Automatic segmentation algorithm for breast cell image based on multi-scale CNN and CSS corner detection

Abstract

As the shapes of breast cell are diverse and there is adherent between cells, fast and accurate segmentation for breast cell remains a challenging task. In this paper, an automatic segmentation algorithm for breast cell image is proposed, which focuses on the segmentation of adherent cells. First of all, breast cell image enhancement is carried out by the staining regularization. Then, the cells and background are separated by Multi-scale Convolutional Neural Network (CNN) to obtain the initial segmentation results. Finally, the Curvature Scale Space (CSS) corner detection is used to segment adherent cells. Experimental results show that the proposed algorithm can achieve 93.01% accuracy, 93.93% sensitivity and 95.69% specificity. Compared with other segmentation algorithms of breast cell, the proposed algorithm can not only solve the difficulty of segmenting adherent cells, but also improve the segmentation accuracy of adherent cells.

Keywords

Breast cell image segmentation staining regularization Multi-scale CNN CSS corner detection

1. Introduction

Nowadays, breast cancer (BC) is the most prevalent type of cancer in women [1]. In recent times, there has been lot of advancement for analyzing of BC progression and molecular markers related discovery, but the histopathological analysis still remains the most widely used for BC diagnosis [2]. The traditional method of manual pathological analysis has strong subjectivity. The diagnostic results depend on the doctor’s visual recognition and experience, which can easily lead to misdiagnosis, so the accuracy of diagnosis is very unstable.

Figure 1.

Total flow chart of the algorithm.

Accurate segmentation for breast cell is the important step towards pathological analysis of microscopic images. The breast cell image has the characteristics of different shapes, staining technique variations and adherent cells, which directly brings difficulty to accurate segmentation. At present, the existing cell segmentation algorithms are not ideal for the segmentation of the cell image with indistinct contour and low color contrast [3, 4]. Especially how to extract cell region from complex background and segmentation of adherent cells, it has always been the focus and difficult point of the current.

In order to improve the accuracy of segmentation, the use of Computer-Aided Diagnosis (CAD) system to help analyze image of breast histological sections is the current development trend [5]. In recent years, many breast cell image segmentation algorithms mainly include threshold algorithm (OTSU) [6], Watershed segmentation algorithm (WATERSHED) [7] and Fuzzy C-means clustering algorithm (FCM) [8], etc. OTSU is simple to implement and the performance is stable, but does not consider the spatial characteristics of the image and is sensitive to noise. WATERSHED is the most widely used algorithm, which has a good response to cell weak edge, but is prone to over-segmentation problems. FCM uses the degree of clustering of each data point to segment the cell, but it is sensitive to the initialization and difficult to converge to the cluster center.

From these studies, it is found that the main reason for the low accuracy of existing algorithms is that they are sensitive to noise and are prone to over-segmentation, which results in inaccurate acquisition dividing lines of adherent cells. In recent years, Convolutional Neural Networks (CNN) [9] have been widely used in various fields of medical image analysis, such as cell counting [10], cell segmentation [11], and cell detection [12]. Compared with the traditional algorithms, CNN can automatically extract the deeper features of the image through multiple nonlinear transformations, and enhance the robustness of feature extraction by local connection and weight sharing [13].

In this study, combining the Multi-scale CNN with the CSS corner detection, this paper puts forward a full automatic segmentation algorithm for breast cell image. The total flow chart of the algorithm is shown in Fig. 1. The staining regularization is used to enhance the cell image before cell segmentation. The following two modules are used for cell segmentation of the enhanced cell image. The first module adopts Multi-scale CNN to extract cell region from the complex background of the image to complete the initial segmentation of cells. The second module adopts CSS corner detection to segment the adherent cells after initial segmentation. The rest of this paper is organized as follows. An overview of cell image staining regularization is given in Section 2. The initial segmentation based on Multi-scale CNN and the segmentation refinement based on CSS corner detection are described in detail in Sections 3 and 4, respectively. Section 5 shows the experimental results. The paper ends with our conclusions in Section 6.

Figure 2.

Cell image staining regularization.

Figure 3.

Multi-Scale CNN model.

2. Cell image staining regularization

When making histological cell sections, the slice image produced by different methods or different conditions have greater heterogeneity, which makes it difficult to make quantitative analysis on slice image. Therefore, this paper uses a staining regularization algorithm proposed by Ruifrok, which is used for enhancement of Hematoxylin and Eosin (H&E) stained breast cell image [14]. By staining regularization, the slice image can be reconstructed separately according to the stained color, thereby promoting the quantitative analysis of the slice image.

To perform the separation of the stains, we have to do an orthogonal transformation of the RGB information, to get independent information about each stain’s contribution. The transformation has to be normalized to achieve correct balancing of the absorption factor for each stain. For normalization, we divide each Optical Density (OD) vector by its total length, resulting in a normalized OD matrix. The literature [14] defines a normalized OD matrix stained with hematoxylin, eosin and DAB. In this paper, H&E stained breast cell image is processed, and the zero components are regularized. The normalized OD matrix $M$ for the above combination of H&E is:

$\displaystyle M=\left({{\begin{array}[]{ccc}{0.6381}&{0.7143}&{0.2892}\\ {0.0919}&{0.9897}&{0.1154}\\ {0.3841}&{0.5617}&{0.7751}\\ \end{array}}}\right)$ (1)

Where, the first row represents hematoxylin stain, the second row represents eosin stain, and the third row is empty. Every column represents the optical density as detected by the red, green and blue channel for each stain.

If $U$ is the 3 by 1 vector for amounts of the three stains at a particular pixel, then the vector of OD levels detected at that pixel is $q=\textit{UM}$ . From the above it is clear that $D=M^{-1}$ , This means, that multiplication of the OD image with the inverse of the OD matrix, which we define as the color-deconvolution matrix $D$ , results in orthogonal representation of the stains forming the image.

$\displaystyle U=Dq$ (2)

The single stained image of H&E were separated by the above method. The separated single stained image are added to obtain a regularized image $G_{1}$ . Take the breast cell image of this experiment as an example, the result is shown in Fig. 2. The cell image used the algorithm of staining regularization has a higher contrast.

3. Initial segmentation for breast cell based on multi-scale CNN

Aiming at the diversity of breast cell shapes, this section uses convolution kernels of different scales to extract image features of different scales, so that the filter extracts and learns more abundant image information. In this work, we make full use of the CNN to learn discriminative features, and build three CNN in parallel to extract multi-scale features from cell image with different sizes. The detailed structure of the Multi-scale CNN model is shown in Fig. 3. Our model is mainly divided into three parts: multi-scale decomposition of cell image, multi-scale feature extraction, pixel classification labeling.

3.1 Multi-scale decomposition of cell image

In order to extract a set of more representative features in the image, the Laplacian pyramid is used to decompose the cell image into three scales of 256 $\times$ 256, 128 $\times$ 128, 64 $\times$ 64 pixels.

The staining regularized cell image $G_{1}(x,y)$ is taken as the bottom layer image of the Gaussian pyramid, which is first Gaussian low-pass filtered, and then down-sampled by interlaced permutation. The resolution of the obtained image is half of the bottom layer image, and the above operation is repeated to form a Gaussian pyramid. The $s$ layer image of Gaussian pyramid is $G_{s}$ :

$\displaystyle G_{s}(x,y)=\sum\limits_{m=-2}^{2}{\sum\limits_{n=-2}^{2}{w^{\ast% }}}(m,n)G_{s-1}(2x+m,2y+n)2\leqslant s\leqslant 3,0\leqslant x<R_{s},0% \leqslant y<C_{s}$ (3)

Where $(x,y)$ is the coordinate value of the pixel. The $w^{\ast}(m,n)$ is a 5 $\times$ 5 window function with low-pass property, the rows and columns of Gaussian pyramid $s$ layer are represented by $R_{s}$ and $C_{s}$ .

In this paper, $s=$ 1, 2, 3 and the decomposed images $G_{1},G_{2},G_{3}$ constitute the Gaussian pyramid of the cell image. We use the interpolation to expand the Gaussian pyramid. The image $G_{s}$ expands to $G_{s}^{\ast}$ , and its size is the same as the image $G_{s-1}$ , as follows:

$\displaystyle G_{s}^{\ast}(x,y)=4\sum\limits_{m=-2}^{2}{\sum\limits_{n=-2}^{2}% {w^{\ast}}}(m,n)G_{s}\left(\frac{x+m}{2},\frac{y+n}{2}\right)$ (4)

Then the Laplace Pyramid decomposition of cell image can be expressed as:

$\displaystyle\left\{\begin{array}[]{l}{\text{X}_{s}=G_{s}-G_{s}^{\ast},1% \leqslant s<3}\\ {\text{X}_{3}=G_{3},s=3}\\ \end{array}\right.$ (5)

Therefore, the original cell image with size 256 $\times$ 256 is decomposed into three different scales by Laplace pyramid. The size of the input cell image is 256 $\times$ 256, 128 $\times$ 128, 64 $\times$ 64 pixels.

3.2 Multi-scale feature extraction

The position of the cell remains unchanged on cell image of different scales. We use the convolutional neural networks of the same structure to extract features from cell image of different scales. The weight sharing is performed by the convolutional neural network of different scales, alleviating the over-fitting problem.

Each convolutional neural network $f_{s}$ ( $s=$ 1, 2, 3) is composed of three modules. The first module consists of 16 convolution kernels with a size of 3 $\times$ 3 and a step of 1, a nonlinear function $\tanh$ , a pooling layer with a size of 4 $\times$ 4 and a step of 2. The second module structure is similar to the first module structure, but it uses 64 convolution kernels with a size of 3 $\times$ 3 and a step of 1. The third module contain 256 convolution kernels with a size of 3 $\times$ 3 and a step of 1. For each scale, the convolutional neural network $f_{s}$ can be described as a sequence of linear transforms, interspersed with nonlinear symmetric squashing units (typically the $\tan$ h function [15]), and pooling/sub-sampling operators. For a network $f_{s}$ with $L$ layers, we have

$\displaystyle f_{s}(X_{s};\theta_{s})=W_{L}H_{L-1}$ (6)

$\theta_{s}$ is the parameter increment of $f_{s}$ , $W_{L}$ is Toeplitz matrices of the $L\text{th}$ layer, $H_{L}$ is the cell image feature vector output by the $L\text{th}$ layer. Where the vector of hidden units at module $l$ is

$\displaystyle H_{l}=\text{pool}(\text{tan}h(W_{l}H_{l-1}+b_{l}))$ (7)

for all $l\in$ (1, 2, $\ldots$ $L-1$ ), with $b_{l}$ a vector of bias parameters, and $H_{0}=X_{s}$ . The matrices $W_{l}$ is Toeplitz matrices.

Finally, the outputs of the 3 networks are up-sampled and concatenated so as to produce $F$ , a map of feature vectors of size 3 times the size of $f_{1}$ , the cell image feature vector is as follows:

$\displaystyle F=[f_{1},u(f_{2}).u(f_{3})]$ (8)

where $u$ is an up-sampling function.

3.3 Pixel classification labeling

Ideally a linear classifier should produce the correct categorization for all pixel locations ( $x, y$ ) from the feature vectors $F_{(x,y)}$ . We compute normalized predicted probability distributions each pixel of cell image using the softmax function, i.e.,

$\displaystyle\hat{c}_{(x,y),a}=\frac{e^{W^{T}_{a}F_{(x,y)}}}{\Sigma_{b\in% \textit{classes}}e^{W^{T}_{b}F_{(x,y)}}}$ (9)

Where $W$ is a temporary weight matrix only used to learn the features, $\hat{c}_{(x,y)}$ is the normalized prediction vector of pixel classifiers for pixel ( $x, y$ ). Then assign each pixel of cell image with the $\arg\max$ of the prediction at its location. More specifically, for each pixel ( $x, y$ ):

$\displaystyle l_{\left({x,y}\right)}=\arg\mathop{\max}\limits_{a\in\textit{% classes}}\hat{c}_{(x,y),a}$ (10)

The resulting labeling $l$ , although fairly accurate, is not satisfying visually as it lacks spatial consistency and precise extraction of cell edge.

4. Segmentation refinement for breast cell based on CSS corner detection

4.1 Judgment of adherent cells

After the cell region is initially positioned, it is necessary to judge the single cell and the adherent cells so that only the adherent cells are segmented. The specific steps for distinguishing single cell from adherent cells is as follows:

Step 1:
Calculate the circularity of each region $R=4\pi A/P^{2}$ , $A$ is the area of the region, $P$ is the perimeter of the region.
Step 2:
Preserve the area of circularity $R>t$ as the divided region $f^{\textit{out}}$ , and others as undivided region $f^{\textit{usg}}$ .
Step 3:
The average area of $f^{\textit{out}}$ is calculated as the initial area threshold Ag.
Step 4:
Calculate the area of each region in $f^{\textit{usg}}$ and compare it with the previous area threshold Ag. If it is $P$ times greater than the initial area threshold Ag, it will be further segmented as the adherent cells region $f^{\textit{usg}}$ . The other cell region will be output as $f^{\textit{out}}$ .

After the analysis of sample images and experimental results, the values of parameters $t$ and $P$ are 0.6 and 1.5 respectively.
4.2 Segmentation of adherent cells

After the single cell was separated from the adherent cells, the adherent cells were segmented by CSS corner detection. The CSS corner detection algorithm is more suitable for corner detection of multi-scale complex images. Its main principle is to determine the position of the corners by the change of curvature, and the corner positioning is very accurate. Hence, this section uses CSS corner detection to segment adherent cells, which can effectively overcome the influence of complex background to obtain accurate cell dividing line. The following steps are used by the CSS corner detection algorithm to detect corners of an image.

4.2.1 The definition of curvature

The curve $\Gamma$ is first parameterized by the arc length parameter $\mu$ :

$\displaystyle\Gamma(\mu)=(x(\mu),y(\mu))$ (11)

Where $x(\mu),y(\mu)$ is the coordinate value of the pixel. An evolved version $\Gamma_{\sigma}$ of $\Gamma$ can then be computed. $\Gamma_{\sigma}$ is defined by:

$\displaystyle\left\{{{\begin{array}[]{l}{\Gamma_{\sigma}=(X(\mu,\sigma),Y(\mu,% \sigma))}\\ {X(\mu,\sigma)=x(\mu)\otimes g(\mu,\sigma)}\\ {Y(\mu,\sigma)=y(\mu)\otimes g(\mu,\sigma)}\\ \end{array}}}\right.$ (12)

$\otimes$ is the convolution operation and $g(\mu,\sigma)$ denotes a Gaussian of width $\sigma$ .

The curvature $\kappa$ is based on the $\Gamma_{\sigma}$ :

$\displaystyle\kappa(\mu,\sigma)=$ (13) $\displaystyle\quad\frac{X_{\mu}(\mu,\sigma)Y_{\mu\mu}(\mu,\sigma)-X_{\mu\mu}(% \mu,\sigma)Y_{\mu}(\mu,\sigma)}{(X_{\mu}(\mu,\sigma)^{2}+Y_{\mu}(\mu,\sigma)^{% 2})^{1.5}}$

Where

$\displaystyle X_{\mu}(\mu,\sigma)=x(\mu)\otimes g_{\mu}(\mu,\sigma)$ $\displaystyle Y_{\mu}(\mu,\sigma)=y(\mu)\otimes g_{\mu}(\mu,\sigma)$ (14) $\displaystyle X_{\mu\mu}(\mu,\sigma)=x(\mu)\otimes g_{\mu\mu}(\mu,\sigma)$ $\displaystyle Y_{\mu\mu}(\mu,\sigma)=y(\mu)\otimes g_{\mu\mu}(\mu,\sigma)$

$g_{\mu}(\mu,\sigma)$ and $g_{\mu\mu}(\mu,\sigma)$ are the 1st and 2nd derivatives of $g(\mu,\sigma)$ respectively.

4.2.2 The CSS corner detector

The following steps are used by the CSS algorithm to detect corners of cell image:

Step 1:
Apply the Canny edge detector to the gray level image and obtain a binary edge-map.
Step 2:
Extract the edge contours from the edge-map, fill the gaps in the contours and find the T-junctions.
Step 3:
Compute curvature at a fixed low scale for each contour to retain all true comers.
Step 4:
All of the curvature local maxima are considered as corner candidates, including the false corners. By classifying the false corners into rounded and boundary noise [16], remove the false corners to get the correct corners.
Step 5:
Track the corners from the highest scale to the lowest scale to improve localization accuracy.
Step 6:
Compare the T-junction with other corners and remove one of the two corners which are very close.

Figure 4.
Workflow of this algorithm.

Among the corner candidates, although some corners are detected numerically as the local absolute maximum, its neighbors in the region of support are often very small. We can remove them by using an adaptive local curvature threshold. In principle, we set the threshold for a candidate according to its neighborhood region’s curvature, and the local curvature maximum smaller than its local threshold is eliminated. This adaptive threshold is given by:

$\displaystyle T(\mu,\sigma)=1.5\times\frac{1}{K1+K2+1}\sum\limits_{\lambda=\mu% -K2}^{\mu+K1}{\kappa(\lambda)}$ (15)

Where $C$ is a coefficient, $K1$ and $K2$ are sizes of the region of support. In general, a well-defined corner should have a relatively sharp angle, if $160^{\circ}\leqslant\angle C_{\lambda}\leqslant 200^{\circ}$ , $C_{\lambda}$ is the false corner, otherwise it is the correct corner.
4.2.3 Determine the best dividing line

After the corners are detected, the dividing line can be obtained by connecting the corners. The selection of the best dividing line is as follows:

Step 1:
The circularity of the cells before segmentation and the circularity and area of the two regions after segmentation were calculated, respectively.
Step 2:
If the circularity and area of both regions are greater than the threshold and the average circularity is greater than the circularity of the cells before segmentation, the corresponding dividing line is used as a candidate dividing line.
Step 3:
Select the shortest dividing line as the best dividing line.

In general, the use of CSS corner detection in cell image segmentation is a very good method, which can effectively detect corners and obtain more accurate cell dividing line.
5. Experimental results and analysis

5.1 Data set extension

Data set contains 58 H&E histopathology images of breast tissue from the Yale Tissue Microarray Facility. There are 26 malignant and 32 benign images respectively. Each image is represented in analog image as 8 bit with 896 $\times$ 768 binary ground truth images. In order to make the best use of limited data sets, the image size can be normalized by third-order difference transformation, and then the data set can be expanded by dividing, rotation, brightness, contrast conversion [17]. Through several methods, the original data set of 58 cell images can be increased to 44544 cell images with 256 $\times$ 256 pixel. Thus, the model will not see two identical images and can effectively suppress over-fitting.

Figure 5.

Segmentation result comparison of different algorithms.

5.2 Implementation details

During the experiment, this paper randomly selects 30000 images of the whole data set as the training samples, and the remaining 14544 images as test samples to evaluate the performance of the algorithm. The Multi-scale CNN model frame implementation is based on the open source software library TensorFlow 1.6 and is trained in end-to-end training on the NVIDIA GeForce GTX 1080Ti GPU. In the Multi-scale CNN structure, the nonlinear activation function of Rectifying Linear Units (ReLUs) and Max-pooling are used in the convolution layer and the pooling layer. The entire structure of the algorithm is shown in Fig. 4. First, the stain regularized breast cell image is input into a multi-scale CNN model for parameter training and model testing. However, it was found through the output results that although cell regions were accurately located, the adherent cells were still exist. The CSS corner detection is used to detect the cell corner to obtain the optimal dividing line, thereby realizing the segmentation of the adherent cells. In the whole segmentation algorithm, the proposed algorithm adaptively selects parameters and has good universality.

5.3 Cell segmentation results

In order to evaluate the effect of algorithm segmentation, sensitivity and specificity are selected as the evaluation indexes of various segmentation algorithms. The calculation formula is as follows:

$\displaystyle\textit{Accuracy}=\!\frac{N_{\text{seg}}}{N_{\text{seg}}\!+\!N_{% \text{spl}}\!+\!N_{\text{mer}}\!+\!N_{\text{add}}\!+\!N_{\text{mis}}}$ (16) $\displaystyle\textit{Sensitivity}=\frac{N_{\text{seg}}}{N_{\text{seg}}+N_{% \text{spl}}+N_{\text{mer}}+N_{\text{mis}}}$ (17) $\displaystyle\textit{Specificity}=1-\frac{N_{\text{spl}}+N_{\text{mer}}}{N_{% \text{seg}}\!+\!N_{\text{spl}}\!+\!N_{\text{mer}}\!+\!N_{\text{mis}}}$ (18)

Here, $N_{\text{seg}}$ is the number of cells correctly segmented, $N_{\text{spl}}$ and $N_{\text{mer}}$ are the number of cells over-segmentation and under-segmentation, $N_{\text{add}}$ is the number of background that are incorrectly segmented into cells, $N_{\text{mis}}$ is the number of cells that are incorrectly segmented into background.

In addition, in order to evaluate the effect of the algorithm, this paper selects some algorithms that are well performed in cell segmentation and compares the results of these algorithm. The segmentation result of each algorithm is shown in Fig. 5. Among them, the first line is a benign breast cell image, and the second line is a malignant breast cell image. The marker in the figure is the object to be compared.

As can be seen from the figure, for images with uniform cell size and simple background, OTSU and FCM have better segmentation effects, but cannot be effectively segmented for adherent cells. However, The proposed algorithm can effectively segment the adherent cells and is basically without error segmentation. As indicated by the red circle mark in the Fig. 5. For images with uneven gray scale and complex background, WATERSHED has more over-segmentation. However, the algorithm still has more accurate segmentation results. As indicated by the red rectangle mark in the figure. CNN uses the same data set as this paper, but the algorithm is not ideal for segmentation of adherent cells. As indicated by the green circle mark in the figure. The superiority of the proposed algorithm in this paper is that convolution kernels of different scales can extract features of different scales, making its location on the cell area more precise. Combined with the CSS corner detection algorithm for segmentation of adherent cells, it can be seen that the overall segmentation effect of the algorithm is better than other algorithms. The results of each algorithm are shown in Table 1.

Table 1

Segmentation results comparison of different methods

Method	$N_{\text{seg}}$	$N_{\text{spl}}$	$N_{\text{mer}}$	$N_{\text{add}}$	$N_{\text{mis}}$
OTSU	1124	2	122	36	16
WATERSHD	1187	107	71	18	6
FCM	1165	62	45	28	43
CNN	1192	99	34	24	16
Proposed	1265	36	22	13	24

Table 1 shows the quantitative statistical analysis of the segmentation results of different algorithms. It can be seen that in the segmentation result of FCM, $N_{\text{mer}}$ and $N_{\text{mis}}$ are higher, indicating that the method cannot effectively separate the adherent cells. The segmentation accuracy of WATERSHED is high, but there are some over-segmentation and under-segmentation. In the segmentation result of CNN, $N_{\text{mer}}$ is smaller and $N_{\text{spl}}$ is larger, indicating that the method can effectively suppress the under-segmentation, but the over-segmentation phenomenon is more serious. However, the algorithm in this paper can effectively avoid over-segmentation and under-segmentation, and has high segmentation accuracy. The segmentation accuracy of different algorithms are shown in Table 2.

Table 2

Comparison of segmentation accuracy

Method	Accuracy (%)	Sensitivity (%)	Specificity (%)
OTSU	86.48	88.92	90.18
WATERSHED	85.42	86.57	87.02
FCM	86.78	88.59	91.86
CNN	87.26	88.82	90.02
Proposed	93.01	93.91	95.69

It can be seen from Table 2 that this algorithm outperforms other algorithms in the three evaluations of Accuracy, Sensitivity and Specificity. The mean accuracy of segmentation is 93.01%. The sensitivity indicates the recall rate of cells. The proposed algorithm has a small amount of over-segmentation and under-segmentation, so the sensitivity is 93.91%. The specificity indicates the recall rate of the background. In this paper, the cell over-segmentation is slightly more than the under-segmentation, and there is a small amount of mis-segmentation, so the specificity is slightly higher than the sensitivity, which is 95.69%. However, the results show that compared with other algorithms, the proposed algorithm has obvious advantages in identifying adherent cells and inhibiting single cell over-segmentation.

6. Conclusion

In this paper, we propose an automatic segmentation algorithm for the breast cell image. The Multi-scale CNN and CSS corner detection methods are combined in the algorithm to make full use of image information, which can effectively overcome the influence of complex background and has high segmentation accuracy. The features of different scale images can be extracted by multi-scale CNN to improve the accuracy of cell localization and complete the initial segmentation of breast cell. The dividing line of adherent cells can be obtained more accurately by the CSS corner detection to improve the segmentation accuracy of adherent cells. The experimental results have demonstrated the proposed algorithm can effectively solve the problem of adherent cells segmentation, and the segmentation accuracy is 93.01%, which is significantly higher than other segmentation algorithms.

Footnotes

Acknowledgments

This work is supported by Shaanxi Provincial International Science and Technology Cooperation Program (No.2018kw-026), Xi’an Science and Technology Plan Program (No.201805040 YD18CG24).

References

Spanhol

F.A.

Oliveira

L.S.

Petitjean

and Heutte

, A dataset for breast cancer histopathological image classification, IEEE Transactions on Biomedical Engineering 63(7) (2016), 1455–1462.

Spanhol

F.A.

Oliveira

L.S.

Petitjean

and Heutte

, Breast Cancer Histopathological Image Classification using Convolutional Neural Networks, International Joint Conference on Neural Networks, IEEE, 2016, pp. 2560–2567.

Yang

H.H.

Zhao

L.L.

Pan

X.P.

et al., Overlapping cell segmentation based on level set and concave area detection, Journal of Beijing University of Posts & Telecommunications 39(6) (2016), 11–16.

Bílková

Soukup

and Kučera

, Cell Segmentation Using Level Set Methods with a New Variance Term, International Conference Image Analysis & Recognition, 2016, pp. 183–190.

Kowal

Filipczuk

Obuchowicz

Jozef

and Monczak

, Computer-aided diagnosis of breast cancer based on fine needle biopsy microscopic images, Computers in Biology and Medicine 43(10) (2013), 1563–1572.

Otsu

, A threshold selection method from gray-level histograms, IEEE Transactions on Systems Man & Cybernetics 9(1) (2007), 62–66.

Buggenthin

Marr

Schwarzfischer

Hoppe

P.S.

and Theis

F.J.

, An automatic method for robust and fast cell detection in bright field images from high-throughput microscopy, BMC Bioinformatics 14(1) (2013), 200–203.

Tang

J.R.

Isa

N.A.M.

and Chńg

E.S.

, A fuzzy C-means clustering approach: quantifying chromatin pattern of non-neoplastic cervical squamous cells, PLOS ONE 10(11) (2015), 1–15.

Hatipoglu

and Bilgin

, Classification of histopathological images using convolutional neural network, International Conference on Image Processing Theory, IEEE, 2015.

10.

Xie

Noble

J.A.

Zisserman

et al., Microscopy cell counting and detection with fully convolutional regression networks, Computer Methods in Biomechanics and Biomedical Engineering. Imaging & Visualization 6(3) (2018), 283–292.

11.

D.J.

Salama

et al., Nuclei detection and segmentation of fluorescence microscopy images using three dimensional convolutional neural networks, International Symposium on Biomedical Imaging, 2018, pp. 418–422.

12.

Xie

Xing

Kong

et al., Beyond Classification: Structured Regression for Robust Cell Detection Using Convolutional Neural Network, medical image computing and computer assisted intervention, 2015, pp. 358–365.

13.

C.Y.

B.S.

Zhang

Y.G.

et al., Retinal vascular image segmentation based on improved convolution neural network, Journal of Optics 38(11) (2018), 133–139.

14.

Ruifrok

A.C.

and Johnston

D.A.

, Quantification of histochemical staining by color deconvolution, Analytical and Quantitative Cytology and Histology 23(4) (2001), 291–299.

15.

LeCun

Bottou

Orr

and Muller

, Efficient Backprop[C], Neural Networks: Tricks of the Trade, Springer-Verlag, 1998.

16.

Rattarangsi

and Chin

R.T.

, Scale-based detection of corners of planar Curves, IEEE Trans on Pattern Analysis and Machine Intelligence 14(4) (1992), 430–449.

17.

Hashemi

Kiani

Noroozi

and Moghaddam

M.E.

, An image contrast enhancement method based on genetic algorithm, Pattern Recognition Letters 31(13) (2010), 1816–1824.