CN112200801B - Automatic detection method for cell nucleus of digital pathological image - Google Patents

Abstract

The invention discloses an automatic detection method for cell nucleus of digital pathological images, which comprises acquiring fragment images of digital pathological images and standardizing them; inputting the normalized fragment images into a trained cell nucleus detection model for detection, Obtain the cell nucleus set S _N1 ; cover the detected cell nucleus in the fragment image with the background color, and use the image gradient energy function to calculate the energy value of each fragment image; judge whether the energy value of each fragment image is greater than the preset threshold, if so, Go to the next step, otherwise discard the corresponding fragment images; input all the fragment images with energy values greater than the preset threshold into the trained cell nucleus detection model for detection to obtain the cell nucleus set S _N2 ; merge the cell nucleus set S _N1 and the cell nucleus set S _N2 , to obtain the final set of nuclei S _N .

Description

A method for automatic detection of cell nuclei in digital pathological images

technical field

The invention relates to a cell nucleus detection technology in an image, in particular to a cell nucleus automatic detection method of a digital pathological image.

Background technique

The occurrence and development of cancer is the result of the interaction between cancer cells and the tumor microenvironment. The changes in the type, number or shape of cells in the tumor stroma have important medical guiding significance. For example, lymphocytic infiltrates in breast cancer generally have a better prognosis, while the presence of tumor-associated fibroblasts suggests a poor prognosis. In routine pathological work, changes in cellular components and extracellular matrix in the tumor stroma are generally described qualitatively. Based on digital pathological image analysis, different components in the interstitium can be automatically segmented and quantitative or qualitative studies can be performed.

In quantitative studies, using quantitative nuclear morphometry, the researchers found that there was a significant difference in the prognosis of nuclei between patients with low nuclear area and high nuclear area. To extract a rich set of quantitative features in breast cancer epithelium and stroma (6642 features), researchers at Stanford University developed the C-Path system (Computational Pathologist) to measure standard morphological descriptors including image objects and higher Levels of contextual, relational, and global image features.

The detection of nuclei in pathological WSI images is the basis for the quantitative analysis of the entire image, and provides reliable support for the quantitative analysis and determination of biological indicators required for various medical research. Although many methods for nuclear detection based on digital pathology have been proposed, it is still a very challenging task. The main reasons are:

(1) The difference between WSI images is subtle, the cells overlap, and the color distribution is uneven;

(2) The lack of large-scale public and labeled data sets brings certain difficulties to algorithm research;

(3) Different from other tumors, the size, density, and shape of the breast have great individual differences, resulting in complicated digital imaging.

(4) The manual labeling of nuclei is very heavy, and it is difficult to form a large labeling database.

To realize the detection of cell nuclei based on traditional machine learning, a large number of parameters need to be manually adjusted, the accuracy is difficult to improve, and it is difficult to adapt to various complex equipment and staining differences; models based on deep learning methods often require a large number of labeled training samples, and need to be carefully adjusted Model parameters and training neural network models, which can take a lot of time.

SUMMARY OF THE INVENTION

Aiming at the above deficiencies in the prior art, the present invention provides an automatic detection method for cell nuclei of digital pathological images with high detection accuracy.

In order to achieve the above-mentioned purpose of the invention, the technical scheme adopted in the present invention is:

Provided is a method for automatic detection of cell nuclei in digital pathological images, comprising:

S1. Acquiring fragment images of digital pathological images and standardizing them;

S2, input the fragmented images after the standardized processing into the trained cell nucleus detection model for detection, and obtain the cell nucleus set S _N1 ;

S3. Cover the cell nucleus detected in the fragment image with the background color, and use the image gradient energy function to calculate the energy value of each fragment image;

S4, determine whether the energy value of each fragment image is greater than the preset threshold, if so, go to step S5, otherwise discard the corresponding fragment image;

S5, input all the fragmented images whose energy values are greater than the preset threshold into the trained cell nucleus detection model for detection, and obtain the cell nucleus set S _N2 ;

S6. Merge the cell nucleus set S _N1 and the cell nucleus set S _N2 to obtain the final cell nucleus set S _N .

Further, the training method of the cell nucleus detection model includes:

Download the MSCOCO data set D ₁ , and use the MSCOCO data set to train the target detection algorithm to obtain the target detection model M ₁ ;

Download the 2018DSB data set D ₂ , and use the 2018 DSB data set to train the target detection model M ₁ whose model parameters only retain the feature extraction layer to obtain the target detection model M ₂ ;

Download the open digital pathological image nucleus labeling dataset D ₃ , and normalize each fragment image in the dataset D ₃ , and then randomly perform pixel values of three channels in the normalized fragment image according to a preset ratio. reverse color;

After merging the data set obtained after the inverse color processing with the data set D ₃ , the target detection model M ₃ whose model parameters are retained only in the feature extraction layer is trained to obtain a cell nucleus detection model.

The beneficial effects of the present invention are: when detecting cell nuclei, the scheme performs two detections on each fragment image, that is, the first time is directly on the fragment image, and the second time is on the RGB three channels of the original image, Set the detected nuclei as the background color respectively; then perform the nuclei detection again; finally, combine the two nuclei detections as the final detection result; this detection method can greatly reduce the missed detection of nuclei in the image.

In addition, when training the cell nucleus detection model, this scheme makes full use of open data sets to provide a large amount of labeling information, and a lot of labeling work can be saved through transfer learning; in addition, the influence of equipment, environment and other objective factors on pathological images can be reduced , which greatly improves the accuracy and robustness of nuclear detection.

Description of drawings

Figure 1 is a flow chart of a method for automatic detection of cell nuclei in digital pathological images.

Detailed ways

The specific embodiments of the present invention are described below to facilitate those skilled in the art to understand the present invention, but it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes Such changes are obvious within the spirit and scope of the present invention as defined and determined by the appended claims, and all inventions and creations utilizing the inventive concept are within the scope of protection.

Referring to FIG. 1 , FIG. 1 shows a flowchart of a method for automatic detection of cell nuclei in digital pathological images; as shown in FIG. 1 , the method S includes steps S1 to S6 .

In step S1, a fragmented image of the digital pathological image is acquired and subjected to standardization processing; wherein the method adopted for the standardization processing is the Vahadane method.

In step S2, the standardized fragmented images are input into the trained cell nucleus detection model for detection, and the cell nucleus set S _N1 is obtained;

In one embodiment of the present invention, the training method of the cell nucleus detection model includes:

A1. Download the MSCOCO data set D ₁ , and use the MSCOCO data set to train the target detection algorithm to obtain a target detection model M ₁ ; the preferred target detection algorithm in this scheme is the Mask-RCNN algorithm.

A2. Download the 2018DSB data set D ₂ , and use the 2018 DSB data set to train the target detection model M ₁ whose model parameters only retain the feature extraction layer to obtain the target detection model M ₂ ;

A3. Download the open digital pathological image cell nucleus labeling dataset D ₃ , and standardize each fragment image in the dataset D ₃ , and then randomly select the pixel values of the three channels in the standardized fragment image according to the preset value. Invert the scale;

The inverse color processing is to use 255 minus the original pixel value to update the original pixel value of the three channels.

A4. After merging the data set obtained after the inverse color processing with the data set _D3 , train the target detection model M3 _whose model parameters only retain the feature extraction layer to obtain a cell nucleus detection model.

This scheme uses the above method to train the nucleus detection model, makes full use of the large amount of annotation information provided by the open data set, and saves a lot of annotation work through transfer learning. After augmenting the dataset with a small number of digital pathological images, and then training the target detection model, the detection accuracy of cell nuclei can be greatly improved.

During implementation, the preferred training method of the cell nucleus detection model in this scheme further includes the following steps:

B1. Acquire multiple digital pathological images stored by the user, and read the ultra-high-definition images at the bottom of the digital pathological images, and then extract the fragmented images of the ultra-high-definition images in an overlapping or non-overlapping manner to form a fragment collection.

When implementing, this scheme can use OpenSlide pathological image reading software to read the bottom image; when extracting fragments, according to the image pixel position coordinates, directly read the fragment image T _i ; can be overlapped and read, then the starting coordinates of T _i Random, non-overlapping reads, then Ti reads in steps of _n .

B2, using the image gradient energy function to calculate the energy value of each fragment image, and selecting fragment images whose energy value is greater than the set threshold to form an image set ST ₁ ;

B3, standardize each fragment image in the image set ST ₁ , and then randomly invert the pixel values of the three channels in the normalized fragment image according to a preset ratio to obtain a new data set ST ₂ ;

B4. After merging the image set ST ₁ and the data set ST ₂ , train the cell nucleus detection model whose model parameters only retain the feature extraction layer to obtain the final cell nucleus detection model.

During model training, this scheme can greatly improve the accuracy of detection by introducing a small amount of labeled data in practice, reduce the impact of equipment, environment and other objective factors on pathological images, and greatly improve the accuracy and robustness of nuclear detection. .

In step S3, the cell nucleus detected in the fragment image is covered with a background color, and the background color covering is to modify all the pixel values corresponding to the position of the cell nucleus detected in the fragment image to 255;

After that, the energy value of each fragment image is calculated using the image gradient energy function:

Among them, T _i is the ith fragment image in the dataset; g(T _i ) is the energy value of T _i ; x is the pixel value in the x direction of the image; y is the pixel value in the y direction of the image.

In step S4, it is judged whether the energy value of each fragment image is greater than the preset threshold, if so, go to step S5, otherwise the corresponding fragment image is discarded;

In step S5, input all fragment images with energy values greater than a preset threshold into the trained cell nucleus detection model for detection, to obtain a cell nucleus set S _N2 ;

In step S6, the set of nuclei S _N1 and the set of nuclei _SN2 are combined to obtain the final set of nuclei _SN .

To sum up, the detection of nuclei using this scheme can reduce the missed detection of nuclei, and at the same time can greatly improve the accuracy and robustness of nuclei detection.

Claims (7)

1. The automatic detection method of the cell nucleus of the digital pathological image is characterized by comprising the following steps:

s1, acquiring a fragment image of the digital pathological image, and carrying out standardization processing on the fragment image;

s2, inputting the normalized fragment images into a trained cell nucleus detection model for detection to obtain a cell nucleus set S_N1；

S3, covering the cell nucleuses detected in the fragment images with background colors, and calculating the energy value of each fragment image by using an image gradient energy function;

s4, judging whether the energy value of each fragment image is larger than a preset threshold value, if so, entering a step S5, and otherwise, discarding the corresponding fragment image;

s5, inputting all fragment images with energy values larger than a preset threshold value into a trained cell nucleus detection model for detection to obtain a cell nucleus set S_N2；

S6 merging cell nucleus set S_N1And cell nucleus set S_N2Obtaining the final cell nucleus set S_N。

2. The method for automatically detecting the cell nucleus of the digital pathological image according to claim 1, wherein the training method of the cell nucleus detection model comprises:

downloading MSCOCO data set D₁Training a target detection algorithm by adopting an MSCOCO data set to obtain a target detection model M₁；

Downloading 2018DSB data set D₂And adopting 2018DSB data set to only reserve the target detection model M of the feature extraction layer for the model parameters₁Training to obtain a target detection model M₂；

Downloading open nuclear annotation data set D of digital pathology images₃And to the data set D₃Normalizing each fragment image, and then randomly reversing the pixel values of three channels in the normalized fragment images according to a preset proportion;

the data set obtained after the reverse color processing and the data set D are processed₃Target detection model M only retaining feature extraction layer for model parameters after combination₃And training to obtain a cell nucleus detection model.

3. The method for automatically detecting the cell nucleus of the digital pathological image according to claim 2, further comprising:

acquiring a plurality of digital pathological images stored by a user, reading the ultra-high-definition image at the bottommost layer of the digital pathological images, and extracting fragment images of the ultra-high-definition image in an overlapping or non-overlapping mode to form a fragment set;

calculating the energy value of each fragment image by adopting an image gradient energy function, and selecting the fragment images with the energy values larger than a set threshold value to form an image set ST₁；

For image set ST₁Normalizing each fragment image, and then randomly reversing the pixel values of three channels in the normalized fragment images according to a preset proportion to obtain a new data set ST₂；

Set of images ST₁And a data set ST₂And training the cell nucleus detection model with model parameters only reserved in the feature extraction layer after combination to obtain the final cell nucleus detection model.

4. The method for automatically detecting the cell nucleus of the digital pathological image according to claim 3, wherein the energy function is calculated by the formula:

wherein, T_iThe ith fragment image in the data set; g (T)_i) Is T_iThe energy value of (a); x is the pixel value in the x direction of the image; y is the image y-direction pixel value.

5. The method for automatically detecting the cell nucleus of the digital pathological image according to claim 2, wherein the normalization process is performed by a Vahadane method; the reverse color processing is to adopt 255 to reduce the original pixel value to update the original pixel value of three channels.

6. The method as claimed in claim 1, wherein the background color is overlaid to modify all pixel values corresponding to the detected cell nucleus positions in the fragment image to 255.

7. The method for automatically detecting the cell nucleus of a digital pathological image according to any one of claims 2-5, wherein the target detection algorithm is Mask-RCNN algorithm.

Images