Effect of grape seed proanthocyanidins on activity of HaCaT cells in mice based on deep learning image processing

Abstract

BACKGROUND:

Grape seed proanthocyanidin extract (GSPE) has a certain resistance to contrast light, which makes the boundary of the imaging image unclear.

OBJECTIVE:

Because of this, an image processing algorithm is needed to process the contrast image to study the role of GSPE in the process of anti-ultraviolet.

METHODS:

In this paper, the fuzzy edges of contrast images were processed by deep learning algorithm, and the changes of VEGF and PEDF expression in HaCaT cells before and after UVA irradiation and after GSPE intervention were studied.

RESULTS:

The experiment results show that after processing, the edge and boundary of the image become clear and separable, which can be used to compare and analyze the test process. The image comparison results show that GSPE can down regulate the expression of VEGF gene and protein, and up regulate the expression of PEDF gene and protein. The synergistic effect of GSPE and GSPE can inhibit angiogenesis. It is confirmed that GSPE has the effect of anti-ultraviolet ray induced early angiogenesis.

Keywords

HaCaT cells SURF deep learning binary image

1. Introduction

Ultraviolet B (UVB) and long wave ultraviolet (UVA) are mainly affected by ultraviolet radiation on the surface of the earth. UVB is mainly related to burn and skin cancer, and UVA is closely related to skin photoaging. UVA is the main ultraviolet ray reaching the ground, and UVB accounts for a small part. UVB can not only penetrate into the basal layer of human skin, but also cause deep damage to the skin. This can cause changes in cytokine secretion and gene expression, resulting in excessive matrix degradation, resulting in connective tissue damage during photoaging [1, 2]. Most people’s skin types are mostly suntan, do not easily sunburn, and the incidence rate of skin cancer is very low [3]. The practical significance is to prevent and cure UVA.

Proanthocyanidin (PC) is a kind of polyphenol compounds widely existing in plants. As an ideal source of PC, grape seed proanthocyanidin extract (GSPE) got the most attention [4, 5]. A large number of experimental studies have confirmed that GSPE has very strong antioxidant and free radical scavenging activities, and is one of the strongest and most effective free radical scavengers found so far. In particular, its activity in vivo is incomparable with other antioxidants. It has even been reported that the antioxidant activity of GSPE in vivo is 20 times of VC and 50 times of VE [6]. As a natural and efficient antioxidant, GSPE has a variety of biological activities, such as inhibiting arteriosclerosis, lowering blood pressure, anti-tumor, anti-aging, anti-radiation and other pharmacological effects. Now it is favored by people in the fields of nutrition and health care, medicine and cosmetics. However, there are few reports about GSPE for external beauty and skin care to resist ultraviolet radiation. This is mainly because GSPE itself has a certain resistance to the contrast light, which causes the boundary of the imaging image unclear. Therefore, an image processing algorithm is needed to process the contrast image to study the role of GSPE in the process of anti-ultraviolet [7, 8, 9, 10].

2. Mathematical model of fuzzy edge detection and edge feature extraction based on deep learning

Based on the above analysis of cell image, in order to achieve the feature extraction of contrast image. It is convenient for the expert to research and evaluate for the analysis of contrast image, which needs to collect the spectral characteristics of cell image. The feature points of image information are extracted, and the edge information collection formula of cell image is as follows [11, 12, 13, 14]:

$\displaystyle{y}=\bar{y}+R_{i}{d}$ (1) $\displaystyle{z}=\bar{z}+R_{h}{d}$ (2)

In the above formula, $R_{t}$ represents the edge change pixel of green landscape structure, and $R_{h}$ is the change rate of style texture. Using the deep learning system evaluation method to design the spectral characteristic statistical model of cell image patches, the formula group of characteristic information evaluation coefficient is shown as follows:

$\displaystyle\psi_{{11}}=PA+A^{T}P+Q_{1}+R_{1}+R_{2}+K_{1}+K_{1}^{T}$ $\displaystyle\psi_{12}=W_{1}-K_{1}+M_{1}+K_{2}^{T}$ $\displaystyle\psi_{13}=PB+L_{1}+M_{1}+K_{3}^{T}$ (3) $\displaystyle\psi_{14}=-L_{1}+K_{4}^{T}$ $\displaystyle\psi_{15}=-W_{1}+K_{5}^{T}$ $\displaystyle\psi_{23}=L_{2}-M_{2}+W_{3}^{T}-K_{3}^{T}+M_{3}^{T}$ $\displaystyle\psi_{24}=-L_{2}+W_{4}^{T}-K_{4}^{T}+M_{4}^{T}$ $\displaystyle\psi_{25}=-W_{2}+W_{5}^{T}-K_{5}^{T}+M_{5}^{T}$ $\displaystyle\psi_{33}=-(1-\mu)Q_{2}+L_{3}-L_{3}^{T}-M_{3}-M_{3}^{T}$ $\displaystyle\psi_{34}=-L{}_{3}+L_{4}^{T}-M_{4}^{T}$ (4) $\displaystyle\psi_{35}=-W_{3}+L_{5}^{T}-M_{5}^{T}$ $\displaystyle\psi_{44}=-R_{1}-L_{4}-L_{4}^{T}$ $\displaystyle\psi_{45}=-W_{4}-L_{5}^{T}$ $\displaystyle\psi_{55}=-R_{2}-W_{5}-W_{5}^{T}$ $\displaystyle U=h_{1}Z_{1}+h_{2}Z_{2}+hZ_{3}$

In the establishment of the relevant edge detection information system, it is important to provide a powerful means of information acquisition. On the pixel correlation $y=y_{i+1}$ , there are two pixels PE1 and PE2 in the interval $({x}_{i,r}-\frac{{1}}{{2}},{x}_{{i},{r}}-\frac{{3}}{{2}})$ . According to the principle of pixel point association, when $({x}_{{i}+{1}},{y}_{{i}+{1}})$ is to the right of the midpoint $M(x_{i}+\frac{1}{2},y_{i+1})$ , the pixel point PE1 is taken, otherwise it is the pixel point PE2, color adjustment and correction. In the random field of urban landscape analysis (RF), it can be expressed as the product of two independent random fields, that is, $C=\{C_{i};i\in I\}$ in the partition of spectral characters. It can be expressed as:

$\displaystyle C=S\cdot\mu=\{S_{i}\cdot U_{i}:i\in I\}$ (5)

For any $i$ and $j$ ( $i$ is not equal to $j$ ), $D_{i}$ is independent of $C_{j}$ , and for a given $S_{i}$ , $C_{i}$ and $D_{i}$ are independent of $S_{j}$ . Because $C_{i}$ and $V_{i}$ obey the progressive rules of information extraction and extraction accuracy, the optimization equation of feature extraction of edge detection information is obtained as follows:

$\displaystyle I_{\textit{GSM}}=I(C^{N};D^{N}|{s^{N}})=\sum_{i=1}^{N}{I(C_{i};D% _{i}|{s_{i}})}$ (6)

In the above framework of detection analysis edge information feature extraction, because the larger feature information local patches obviously contain more spectral local information, can better describe the local characteristics of image information. But it brings the complexity of parameter estimation of local model. The pattern is planned to be 4 $\times$ 4 pixel blocks, and then the 4 $\times$ 4 sub blocks are further divided into 4 overlapping 3 $\times$ 3 sub blocks to obtain the gray scale information extraction area division map of image edge characteristics. Each pixel node generates a remote sensing texture number, which realizes the analysis of edge detection information extraction.

Through the analysis, it can be seen that through the partition treatment of edge information, the edge patches can keep a more complex edge as much as possible, which is conducive to play the edge effect and improve the image definition of fuzzy edge.

Speed up feature (SURF) is a fast robust feature extraction and registration algorithm based on SIFT (scale in variant feature transformation) algorithm. SIFT algorithm was proposed by Lowe in 1999. It is robust to image rotation, translation, zooming and denoising, but it takes a long time. Surf algorithm inherits the strong robustness of SIFT algorithm, and has high matching accuracy and fast speed. Therefore, surf algorithm is selected for cell image registration. The flowchart of the panorama registration algorithm is shown in Fig. 1.

Figure 1.

Algorithm flowchart.

In order to improve the matching efficiency and ensure the scale invariance of the algorithm, the image must be layered first. The surf algorithm filters the integral image of the original image through different size box filters, thus forming different scale space. Then the Hessian matrix is used to detect the image extreme points on each layer of image, and the threshold value of Hessian matrix is set to filter the extreme points. Finally, the location and scale value of the feature points are obtained.

In order to ensure the rotation invariance of the registration algorithm, the Haar wavelet response value of the pixel point in the X and Y direction is calculated in the circular region with the feature point as the center and the radius as 6. Then the response value is histogram counted, and the longest vector direction is selected as the main direction of the point in the whole circular region.

After the feature location and its corresponding scale value are obtained, the feature points need to be described; take the feature point as the center, select the 20” $\times$ 20” area within the neighborhood of the center point, and rotate its main direction to be consistent with the main direction of the feature point. The response values $d_{x}$ and $d_{y}$ of the Haar wavelet with Gaussian weighting ( $\sigma=$ 0.33”) in $x$ and $y$ directions are calculated. Sum the response values of all pixels in the region, which are expressed as $\sum{d_{x}}$ and $\sum{d_{y}}$ . In addition, the absolute values of $d_{x}$ and $d_{y}$ in each sub region are summed and expressed as $\sum{|{d_{x}}|}$ and $\sum{|{d_{y}}|}$ . Establish four-dimensional eigenvector: $v=(\sum{d_{x},}\sum{d_{y},}\sum{|{d_{x}}|,}\sum{|{d_{y}}|})$ . In addition, in order to make the descriptor scale invariant, the description vector needs to be normalized.

After generating the surf feature vectors of two images, the Euclidean distance between the vectors is used as the basis for measuring the similarity of feature points. If the ratio of the nearest distance to the next nearest distance of two feature points is less than 60%, the nearest pair of feature points is considered as matching points. However, the points obtained by rough matching cannot fully meet the needs of panoramic image mosaic accuracy, so it is necessary to further eliminate the wrong matching points and refine the matching results.

Deep learning algorithm combined with SURF proposed in this paper is a data fitting algorithm with fault tolerance and high robustness. The basic idea is to first design the objective function, then estimate the initial value of the parameters in the function by repeatedly extracting the minimum set of points, and then divide all the data into “inner point” (inlier) and “outer point” (outlier) by using the initial value. Finally, recalculate according to all the inner points to get the parameters of the function. The feature matching can effectively exclude the outliers and realize the fine matching of image.

3. Experiment result and analysis

3.1 Experimental steps

Each sample is provided with 3 complex holes. The relative content of PEDF protein in HaCaT cells was detected by indirect immunocytochemistry and contrast technique.

1)
HaCaT cells were digested and blown into single cells, and then seeded in a 6-well plate with a pre placed sterile cover glass by 1 $\times$ 10 ${}^{4}$ ;
2)
After about 50–60% of the cells climbing on the slide, take out the cover glass, mark it, put the cell face up and put it into a clean 6-well plate;
3)
4% polymethylmethacrylate was fixed at room temperature for 20 min;
4)
Was PBS three times, each time 5 min;
5)
PBS containing 0.1% Triton X-100 was used for 15 min at room temperature;
6)
10% sheep serum was added and sealed at room temperature for 1 h;
7)
First antibody incubation: the serum was aspirated by filter paper and diluted with 50 $\mu$ L/tablet, 37 ${}^{\circ}$ C 30–60 min;
8)
The second antibody was incubated at 37 ${}^{\circ}$ C for 30 min in a wet box, washed with PBS for 3 times for 5 min, and then placed on the dried slide; the second antibody labeled with biotin was added to the slide for 30 min at 37 ${}^{\circ}$ C, and washed with PBS for 3 times for 5 min;
9)
Add 50 PL/tablet of diluted SP (streptomycin and ubiquitin peroxidase), incubate in 37 ${}^{\circ}$ C constant temperature oven for 30 min-1 h, wash 5 times with PBS and for 5 min;
10)
Color development: add 50 $\mu$ L of freshly prepared 0.05% DAB solution (add DAB quickly when adding DAB, observe the dyeing situation, pour out the dye solution when the specimen is dyed brown, rinse with distilled water, and stop staining);
11)
Re-dyeing, gradient alcohol dehydration, transparent, neutral gum seal;
12)
They were observed and photographed under an inverted microscope.

Results: the positive staining cell count method was used. Under 40 times light microscope, 10 non overlapping visual fields were randomly selected, and positive staining cells were counted manually or by machine. In order to exclude the difference of cell density, the results were expressed by percentage. There were 3 slices in each group, and then the comparison was made between groups.

All data were analyzed by SPSS17.0, which was in accordance with normal distribution and expressed as $\bar{x}\pm s$ . The mean of multiple samples was compared by one-way ANOVA, and the comparison between pairs was conducted by LSD method, and the significance test level was $a=$ 0.05.
3.2 Experimental results

The pictures of immunocytochemistry results of each group are shown in Figs 2–6.

Figure 2.

Original picture for blank control group.

Figure 3.

Original picture for pure illumination group.

Figure 4.

Original picture for illumination plus medicine group 1 (10 $\mu$ g/mL).

Figure 5.

Original picture for illumination plus medicine group 2 (50 $\mu$ g/mL).

Figure 6.

Original picture for illumination plus medicine group 3 (100 $\mu$ g/mL).

It can be seen from the images that no matter the blank control group, the simple light group or the light plus drug group, the image is fuzzy and the boundary is not clear, so it is impossible to analyze the test results.

Firstly, the image is binarized. The results are shown in Fig. 7.

Figure 7.

The binary image.

The method of deep learning and surf proposed in this paper is used to process the binary image, the experimental results with clear edge can be obtained, as shown in Figs 8–12.

Figure 8.

The blank control group after deep learning algorithm.

Figure 9.

The pure illumination group after deep learning algorithm.

Figure 10.

The illumination plus medicine group 1 (10 $\mu$ g/mL) after deep learning algorithm.

Figure 11.

The illumination plus medicine group 2 (50 $\mu$ g/mL) after deep learning algorithm.

Figure 12.

The illumination plus medicine group 3 (100 $\mu$ g/mL) after deep learning algorithm.

It can be seen from the results that after processing, the edge and boundary of the image become clear and separable, which can be used to compare and analyze the test process. After image comparison, the following conclusions can be obtained:

The brown granules were mainly concentrated in the cytoplasm, indicating that PEDF protein was mainly expressed in the cytoplasm of HaCaT cells;

Compared with the blank control group, the number of brown granules in the light group was significantly decreased ( $P<$ 0.05), indicating that light can reduce the expression of PEDF protein;

The staining of PEDF protein was deeper and brown granules were more in the treatment group than in the light group, indicating that the expression of PEDF protein was increased in a concentration dependent manner (except for drug group 1, $P<$ 0.05).

Compared with the blank control group, $P<$ 0.05 was statistically significant, indicating that although the drug increased the expression of PEDF protein after light exposure, it did not return to the normal level.

Comparison of relative expression (percentage) of PEDF protein in each group is shown in Table 1.

Table 1

Comparison of relative expression (percentage) of PEDF protein in each group

Group	$\bar{x}\pm s$	$F$ value	$P$ value
Blank control group	70.75 $\pm$ 3.30 ${}^{*}$	172.77	0.00
Simple illumination group	37.25 $\pm$ 1.5
Group 1	39.75 $\pm$ 0.96
Group 2	53.00 $\pm$ 1.83
Group 3	57.25 $\pm$ 3.10

Note: group 1 was treated with 10 $\mu$ g/mL, 50 $\mu$ g/mL with group 2 and 100 $\mu$ g/mL with group 3.

LSD method was used to analyze the difference between the two groups. The other groups were compared with the simple light group, except for the comparison with the drug group 1 ( ${}^{*}P<$ 0.05). The other groups compared with the blank control group $P<$ 0.05, with statistical significance. There was statistical significance between groups 1 and 2, 1 and 3.

3.3 Discussion

Immunocytochemistry is a new technique which combines the basic principles of immunology and cytochemistry. According to the characteristics of specific binding of antigen and antibody, it can detect the existence and distribution of some polypeptide, protein, membrane surface antigen and receptor in cells. There are many kinds of peptides and proteins with antigenicity. When a peptide or protein of human or animal is injected into another animal as an antigen, the specific antibody (immunoglobulin) corresponding to the antigen will be produced. After the antibody is put forward from the serum and combined with a certain marker, it becomes the labeled antibody. When the labeled antibody was incubated with the tissue section specimen, the antibody specifically bound to the corresponding antigen in the cell, and the binding site was shown by the marker, the distribution of the peptide or protein was observed under the microscope.

As can be seen in the above pictures, the intercellular space becomes larger, the overall staining of cells becomes lighter, and the number of brown cells is significantly reduced. The number of positive cells increased with the increase of GSPE concentration. From the point of view of protein, immunocytochemistry was used to prove that the content of PEDF factor in the cytoplasm of HaCaT cells was decreased after UVA irradiation. GSPE can increase the protein content of PEDF decreased after UVA irradiation. This is consistent with the results of RT-PCR. However, due to the different levels of gene and protein, the same drug concentration has different effects on them. This led to the fact that although 5 (WML) GSPE could restore its gene expression level, the GSPE with 50 $\mu$ g/mL or even 100 $\mu$ g/mL did not completely restore its protein expression level. The expression of PEDF protein was detected by immunocytochemistry because of the limitation of conditions. Only by observing and comparing the number of positive cells in each sample, the expression of PEDF protein could be reflected. If possible, Western blot hybridization and other methods can be used to study the expression changes at the protein level.

4. Conclusion

After deep learning and surf algorithm, the edge of cell image is clear. The effect of grape seed proanthocyanidins on the activity of HaCaT cells can be analyzed by this image. In this paper, HaCaT cells were used as target cells to study the early effects of UVA irradiation on skin angiogenesis. RT-PCR, enzyme-linked immunosorbent assay and immunocytochemistry were used to verify the effect of UVA and GSPE on the expression of VEGF and PEDF in HaCaT cells. It is further proved that UVA irradiation on HaCaT cells can increase VEGF and decrease PEDF expression, which can promote angiogenesis, and eventually lead to the increase of capillary in skin at the early stage of photoaging. Grape seed proanthocyanidins GSPE can down regulate the expression of VEGF gene and protein, and up regulate the expression of PEDF gene and protein. The synergistic effect of GSPE and GSPE can inhibit angiogenesis. It is confirmed that GSPE has the effect of anti-ultraviolet ray induced early angiogenesis.

Footnotes

Conflict of interest

None to report.

References

Al-Maweri

NAAS

Sabri

AQM

Mansoor

, et al. Metadata hiding for UAV video based on digital watermarking in DWT transform. Multimedia Tools and Applications, 2017, 76(15): 16239–16261.

Jwa

, Ozguner Tang

. Information-theoretic data registration for UAV-based sensing. IEEE Transactions on Intelligent Transportation Systems, 2008, 9(1): 5–15.

Corrado

. Southern promises: a snapshot of the microbiology research landscape in South America based on bibliometric data. FEMS Microbiology Letters (16): 16–21.

Deng

Song

, et al. Multi-UAV counter-game model based on uncertain information. Applied Mathematics and Computation, 2020, 366: 124–129.

Heng

Xiao

Chao

, et al. Cultivated land information extraction in UAV imagery based on deep convolutional neural network and transfer learning. Journal of Mountain Science, 2017, 14(4): 731–741.

Hirota

Hirohata

Mashima

, et al. Population structure of the large Japanese field mouse, apodemus speciosus (rodentia: muridae), in suburban landscape, based on mitochondrial d-loop sequences. Molecular Ecology, 2004, 13(11): 3275–3282.

Crampton

. A landscape zonation for the southern and central mackenzie river valley based on terrain permafrost characteristics. Canadian Journal of Earth Sciences, 1973, 10(12): 1843–1854.

Newman

Liu

. Perspective on BRCA1, Breast Disease 10 (1998), 3–10.

Xiangyu

Friedrich

Franz

, et al. Optimization of OpenStreetMap building footprints based on semantic information of oblique UAV images. Remote Sensing, 2018, 10(4): 624–630.

10.

Wang

Luo

, et al. UAV tracking based on saliency detection. Soft Computing, 2020: 1–14.

11.

Steinhardt

Volk

. Meso-scale landscape analysis based on landscape balance investigations: problems and hierarchical approaches for their resolution. Ecological Modelling, 2003, 168(3): 251–265.

12.

Gong

Ibrahim

, et al. Orchard identification using landform and landscape factors based on a spatial-temporal classification framework. International Journal of Remote Sensing, 2014, 35(5–6): 2118–2135.

13.

Wolf

Meyer

. Suburban scenario development based on multiple landscape assessments. Ecological Indicators, 2010, 10(1): 74–86.

14.

Wang

Zhang

. Automatic detection of wind turbine blade surface cracks based on UAV-taken images. IEEE Transactions on Industrial Electronics, 2017, 64(9): 7293–7303.