CN104715484B - Automatic tumor imaging region segmentation method based on improved level set - Google Patents

Abstract

The present invention provides an automatic tumor image region segmentation method based on an improved level set, comprising: acquiring an original PET image including a lesion area to be segmented and performing preprocessing and positioning to determine the preprocessed PET image of a lesion area to be segmented; Construct a hypergraph according to the CT image of the lesion area and the PET image of the lesion area to be segmented after the preprocessing, thereby preliminarily determining the rough tumor area in the PET image as an initial zero level set; performing an improved level on the initial zero level set A set method is used to determine the tumor area; edge smoothing is performed on the tumor area according to the morphological operation. The method of the invention can realize rapid and accurate segmentation of tumor regions, thereby assisting surgeons in diagnosis, treatment and curative effect evaluation.

Description

Automatic Tumor Image Region Segmentation Method Based on Improved Level Set

technical field

The invention relates to image processing technology, in particular to an automatic tumor image region segmentation method based on an improved level set.

Background technique

Cervical cancer is one of the three most common malignant tumors of female reproductive organs, and one of the main malignant tumors that threaten women's life and affect the quality of life, ranking first among female reproductive organ malignant tumors. According to the 2014 World Cancer Report published by The International Agency for Research on Cancer under the World Health Organization at its headquarters in Lyon, France on February 3, the number of new cases of cervical cancer worldwide in 2012 reached more than 500,000. Among female malignant tumors, it ranks fourth after breast cancer, rectal cancer, and lung cancer. Over the same period, cervical cancer caused more than 260,000 deaths, and its mortality rate is second only to breast cancer, lung cancer, and rectal cancer. The fourth highest cancer death rate among women. Cervical cancer is the most common cancer among women in less developed countries. In recent years, the incidence of cervical cancer in young women has been increasing, and it has become one of the three major cancers that young women are prone to suffer from. my country is a big country in the incidence and death of cervical cancer, accounting for about one-third of the world's incidence and mortality. Therefore, accurate diagnosis of cervical cancer patients is very important. Positron emission tomography (PET) and computerized tomography (Computed Tomography, CT) as molecular imaging methods are commonly used detection methods in the field of clinical tumors. Quantitative analysis of tumors using PET/CT can provide clinical Provide accurate diagnostic information and assist in formulating treatment plans.

At present, the most commonly used quantitative analysis index in clinical practice is the standard uptake value (SUV), which is equal to the radioactive concentration of the lesion (kBq/ml) divided by the injected dose (MBq) and then divided by the body weight (kg), followed by tumor Volume, or MTV value, has also been shown to predict tumor recurrence and assess prognosis. However, these quantitative indicators all depend on the accurate delineation of the tumor area. In addition, in the radiotherapy plan for cervical cancer, it also depends on the accurate delineation of the target area. Considering the inefficiency and high subjectivity of manual segmentation, automatic and accurate segmentation of cervical cancer tumors is very necessary. However, compared with other tumors, the delineation of the tumor area of cervical cancer faces more challenges: on the one hand, because the attenuation coefficients of the tumor and the cervical parenchyma are the same, it is difficult to distinguish accurately on CT images; The location of the cervix is very close, and the radioactive activity of the urine in the bladder is greater than or approximately equal to that of the tumor, so it is difficult to automatically extract it from the PET image.

Contents of the invention

The invention provides an automatic tumor image region segmentation method based on an improved level set, which is used to solve the problem that it is difficult to automatically distinguish the tumor region and the bladder region in cervical cancer tumor segmentation, so that users can quickly and accurately segment cervical cancer tumors to assist Surgeons conduct diagnosis, treatment and evaluation of curative effect.

The automatic tumor image region segmentation method based on the improved level set of the present invention includes:

Obtain the original PET image containing the lesion area to be segmented and perform preprocessing and positioning to determine the preprocessed PET image of the lesion area to be segmented;

Constructing a hypermap according to the CT image of the lesion area and the PET image of the lesion area to be segmented after the preprocessing, so as to preliminarily determine that the rough tumor area in the PET image is an initial zero level set;

performing a modified level set method on said initial zero level set to determine tumor regions;

Edge smoothing is performed on the tumor region according to the morphological operation.

The beneficial effects of the present invention are:

The present invention proposes an automatic tumor image region segmentation method based on an improved level set, which solves the problem that it is difficult to automatically distinguish the tumor region and the bladder region in cervical cancer tumor segmentation, and enables users to quickly and accurately segment cervical cancer tumors to assist surgery. Doctors carry out diagnosis, treatment and curative effect evaluation. The method of the present invention has the advantages of fast speed, high precision and strong robustness. Experimental results show that this technology can accurately and automatically outline cervical cancer tumors, and realize automatic distinction between tumors and bladders. It has great practical value in clinical diagnosis and treatment.

Description of drawings

Fig. 1 is the flowchart of the automatic tumor image region segmentation method based on the improved level set of the present invention;

Fig. 2 is a schematic diagram of positioning the lesion region to be segmented in the automatic tumor image region segmentation method based on the improved level set of the present invention;

Fig. 3 is a schematic diagram of the improved level set method iteration in the automatic tumor image region segmentation method based on the improved level set in the present invention;

Fig. 4 is a schematic diagram of the segmentation results and the gold standard of tumor regions in three typical cervical cancer data using the improved level set-based automatic tumor image region segmentation method of the present invention;

5 is a schematic diagram of the consistency between the segmentation results of the automatic tumor image region segmentation method based on the improved level set of the present invention and the gold standard in quantitative analysis.

Detailed ways

Fig. 1 is the flowchart of the automatic tumor image region segmentation method based on the improved level set of the present invention, as shown in Fig. 1, the automatic tumor image region segmentation method based on the improved level set of the present invention comprises:

S1. Acquiring the original PET image containing the lesion area to be segmented and performing preprocessing and positioning to determine the preprocessed PET image of the lesion area to be segmented;

Preferably, the acquisition of the original PET image containing the lesion area to be segmented and performing preprocessing and positioning so as to determine the preprocessed PET image of the lesion area to be segmented comprises:

Divide the voxel gray value in the original PET image containing the lesion area by the injected contrast agent dose and the patient's body weight to convert it into an SUV value, and then perform Gaussian filtering and up-sampling, so that the to-be-segmented The resolution of the PET image is the same as that of the CT image, and finally the preprocessed PET image of the lesion area to be segmented is located and determined according to the SUV value.

Preferably, the locating and determining the preprocessed PET image of the lesion region to be segmented according to the SUV value includes:

preprocessing, including:

The gray value of each voxel in the PET image was converted into an SUV value by dividing by the dose of injected 18F-FDG and the patient's body weight, and then Gaussian filtering and upsampling were performed to make the resolution the same as that of the CT image. At the same time, the CT image is also subjected to the same Gaussian filtering;

and positioning process, including:

Calculate the SUV peak value of each slice in the original PET image to be segmented including the lesion area (SUVpeak, indicating the average value of the SUV value of each voxel in the 26 neighborhoods of the voxel corresponding to SUVmax, that is, the maximum SUV value), and select the foot The slices corresponding to the adjacent two minimum SUV values of the largest SUV peak above are used as the slices where the lesion is located, so as to determine the preprocessed PET image of the lesion area to be segmented.

S2. Construct a hypermap according to the CT image of the lesion area and the preprocessed PET image of the lesion area to be segmented, so as to preliminarily determine that the rough tumor area in the PET image is an initial zero level set;

Preferably, the constructing a hypermap according to the CT image of the lesion area and the preprocessed PET image of the lesion area to be segmented, so as to preliminarily determine the rough tumor area in the PET image as an initial zero-level set includes:

The preprocessed PET image of the lesion area to be segmented and the CT image of the lesion area are normalized and multiplied to construct a hypergraph, and then the fuzzy C-means clustering method, morphological erosion and general threshold method are used to preliminarily determine The rough tumor area in the PET image is an initial zero level set, including:

First build a hypergraph, including:

Each voxel of the hypermap consists of three features, namely: the normalized SUV value of the corresponding voxel in the PET image (ie SUV/SUVmax), and the normalized HU value of the corresponding voxel in the CT image (HU/HUmax , HU represents the CT value of each voxel), and their product. According to root tissue specificity, the hypermap can be divided into four parts: (a) large SUV and HU represent tumors; (b) high SUV but low HU represent bladder; (c) low SUV but high HU represent other Soft tissue; (d) Representative background with low SUV and HU.

Then use the fuzzy C-means clustering method to divide the hypergraph into 4 categories, among which the three features are relatively large, which represent tumors, and use the method of morphological corrosion to corrode the bladder wall that may be misclassified as a tumor area, and then use The threshold of 40% SUVmax commonly used in clinical practice is used to obtain a rough tumor area.

S3. Performing an improved level set method on the initial zero level set to determine a tumor region;

Preferably, performing the improved level set method on the initial zero level set so as to determine the tumor area comprises:

The gradient field information of the preprocessed PET image of the lesion area to be segmented is added to the initial zero-level set, and a new evolution equation is constructed, and the rough tumor area in the PET image is compared with the resolution of the original PET image Perform downsampling at the same resolution to obtain an initial zero level set;

Performing multiple iterations on the evolution equation for the initial zero level set according to the finite difference method to determine the tumor area that is finally segmented;

Preferably, the improved level set method includes:

Considering that both the tumor and the bladder after Gaussian filtering have the characteristic of being bright in the middle and dark at the edge, so the direction of the gradient field at the edge of the tumor and the bladder is opposite, so the new level set evolution equation that can be constructed is as follows:

Among them, I _σ is the PET image of the lesion area to be segmented after preprocessing, (the variance of the Gaussian kernel is σ), φ is the function of the initial zero level set or the level set after several iterations of the initial zero level set, <*> Represents the angle between the gradient vectors of I _σ and φ, |*| represents the magnitude of the vector, c ₁ and c ₂ represent the voxels inside and outside the initial zero level set (or the initial zero level set) respectively The average gray value, namely:

Among them, Ω represents the image area,

The delta function is approximated by the following smoothing function δε:

Φ is a function to adjust the initial zero level set or the level set after several iterations of the initial zero level set and make it continuous. After each iteration, the following Gaussian filtering needs to be performed on the level set function:

φ=G _σ *φ,

Among them, G _σ is the Gaussian kernel with variance σ, * represents the convolution operation, and the definition of the initial function φ0 is as follows:

Among them, c0 is a normal number, and R0 is the rough tumor area obtained by downsampling to the resolution of the original PET image;

Considering that for 3D images, the level set function φ(x, y, z, t) can be discretized as Where (i, j, k) is the space coordinate, n is the time coordinate, then the evolution equation can be discretized as:

Where L is the right side of the equal sign of the evolution equation, then the evolution process can be iterated as follows:

Figure 3 is a schematic diagram of the iterative process.

S4. Perform edge smoothing processing on the tumor region according to the morphological operation.

Preferably, the performing edge smoothing processing on the tumor region according to the morphological operation includes:

performing morphological opening and closing operations on the tumor region using spherical structuring elements to eliminate edge protrusions and fill cavities;

Preferably, the specific morphological operation refers to the use of spherical structural elements to perform morphological opening and closing operations on the tumor region obtained in S103 to eliminate edge protrusions, and the spherical structural element s(x, y, z )as follows:

The use of spherical structural elements to perform morphological opening and closing operations on the tumor region to eliminate edge protrusions and fill cavities includes: using formulas (6) and (7) to perform opening and closing operations respectively:

Open operation processing:

Closing operation processing:

in, represents the dilation operator symbol, Indicates the erosion operation symbol, and M indicates the binary image of the tumor area.

It should be noted that, in order to verify the effectiveness and practicability of the present invention, we conducted experiments on clinical PET/CT images, and the gold standard was the average value of manual segmentation results by two experts.

Through a large number of experiments, the Dice similarity coefficient (that is, the Dice similarity coefficient, which measures the overlap rate between the segmentation result and the gold standard) is 91.80±2.46%, and the Hausdorff distance (measures the maximum degree of mismatch between the segmentation result and the gold standard) is 77.79±2.18mm, and Figure 4 shows the gold standard and segmentation results of three typical images; Figure 5(a) and (b) show the quantitative indicators SUVmean, MTV and Bland between the gold standard and the tumors segmented using this technology -Altman plot illustrating the high agreement of this technique with the gold standard. Experiments show that our method can well meet the needs of clinical diagnosis and auxiliary treatment planning, and has great practical value.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: It is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the various embodiments of the present invention. scope.

Claims (5)

1. an automatic tumor image region segmentation method based on improved level set, it is characterized in that, comprising: Obtain the original PET image containing the lesion area to be segmented and perform preprocessing and positioning to determine the preprocessed PET image of the lesion area to be segmented; Constructing a hypermap according to the CT image of the lesion area and the PET image of the lesion area to be segmented after the preprocessing, so as to preliminarily determine that the rough tumor area in the PET image is an initial zero level set; performing a modified level set method on said initial zero level set to determine tumor regions; Edge smoothing is performed on the tumor region according to the morphological operation. 2. The automatic tumor image region segmentation method based on the improved level set according to claim 1, characterized in that, the acquisition includes the original PET image to be segmented of the lesion area and performs preprocessing and positioning so as to determine the preprocessing The PET image of the lesion area to be segmented includes: Divide the voxel gray value in the original PET image containing the lesion area by the injected contrast agent dose and the patient's body weight to convert it into an SUV value, then perform Gaussian filtering and up-sampling, and finally according to the SUV Value positioning and determination of the PET image of the lesion area to be segmented after preprocessing; The step of constructing a hypermap according to the CT image of the lesion area and the preprocessed PET image of the lesion area to be segmented, so as to preliminarily determine the rough tumor area in the PET image as an initial zero-level set includes: The preprocessed PET image of the lesion area to be segmented and the CT image of the lesion area are normalized and multiplied to construct a hypergraph, and then the fuzzy C-means clustering method, morphological erosion and general threshold method are used to preliminarily determine The rough tumor area in the PET image is an initial zero level set; The performing the improved level set method on the initial zero level set to determine the tumor area comprises: The gradient field information of the preprocessed PET image of the lesion area to be segmented is added to the initial zero-level set, and a new evolution equation is constructed, and the rough tumor area in the PET image is compared with the resolution of the original PET image Perform downsampling at the same resolution to obtain an initial zero level set; The tumor area is determined by performing multiple iterations on the evolution equation on the initial zero level set according to the finite difference method. 3. The automatic tumor image region segmentation method based on the improved level set according to claim 2, wherein said positioning according to the SUV value and determining the preprocessed PET image of the lesion region to be segmented comprises: Calculate the SUV peak value of each slice in the original PET image containing the lesion area to be segmented, and select the corresponding slice between the two adjacent minimum SUV values of the largest SUV peak value above the foot as the slice where the lesion area is located to determine the post-processing The PET image of the lesion area to be segmented; Correspondingly, the evolution equation is: Among them, I _σ is the PET image of the lesion area to be segmented after preprocessing, (the variance of the Gaussian kernel is σ), φ is the function of the initial zero level set or the level set after several iterations of the initial zero level set, <*> Represents the angle between the gradient vectors of I _σ and φ, |*| represents the magnitude of the vector, c ₁ and c ₂ represent the voxels inside and outside the initial zero level set (or the initial zero level set) respectively The average gray value, namely: Among them, Ω represents the image area, The delta function is approximated by the following smoothing function _δε : Φ is a function to adjust the initial zero level set or the level set after several iterations of the initial zero level set and make it continuous. After each iteration, the following Gaussian filtering needs to be performed on the level set function: φ=G _σ *φ, Among them, G _σ is the Gaussian kernel with variance σ, * represents the convolution operation, and the initial function φ ₀ is defined as follows: Among them, c ₀ is a normal number, and R ₀ is the rough tumor area obtained by downsampling to the resolution of the original PET image. 4. The automatic tumor image region segmentation method based on the improved level set according to claim 1, wherein the performing edge smoothing processing on the tumor region according to the morphological operation comprises: The spherical structure elements are used to perform morphological opening and closing operations on the tumor region to eliminate edge protrusions and fill cavities. 5. The automatic tumor image region segmentation method based on improved level set according to claim 4, characterized in that, the use of spherical structural elements performs morphological opening and closing operations on the tumor region to eliminate edge protrusions And filling the void includes: using formulas (6) and (7) to perform opening and closing operations respectively: Open operation processing: Closing operation processing: in, represents the dilation operator symbol, Indicates the erosion operation symbol, and M indicates the binary image of the tumor area.