CN111882561A - Cancer cell identification and diagnosis system - Google Patents

Abstract

The invention discloses a cancer cell identification and diagnosis system, which comprises: a cell image preprocessing module: an improved mathematical morphology on-off filtering algorithm is adopted, and a proper structural element is selected, so that background impurities and small normal discrete cells of a cell image are removed; cell image segmentation module: an improved region growing algorithm is adopted for extracting a cell nucleus region, a watershed algorithm based on price marker control is adopted for extracting the outline of an unlaminated cell body, and a segmentation algorithm based on a snake model is adopted for extracting the outline of the overlapped cell; cancer cell image feature extraction module: and calculating the nuclear-to-cytoplasmic ratio of each segmented region, whether the cell nucleus is uniform in size, abnormal nuclear staining and whether the nuclear distance is uniform. Cell image classification and identification module: cancer cells are identified by neural network identification techniques. The system can efficiently and accurately carry out quantitative analysis and detection identification on the cell image.

Description

A cancer cell identification and diagnosis system

technical field

The invention relates to the technical field of cell image recognition, in particular to a cancer cell recognition and diagnosis system.

Background technique

Cancer is a common disease that endangers human health, and early diagnosis of cancer is the key to treatment. The use of cytology computer-aided diagnosis technology can effectively reduce the work intensity of doctors and improve the accuracy of diagnosis. Due to the complex background of cell images, diverse cell shapes and overlapping problems, it is difficult to detect and identify by computer. The present invention makes new technical improvements to the typical medical image segmentation algorithm, and provides improved algorithms for non-overlapping cell images and overlapping cell images respectively.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a system for identifying and diagnosing cancer cells in view of the deficiencies in the prior art. This diagnostic system can achieve more accurate extraction of target areas, and perform more efficient and accurate quantitative analysis, detection and identification of cell images.

The technical scheme that realizes the object of the present invention is:

A cancer cell identification diagnostic system comprising,

Cell image preprocessing module: First, convert the cell image into a digital image and perform grayscale processing on it. In order to more accurately extract useful cell features in the image, it is necessary to remove small, discrete normal cells and additive noise in the image together. In view of this situation, an improved open-close filter algorithm is used to denoise the image, and the structural element B is selected to be twice that of normal lymphocytes. For the noise smaller than the structuring element B, the open operation is used to eliminate the pepper noise in the background, and then the closed operation is used to eliminate the trachoma noise. For the noise larger than the structural element B, the filtered image is first binarized, and the original image is superimposed and extracted. Image binarization is to convert the original image into a binary image, "0" represents the target image, "1" represents the background, and the binary image is used to select the content of the original image. The value in the binary image is "0". The pixel position retains the original gray value, and the gray value of the pixel position with a value of "1" is "255".

Cell Image Segmentation Module:

The first step: extracting the nucleus region adopts the region growing algorithm of automatic seed points, and combines the threshold segmentation and the region growing algorithm. First, the threshold value T=115 is set, and the image is divided into two parts: the object points larger than the threshold value T and the background points smaller than the threshold value T. The original image f(x,y) is converted into the output image g(x,y), when f(x,y)>T, g(x,y)=1; when f(x,y)≤T, g( x, y)=0. This method achieves "coarse outline" extraction of most regions of the nucleus. Set the growth rule of seeds as the absolute value of the difference between the gray value of adjacent pixels and the average gray value of the seed area is less than the threshold value of 0.15. The centroid of the thick outline is used as the growth point of the seed, and each seed grows iteratively according to the growth rule. Seeds with a value less than the threshold are added until no point that satisfies the rule is found, and the growth ends. This method achieves "fine contour" extraction of cell nuclei. Finally, the Sobel operator is used for edge detection. The template and corresponding formula of the Sobel operator are as follows:

The final output image.

The second step: extracting a single cell body region adopts a marker-controlled watershed algorithm. In order to facilitate the determination of internal and external markers, combined with the background volume of the cell microscope image and the characteristics of the dark target, the grayscale image of the cell was first inverted to obtain an image with a brighter target object and a darker background area. The image is binarized, and the distance of the binary image is changed, so that the dividing line can keep the target object without being too close to the edge of the object. Finally, a watershed transformation is used, and the obtained dividing line is used as an external marker. Reconstruction-based opening and closing operations are performed on the inversely processed image to remove the internal details of the target and use the local maximum area as an internal marker. Combined with the inner and outer markers, the gradient magnitude is calculated from the inversely processed image, and the watershed transformation is performed on the improved gradient image, and the cell body segmentation contour is displayed on the cell grayscale image.

Step 3: Extract the cell contours in the stacked cell images using a segmentation algorithm based on the snake model. In view of the characteristic that the snake model is sensitive to the initial position, the sparse contour point model of the cell is used, and the circular dynamic contour search algorithm is used to automatically locate the contour points of the main cell. Finally, the outline of the cell body is obtained. Its segmentation accuracy is about 94%, the over-segmentation rate is about 2.5%, and the under-segmentation rate is about 3.5%.

Cancer cell image feature extraction module: Calculate the nuclear-to-cytoplasmic ratio of each region after segmentation, whether the size of the nucleus is uniform, whether the nuclear staining is abnormal, and whether the nuclear spacing is uniform.

Nucleocytoplasmic ratio: For non-overlapping cell images, the segment subroutine divides the image into a block and a single cell, and calls the ncratio subroutine to calculate the nucleocytoplasmic ratio of each cell. When the nucleocytoplasmic ratio of the cell is much higher than the normal cell nucleocytoplasmic ratio of 0.5:1, it is determined to be a suspected cancer cell. For the stacked cell images, the segment subroutine is called, the image is divided into sections, and the nuclear-cytoplasmic ratio of the entire area is calculated. When the ratio is much higher than the normal cell nuclear-cytoplasmic ratio of 0.5:1, a suspected cancer cell can be determined.

Nucleus size: The image obtained after segmentation is processed into segments to obtain the size of a single nucleus. If the nuclear size distribution is relatively scattered, it is consistent with the characteristics of cancer cells.

Nuclear chromosome abnormality: By calculating the ratio of nuclear hyperchromatic area to nuclear area, if the proportion is greater than 31.91%, it can be determined as a suspected cancer cell.

The distance between adjacent nuclei of agglomerated cells: Calculate the distance between adjacent nuclei of agglomerated cells by calling distance. If the nuclei of agglomerated cells are not uniformly distributed, a suspected cancer cell can be determined.

Cell image classification and recognition module: using BP neural network recognition technology to identify cancer cells. First initialize the weights and thresholds of the network. The initial value of the network weight is the point at which the network starts to train on the error surface, which affects the training time of the network. Start computing forward, finding all outputs of all neurons. Calculate the output layer error gradient. Then calculate the error gradient of each hidden layer backward, calculate and save the correction amount of each weight, and correct the weight. Determine whether the training goal is achieved. If it is reached, the training is over. If it is not reached, it will be transferred to the forward calculation.

Description of drawings

FIG. 1 is a schematic structural diagram of an embodiment.

FIG. 2 is a schematic flow chart of extracting the nucleus region in the embodiment.

FIG. 3 is a schematic flowchart of the BP neural network identification algorithm in the embodiment

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments, but it is not intended to limit the present invention.

Example:

Referring to Figure 1, a cancer cell identification and diagnosis system includes,

Cell image preprocessing module:

The cell pictures are first converted into digital images and grayscaled. In order to more accurately extract useful cell features in the image, it is necessary to remove small, discrete normal cells and additive noise in the image together. In view of this situation, an improved open-close filter algorithm is used to denoise the image, and the structural element B is selected to be twice that of normal lymphocytes. For the noise smaller than the structuring element B, the open operation is used to eliminate the pepper noise in the background, and then the closed operation is used to eliminate the trachoma noise. For the noise larger than the structural element B, the filtered image is first binarized, and the original image is superimposed and extracted. Image binarization is to convert the original image into a binary image, "0" represents the target image, "1" represents the background, and the binary image is used to select the content of the original image. The value in the binary image is "0". The pixel position retains the original gray value, and the gray value of the pixel position with a value of "1" is "255". As shown in Figure 2, the cell image segmentation module:

The first step: extracting the nucleus region adopts the region growing algorithm of automatic seed points, and combines the threshold segmentation and the region growing algorithm. First, the threshold value T=115 is set, and the image is divided into two parts: the object points larger than the threshold value T and the background points smaller than the threshold value T. The original image f(x,y) is converted into the output image g(x,y), when f(x,y)>T, g(x,y)=1; when f(x,y)≤T, g( x, y)=0. This method achieves "coarse outline" extraction of most regions of the nucleus. Set the growth rule of seeds as the absolute value of the difference between the gray value of adjacent pixels and the average gray value of the seed area is less than the threshold value of 0.15. The centroid of the thick outline is used as the growth point of the seed, and each seed grows iteratively according to the growth rule. Seeds with a value less than the threshold are added until no point that satisfies the rule is found, and the growth ends. This method achieves "fine contour" extraction of cell nuclei. Finally, the Sobel operator is used for edge detection. The template and corresponding formula of the Sobel operator are as follows:

The final output image.

The second step: extracting a single cell body region adopts a marker-controlled watershed algorithm. In order to facilitate the determination of internal and external markers, combined with the background volume of the cell microscope image and the characteristics of the dark target, the grayscale image of the cell was first inverted to obtain an image with a brighter target object and a darker background area. The image is binarized, and the distance of the binary image is changed, so that the dividing line can keep the target object without being too close to the edge of the object. Finally, a watershed transformation is used, and the obtained dividing line is used as an external marker. Reconstruction-based opening and closing operations are performed on the inversely processed image to remove the internal details of the target and use the local maximum area as an internal marker. Combined with the inner and outer markers, the gradient magnitude is calculated from the inversely processed image, and the watershed transformation is performed on the improved gradient image, and the cell body segmentation contour is displayed on the cell grayscale image.

Step 3: Extract the cell contours in the stacked cell images using a segmentation algorithm based on the snake model. In view of the characteristic that the snake model is sensitive to the initial position, the sparse contour point model of the cell is used, and the circular dynamic contour search algorithm is used to automatically locate the contour points of the main cell. Finally, the outline of the cell body is obtained. Its segmentation accuracy is about 94%, the over-segmentation rate is about 2.5%, and the under-segmentation rate is about 3.5%.

Cancer cell image feature extraction module: Calculate the nuclear-to-cytoplasmic ratio of each region after segmentation, whether the size of the nucleus is uniform, whether the nuclear staining is abnormal, and whether the nuclear spacing is uniform.

Nucleocytoplasmic ratio: For non-overlapping cell images, the segment subroutine divides the image into a block and a single cell, and calls the ncratio subroutine to calculate the nucleocytoplasmic ratio of each cell. When the nucleocytoplasmic ratio of the cell is much higher than the normal cell nucleocytoplasmic ratio of 0.5:1, it is determined to be a suspected cancer cell. For the stacked cell images, the segment subroutine is called, the image is divided into sections, and the nuclear-cytoplasmic ratio of the entire area is calculated. When the ratio is much higher than the normal cell nuclear-cytoplasmic ratio of 0.5:1, a suspected cancer cell can be determined.

Nucleus size: The image obtained after segmentation is processed into segments to obtain the size of a single nucleus. If the nuclear size distribution is relatively scattered, it is consistent with the characteristics of cancer cells.

Nuclear chromosome abnormality: By calculating the ratio of nuclear hyperchromatic area to nuclear area, if the proportion is greater than 31.91%, it can be determined as a suspected cancer cell.

The distance between adjacent nuclei of agglomerated cells: Calculate the distance between adjacent nuclei of agglomerated cells by calling distance. If the nuclei of agglomerated cells are not uniformly distributed, a suspected cancer cell can be determined.

As shown in Figure 3, cell image classification and recognition module: using BP neural network recognition technology to identify cancer cells. First initialize the weights and thresholds of the network. The initial value of the network weight is the point at which the network starts to train on the error surface, which affects the training time of the network. Start computing forward, finding all outputs of all neurons. Calculate the output layer error gradient. Then calculate the error gradient of each hidden layer backward, calculate and save the correction amount of each weight, and correct the weight. Determine whether the training goal is achieved. If it is reached, the training is over. If it is not reached, it will be transferred to the forward calculation.

Claims (1)

1. A cancer cell identification and diagnosis system, comprising,

a cell image preprocessing module: firstly, converting a cell picture into a digital image and carrying out gray processing on the digital image. In order to extract useful cell features in an image more accurately, small and discrete normal cells in the image need to be removed together with additive noise. Aiming at the situation, an improved on-off filtering algorithm is adopted to carry out image denoising, and the structural element B is selected to be twice of that of a normal lymphocyte. And for the noise smaller than the structural element B, the opening operation is used for eliminating the pepper-shaped noise in the background, and then the closing operation is used for eliminating the sand hole noise. And for the noise larger than the structural element B, firstly, binarizing the filtered image, and superposing and extracting the original image. The image binarization is to convert an original image into a binary image, wherein '0' represents a target image, and '1' represents a background, the content of the original image is selected by using the binary image, the original gray value is reserved at the pixel position with the value of '0' in the binary image, and the gray value at the pixel position with the value of '1' is '255'.

Cell image segmentation module:

the first step is as follows: and extracting a cell nucleus region, and combining threshold segmentation and a region growing algorithm by adopting a region growing algorithm of an automatic seed point. First, a threshold value T is set to 115, and the image is divided into two parts, i.e., an object point larger than the threshold value T and a background point smaller than the threshold value T. Converting an original image f (x, y) into an output image g (x, y), wherein when f (x, y) > T, g (x, y) is 1; when f (x, y) is less than or equal to T, g (x, y) is 0. The method realizes the extraction of most areas of cell nucleuses by 'coarse contour'. Setting the growth rule of the seeds to be that the absolute value of the difference between the gray value of the adjacent pixel point and the average gray value of the seed area is smaller than the threshold value 0.15, taking the mass center of the coarse contour as the growth point of the seeds, iteratively growing each seed according to the growth rule, adding the seeds of which the difference value is smaller than the threshold value, and ending the growth until the point meeting the rule cannot be found. The method realizes the extraction of cell nucleus with 'fine contour'. And finally, using a Sobel operator for edge detection, wherein the template and the corresponding formula of the Sobel operator are as follows:

and finally, outputting the image.

The second step is that: and extracting a single cell body region and adopting a marker-controlled watershed algorithm. In order to conveniently determine the internal and external marks, the characteristics of background quantity and dark target of the cell microscope image are combined, the cell gray level image is subjected to phase inversion processing, and an image with a bright target object and a dark background area is obtained. And (4) carrying out binarization processing on the image, and carrying out distance change on the binary image so that the boundary line can keep the target object and cannot be too close to the edge of the object. And finally, adopting watershed transformation to obtain a boundary as an external mark. And performing opening operation and closing operation based on reconstruction on the reversely processed image, removing internal details of the target and taking the local maximum value area as an internal mark. And combining the internal and external marks, calculating a gradient amplitude value of the image after the reverse processing, performing watershed transformation on the improved gradient image, and displaying a cell body segmentation contour on the cell gray level image.

The third step: cell contours in the cell images of the stack are extracted using a snake model based segmentation algorithm. Aiming at the characteristic that a snake model is sensitive to an initial position, a sparse contour point model of cells is adopted, a circular dynamic contour searching algorithm is utilized, contour points of a main body cell are automatically positioned, then a snake model method is utilized for segmentation, and a cell body contour is finally obtained through fitting of a curve to the edge of a cell body. The segmentation precision is about 94%, the over-segmentation rate is about 2.5%, and the under-segmentation rate is about 3.5%.

Cancer cell image feature extraction module: and calculating the nuclear-to-cytoplasmic ratio of each segmented region, whether the cell nucleus is uniform in size, abnormal nuclear staining and whether the nuclear distance is uniform.

Nuclear-to-cytoplasmic ratio: for the image of the cells without overlapping layers, the image is divided into blocks and single cells by a segment subroutine, and the ncratio subroutine is called to calculate the nuclear-to-cytoplasmic ratio of each cell. When the nuclear-to-cytoplasmic ratio of the cell is far higher than that of the normal cell by 0.5:1, the cell is judged to be a suspected cancer cell. For the laminated cell image, a segment subroutine is called to divide the image into regions, the nuclear-to-cytoplasmic ratio of the whole region is calculated, and when the ratio is far higher than the normal nuclear-to-cytoplasmic ratio by 0.5:1, the suspected cancer cell can be judged.

Size of cell nucleus: and (4) processing the segmented image by segment partitioning to obtain the size of a single cell nucleus. If the size distribution of the cell nucleus is dispersed, the characteristics of the cancer cells are met.

Nuclear chromosomal abnormalities: and (3) judging the suspected cancer cells by calculating the ratio of the deep staining area of the nuclei to the area of the nuclei if the ratio is more than 31.91 percent.

Adjacent nuclear spacing of clumped cells: and (4) calculating the distance between adjacent nuclei of the clustered cells by calling distant, and if the nuclei of the clustered cells are not uniformly distributed, judging the suspected cancer cells.

Cell image classification and identification module: and (3) identifying the cancer cells by adopting a BP neural network identification technology. The weights and thresholds of the network are first initialized. The initial value of the network weight is the point from which the network starts training on the error surface, and the point influences the training time of the network. The forward calculation is started and all outputs of all neurons are found. An output layer error gradient is calculated. And then calculating the error gradient of each hidden layer backwards, calculating and storing the correction quantity of each weight and correcting the weight. And judging whether the training target is reached. If the result is reached, the training is finished, and if the result is not reached, the calculation is carried out forward.

Images