Sperm morphology is an important technique in identifying the health of sperms. In this paper we present a new system and novel approaches to classify different kinds of sperm images in order to assess their health. Our approach mainly relies on a onedimensional feature which is extracted from the sperm’s contour with gray level information. Our approach can handle rotation and scaling of the image. Moreover, it is fused with SVM classification to improve its accuracy. In our evaluation, our method has better performance than the existing approaches to sperm classification.
With the development of modern computer technology, medical imaging has played an important role in clinical diagnosis and treatment. Medical image analysers are facing the challenge of precisely extracting information from the medical image with the help of computerassisted systems. Since people have become more and more concerned about the health of the next generation, morphology would be one important technique to identify the health of sperms. To examine whether or not the sperms are healthy, it is essential to inspect the sperms to assess their appearance. Currently, sperm quality is mostly judged by experts and doctors. Because of the numerous types of sperm shape, the efficiency and accuracy relying on human assessment are not ideal. As computer morphology technology develops, quantitative analysis of sperm morphology is demanded to assist doctors in their diagnoses. Thus, this research is intended to design a helpful sperm classification system.
Sperm morphology is an image classification problem in sperm imaging. It first detects a segment of the sperm image, after which feature extraction and analysis is possible, for example, sperm length, width, and size, followed by further classification according to sperm features [
Our sperm morphology system is equipped with a microscope connected to a computer to observe the realtime sperm image. The microscope helped us to take photos of the sperm images and input them into our computer. With the input we managed to obtain all the results and conclusions. The system and its equipment are shown in Figure
The sperm morphology diagnosis system with a microscope.
In addition to system implementation, this research has made the following contributions.
We proposed two approaches to transform the sperm contour into a onedimensional waveform as an analysis feature. The first algorithm takes advantage of the distance between two points on the edge to produce a waveform. The second computes the distance from the geometric centre to the edge as the vertical value of the waveform.
After extraction, we proposed an SVM classification on these waveforms with rank and grey level features. According to our comprehensive survey, this has not yet been used in sperm classification.
We also conducted a complete comparison. We compared our approaches with the Knearest neighbour, ScaleInvariant Feature Transform (SIFT), and the elliptic model. The experiment results show better performance than previous methods.
In our evaluation, we applied our approach to a sperm database. The results show that our idea is feasible and gives better performance than the existing approaches.
The rest of this paper comprises four parts. Section
For this research on sperm morphology, we reviewed the related works on segmentation, extraction, shape descriptor, and the classification algorithm, as shown in Figure
Related work.
First, segmentation takes place, so we have to look at several pieces of the literature [
Considering that there is no guarantee that the sperms we observed will appear with the posture and position we need, it is absolutely necessary for us to investigate how to deal with active contours. In research [
To obtain a better result of image segmentation, paper [
Furthermore, in paper [
The next topic is extraction. Because the results of the spermhead contour extraction have an essential influence on the classification, we studied some issues which provide further information. As it is of great importance to obtain the spermhead contour precisely, we studied articles on how to abstract contours. The first was “Sperm morphology assessment using David’s classification” [
After the DC analysis, we reviewed the paper on the technology of extracting objects’ edges. The first article [
As the snake may not do its job well enough for current research, we reviewed some improved algorithms such as Tsnakes [
Moreover, other authors [
The third topic, the most related work, is called shape descriptor. We focused on how to transform it into a onedimensional feature. To achieve the goal we studied further related articles. The first, paper [
For additional study of the shape descriptor, paper [
Another common image matching approach, the Knearest neighbour method as a compared target, proposes a method to fuse realvalue Knearest neighbour classifiers by feature grouping [
As we have been using the SVM as an advanced method to improve the performance of our approach, in order to achieve a deep understanding of SVM, we paid attention to related SVM research. Research work [
Research [
Once we had finished reading about the sperm related SVM, we moved on to research [
We then tried to look for research on feature extraction and the SVM classifier [
One popular robust imagematching approach is SIFT. In paper [
With regard to the elliptic model, it tried to estimate the contour of a sperm by an ellipse shape. Research work [
Although some researchers focused on sperm classification, none has used our approach for sperm imaging. This paper presents a system and a novel approach which uses a onedimensional contour and gray level features to diagnose different sperms according to their characteristics.
In this section, we present the algorithm of our work. First of all, we provide the flow of our approach as Figure
The flow chart of the sperm classification system.
For the classification, we applied two methods. In the first, the bilateral symmetrical function of
The symmetry function is proposed as follows:
In our second method, the geometric centre is calculated. We used the following equation to compute the horizontal coordinate:
To obtain the distance from edge to the centre, we take advantage of the following format:
Considering the difficulty in choosing the starting point, we had to avoid the problem. Therefore, we took rank algorithm [
Therefore, we achieved a binary system with 1 and 0. We transformed 5 consecutive numbers in
By calculating the
Using the rank algorithm [
To avoid the problem of a single criterion being too lopsided, we took the gray level value of the sperm into consideration. First, we calculated the gray level value of all points within the sperm using the following format:
Then we computed the gray percentage using the following format:
By calculating the average rank difference of normal and abnormal sperms, we found the dividing line between them. Thus, the rank difference itself can work as a judgment as to whether or not the test sperm is normal. We combined the two rank differences originating from the distance from centre to the contour and the grey level value of the pixels in a sperm by calculating the sum of the test sperm’s distance average rank difference and the average grey level value rank difference while each of them takes a certain weight. The format is as follows:
With the combination value we enhanced the original judgment by considering more elements and the importance of each. To find the dividing line, we collected all the combination values of the sperms and chose one of their average values to provide the best accuracy for the dividing line.
To achieve a better result, we fused the SVM method as an advanced classifier. SVM Methods are supervised learning models with associated learning algorithms that analyse data and recognize patterns and are used for classification and regression analysis. The basic SVM takes a set of input sperm features and predicts, for each given input, the possible class form, normal or abnormal, making it a nonprobabilistic binary linear classifier.
Given a training dataset
The SVM requires the optimal solution. We use the LIBSVM [
As for the input of the SVM approach, we fused the grey level value and
Grey level value extraction path.
43  42  41  40  39  38  37  64 
44  21  20  19  18  17  36  63 
45  22  7  6  5  16  35  62 
46  23  8  1  4  15  34  61 
47  24  9  2  3  14  33  60 
48  25  10  11  12  13  32  59 
49  26  27  28  29  30  31  58 
50  51  52  53  54  55  56  57 
When the path collides with the contour, the sequence of grey level values takes advantage of the Rank algorithm in transforming 8 consecutive numbers
With the rank sequences, we added the
In this fundamental evaluation, we have undertaken various experiments on 80 normal and 80 abnormal sperms. They are from a hospital; the images are provided on
Figures
Extract sperm head from normal sperm.
Extract sperm head from abnormal sperm.
In this way, we segmented all the 80 normal and 80 abnormal sperm heads from 80 pieces of sperm images.
In this section, we present part of the results of the steps in our experiments which explain how we made a choice and their performances. In Table
Waveform extraction from both methods.
Num  Normal  Symmetry  Mid 
 
1 



2 



3 



4 



5 



6 



7 



8 



9 



10 



 
Num  Abnormal  Symmetry  Mid 
 
81 



82 



83 



84 



85 



86 



87 



88 



89 



90 



Table
The waveform of different starting points from the first method.
Start 
Result  

Start 1 




 
Start 2 




In Figure
Normal and abnormal sperms and their corresponding data.
Normal sperms
Abnormal sperms
Then we focused on the influence of the grey level value. Parts of the results are as shown in Figures
Grey level value and percentage of normal sperms.
Grey level value and percentage of abnormal sperms.
In Figures
According to the percentage results, we can ascertain the percentage dividing line between normal and abnormal sperms to be 0.268228785. With the division we can achieve 72.5% accuracy.
We then tried to combine the two aspects as a judgment. First of all, we found the new parameter through format (
The result of joint rank and gray level features of normal and abnormal sperms.
In Table
Dividing line of C.

0.1  0.2  0.3  0.4  0.5  0.6  0.7  0.8  0.9 
DL  177  165  160  149  134  123  114  101  95 
Accuracy of different

0.1  0.2  0.3  0.4  0.5  0.6  0.7  0.8  0.9 
Acc  0.61  0.62  0.64  0.66  0.69  0.72  0.74  0.77  0.73 
In Table
With all the data provided above, we took advantage of the SVM method to undertake the classification and used LIBSVM Software from
We input the training samples and test samples as mentioned in Section
In this section, we introduce the three methods we used as comparisons to our method: Knearest neighbor method works quite effective as well, the other two fail to satisfy us.
The Knearest neighbour method showed the following differences.
We took the 160 images of sperms as the original classifications, and then we input the test image, transformed all the images into gray, and then calculated the difference between the test image and the original images. Then we picked the ten images with the smaller differences. After counting the classifications of the ten images, we specified the one occurring most often to be the classification of the test image. The classification of the top ten nearest neighbours is as in Figure
The classification of the top ten neighbours of (a) normal and (b) abnormal sperms.
Normal sperms
Abnormal sperms
Taking the sperms whose sum value is larger than 3 as normal and the sperms whose sum value is smaller than 4 as abnormal, we can tell that only 39 sperms are mistakenly classified, providing accuracy of 75.625%.
Lowe summed up the existing feature detection method based on invariants technology, in 2004, and formally proposed an image scaling, rotation, and even affine transformation for invariant image with local feature description operator based on scale space SIFT [
In format (
The SIFT feature vector has the following features: (a) it is the local feature of an image which maintains invariance not only on rotation, scale, and brightness variation but also on the viewing angle, the affine transformation, and the noise; (b) it is distinctive and informative and suitable for fast, accurate matching in a mass signature database; (c) it can produce a large number of SIFT feature vectors with few objects; (d) it is of high speed.
Matlab source code of SIFT is from
With the image of a single sperm, we can use the ellipse model to estimate the contour of the sperm. First of all, we used a rectangular segment to cut out the sperm. In order to make the rectangle close to the sperm edge, we needed to choose the rectangle with the smallest area; we rotated the sperm so that the segmentation would be easier. We calculated the leftmost, uppermost, downwardmost, and rightmost points on the edges and built the rectangular segment based on them as depicted in Figure
The ellipse model to estimate the contour of the sperm.
With the rectangle we achieved an ellipse to estimate the sperm whose long axis radius is half the length of the rectangle’s long side and whose short axis radius is half the length of the rectangle’s short side. As a result, we ascertained the ovality, which is the ratio of the length of the long axis to the short axis. Through our experiment on 80 sperms, we could tell whether the sperm was too close to a circle, whose ratio is close to 1, and whether the sperm was too slim, whose ratio is close to 2 or even larger than 2. We took those whose ratio was larger than 1.2 and smaller than 1.8 as normal, and thus we achieved accuracy of 66.25%.
As shown in Figure
Accuracy comparison of all methods.