 400-450HU and the window position of 20-50HU, and the lung windows with the window width of 1000-1500HU and the window position of-650-450 HU; a 0.5-1.25mm thin slice image reconstructed at the mediastinal window was acquired.

3. The method of claim 2 for automatic segmentation of anterior mediastinal lesion of breast based on CT image, wherein: in the step (2), firstly, based on a method for calculating an individualized lung tissue threshold value and searching a maximum connected region, a mask file of a two-dimensional lung tissue matrix is designed, the mask and an original image to be processed are subjected to point multiplication, the same algorithm generates a mediastinum 3D image after layer-by-layer cyclic processing, the chest peripheral tissue structure of a non-mediastinum region of the image is removed, and only the mediastinum region including a focus is reserved.

4. The method of claim 3 for automatic segmentation of anterior mediastinal lesion of breast based on CT image, wherein: in the step (3), the image is preprocessed before being input into a network, and the preprocessing comprises normalization of a gray average value and a variance; in the network output up-sampling process, bicubic interpolation is adopted to make up for image dimension change generated in the pooling process; and after the segmentation result output by the V _ Net is obtained, smoothing the selected lesion area.

5. The method of claim 4 for automatic segmentation of anterior mediastinal lesion of breast based on CT image, wherein: in the step (3), the network structure adopts an end-to-end learning mode, the size of an input image is 256 × 256 × 32, the number of channels is 1, the network is divided into four stages during encoding, each stage includes 1 to 3 convolutional layers, the size of a convolutional kernel used in convolutional operation of each stage is 3 × 3 × 3, the step size is 1, a PReLu nonlinear activation function is used in the whole network structure, each stage of encoding is downsampled through a convolutional kernel with the size of 2 × 2 × 2 and the step size of 2, the network adds the input and the output of each stage to obtain the learning of a residual error function, and after the first stage, the dimension of a feature map is changed into 16 × 256 × 256 × 32; after the second stage, the dimension of the feature map is changed into 32 × 128 × 128 × 16; through the third stage, the dimension of the feature map is changed into 64 × 64 × 64 × 8; after the fourth stage, the dimension of the feature map becomes 64 × 32 × 32 × 4; the decoding stage corresponds to the encoding stage, the up-sampling process combines the characteristics extracted by the corresponding layer in the encoding process, and the size of the convolution kernel used by the last convolution layer is 1 multiplied by 1, so that the size of the output image is kept consistent with the size of the original input image; finally, softmax is used to generate a segmentation probability map of foreground and background.

6. The method of claim 5 for automatic segmentation of anterior mediastinal lesion of breast based on CT image, wherein: in the step (3), V _ Net is trained by using a volume convolution method and a Diceloss function based on a new objective function of a Dice coefficient and optimized during the training, where the Dice coefficient D is formula (1):

wherein the sum runs on N voxels, and the predicted binary segmentation body pi belongs to P and the gold standard binary body gi belongs to G; the Dice coefficient is expressed by a gradient as formula (2):

7. the method of claim 6, wherein the method comprises automatically segmenting the anterior mediastinal lesion based on the CT image, wherein the method comprises: in the step (4), on the basis of primary segmentation of the focus, a Morphological Snakes algorithm is adopted for fine segmentation, so that the segmentation accuracy is further improved, and a pixel level set solved by the image to be segmented is set as a formula (3):

in the level set framework, the conventional partial differential equation represents a curve evolution equation as formula (4):

when F is ═ 1 and

the above equation is expressed as formula (5) using mathematical morphology:

wherein u represents a level set, represents a gradient operator, div represents a divergence operator, v represents a constant, and g () represents a boundary stopping function, which results in formula (6) after n times of contour boundary evolution:

the result after n times of profile evolution is formula (7):

the number of consecutive applications of the smoothing operator controls the strength of the smoothing step, this number being represented by the parameter mu as N,

the final variation after n iterations is expressed as equations (8) - (10):

wherein D is_dTo expand the operator, E_dIn order to erode the operator,

is a smoothing operator.

8. The method of claim 7 for automatic segmentation of anterior mediastinal lesion of breast based on CT image, wherein: in the step (5), the segmentation result is evaluated by using three indexes of a Dice coefficient, a precision rate P and a recall rate R, wherein the Dice coincidence rate is defined as a formula (11):

in the formula, A represents a segmentation result, B represents a manual drawing standard corresponding to the segmentation result, A ^ B is a part belonging to a focus in a segmentation graph and is a correctly predicted part, and the closer to 1 the Dice is, the more accurate the segmentation result is;

the precision is defined as formula (12):

in the formula, TP represents a region in which a lesion is correctly divided, FP represents a region in which a non-lesion is divided into lesion regions, and P represents a percentage of the correctly divided region in the division result.

The recall is defined as equation (13):

where TP represents the region where the lesion is correctly segmented, FN represents the region where the lesion is not correctly segmented, and R represents the percentage of the region where segmentation is correct to the original lesion region.

Images