efficient CASA system is presented from the computer vision researcher point of view to ... which are converted in a digital bitmap format with BMP 8 bits for each ...
A SIMPLE AND EFFECTIVE SYSTEM FOR COMPUTERASSISTED SEMEN ANALYSIS J.M.Pascual-Gaspar1, H.Olmedo1, A.I.Exposito2, A. Exposito3, G.Bermudez1, and J.Finat1 1
MoBiVA Group, Lab 2.2, Univ. of Valladolid, Spain 2 Hospital Clinico Universitario, Valladolid, Spain 3 Hospital de Cruces, Bilbao, Spain
Keywords: Computer vision, WHO standards, CASA method.
Abstract During the last years some Computer-Assisted Semen Analysis (CASA) software have appeared to assist the laboratory andrologist labour, but the detailed method of these systems have been hidden to the scientific community because of commercial interests. In this work a simple and efficient CASA system is presented from the computer vision researcher point of view to amend the absence of technical papers in this field. The proposed system measures the concentration and motility of spermatozoa using simple but effective computer vision techniques and generates its results according to the WHO standards in Andrology.
1 Introduction The analysis of spermatozoa is an important topic for fertility issues and for assisted reproduction. Correlation with pregnancy rates shows that the evaluation of sperm motility by Computer Assisted Semen Analysis (CASA) has predictive value for fertility [3]. The most important aspects for male infertility concern to the concentration, size, shape and motility of spermatozoa. Visual tracking of several dozens of spermatozoa is a difficult task due to the limitations of the human visual perception and ocular fatigue. Ordinary samples can have between a small number of dozens and about a hundred of spermatozoa with different activity levels along short video sequences. Motility is measured often by simple ocular inspection or using a CASA system [9]. In this paper a simple CASA system is presented with a computer vision point of view. The system presented in this work provides information in the form of spermogram, i.e., a set of histograms (both static and kinematic) for measuring the potential fertility in terms of progressive motility. Besides static spermograms, a novel contribution of this work is the generation of kinematic spermograms including a mobile segmentation in terms of the recommendations of the World Health Organization (WHO) [7]. The structure of this work is: preprocessing tasks of the image samples to enhance visualization is described in the second section. Third and fourth sections describe the static and
kinematics segmentations approaches respectively. Fifth section presents the experiment results and finally, some conclusions and future work are sketched at last section of the paper.
2 Problem analysis and image processing The activity analysis of spermatic motility is far from being elementary. Some troubles concern to the filamentous character of each spermatozoid and the relatively high number of objects to be tracked along the video sequence. In this work a serial processing is performed with two main steps which are named as static segmentation and motion tracking. Static segmentation is focused on the identification and labelling of spermatozoa with respect to another bodies contained in the mucus system. After labelling, motion tracking is performed by applying a method of finite differences between two consecutive images along the video sequence. A general strategy for planning the image analysis is based on an iterative method applied to individual views and the tracking of spermatozoa which are common to several video frames. The iterative method follows a typical pipeline with three basic processes which are labelled as sampling, definition of the strategy for image processing, and the development of the application.
3 Experimental set-up The assisted computer analysis of video sequence starts with a small sample of seminal plasma. Each second has 25 views, which are converted in a digital bitmap format with BMP 8 bits for each channel of a typical colour scale. Views have been sampled from the original video sequence in AVI format, keeping the original rate of 25 images per second. Besides to spermatozoa, microscopic analysis displays other cells such as bacteria, leukocytes and epithelial cells, etc. In the static analysis stage, it is necessary to eliminate all of them for giving an account as accurate as possible. Thus, it is necessary to use additional geometric and radiometric information relative to the shape, by introducing thresholds for eccentricity and reinforcing the small differences of colour
in the mucus system. Main features concerning to the head shape of normal spermatozoa are circularity and eccentricity. We are not considering abnormalities such as spermatozoa with two heads or two tails. Relative to motility aspects, it is necessary to implement an automatic discrimination between the immobile spermatozoa and some objects lying in the mucus system supporting them. The main problem for the kinetic analysis concerns to the automated identification of tails, and their shape variation corresponding to the motion. In the generic case, further research concerning to the correlation between head shape and velocity is desirable, and it is currently under development. 2.1 Binarization and labelling The input is given by a greyscale image. The Fig.1 shows a subsample with 5 spermatozoa and a round cell, jointly with smaller black elements which arise from the preparation. Labelling process of spermatozoa is automatic and it follows from a left-right and top-down sweep out performed on the original image, and the introduction of range thresholds for the allowed size.
However, after the application of binary filtering, there persist some undesirable particles such as round cells (possibly leukocytes), air bubbles and spurious pixels. To eliminate them, we have used a morphologic discrimination filter. The performed adaptation uses a threshold for the allowed area measured in terms of the number of pixels occupied by the spermatozoid head. The extremal values for the allowed area occupied by spermatozoid head depends on the augmentation level (400£ in our case), and they are given by parameters which can be reconfigured by the user (in our case they are comprised between 7 and 35 pixels). So, after applying this filter, round cells and spurious information are deleted. A first information concerns to the computation of global area corresponding to spermatozoa heads (see Fig.3), which is correlated with the concentration of spermatozoa in the examined sample.
Fig. 3: Second stage of the binarization filter. Furthermore, it is possible to perform a count of true pixels for each spermatozoid head, to identify coordinates of centroid (see Fig.4), to compute eccentricity of optimally adjusted elliptical shape and to estimate the area in terms of black pixels for each spermatozoid.
Fig. 1: Spermatic sample viewed at microscope (400x). A direct inspection of histogram displays a typical Gaussian distribution for the image pixels, with grey values in the interval [205; 255] for spermatozoa. The identification of extremal values allows the design and implementation of a filter for eliminating the background and irrelevant particles in the preparation. The binarization filter gives as output an identification of each head which are displayed in the Fig.2.
Fig. 4: Final spermatozoa centroids after binarization filter. As resume, an elementary morphological statistics of data relative to each sampled view (see Fig.5) provides the key data for mobile segmentation and tracking.
(a) Gray levels thresholds
(b) Binarized image
(c) Non-spermatic cell
Fig. 2: First stage of the binarization filter.
Fig. 5: Morphologic spermatozoa properties.
4 Algorithm for characterization of motility There are several possible approaches for characterizing the spermatic motility and activity of a given sample from a video sequence. In our work we adopt the classification criterium of the World Health Organization described in [7], which is the ”de facto” standard criterium in the Andrology laboratory. This classification criterium categorizes the spermatic motility according four types of individuals, labelling as follows: Type-A) Individuals with fast and progressive motility (at least 25 micrometers at 37 C or 20 micrometers at 20 centesimal degrees or at least 5 heads per second), TypeB) Slow progressive motility (between 5 and 25 micrometers, i.e., between 1 and 5 heads per second), Type-C). Nonprogressive motility (less than 5 micrometers, i.e. less than 1 head per second), Type-D) immobile individuals. To implement the particle tracking (kinematic segmentation) is needed to delimit the maximal displacement of spermatozoa over two consecutive frames. This displacement depends on the video images frequency (25fps in our case) resulting enough for this ratio to choose an upper bound of 5 micrometers for the covered distance per frame by each spermatozoid. From this bound decision, it is constructed a ‘tracking window’, where to locate each spermatozoid between frames allowing thus the motion tracking (Fig.6).
Fig. 6: Spermatozoid tracking window. Our strategy has been applied to four series of AVI videos with different lengths. The velocity of spermatozoa is bounded by five pixels per each 0.04 seconds, measured with respect to Euclidean distance in the plane of the sample. Collision phenomena are not considered in this work. The whole application has been coded in Java language with a modular architecture (Fig.7) in order to let any further module improvements or even module replacement without affecting the other components. As it can be seen, there exist three modules: a) Binary filter module, b) Particle analyzer module and c) Motility tracking module. The a) and b) modules implements the static segmentation, i.e. binarization and labelling and they are applied to individual images of the frame sequence. Module c) carries out the dynamic segmentation (motion tracking) and it takes as input a pair of consecutive images. Its function is to establish the labels correspondence between particles present in both frames. Also keeps track of new particles in scene, and saves the history of the disappeared ones for final spermogram statistics.
Fig. 7: Application algorithm schema and modules.
4 Experimental results The goal of this section is the description of software tools for processes concerning to classification and tracking of populations.
4.1 Static contents retrieval The implementation of search procedures in large datasets is easy when simple patterns can be identified from different kinds of histograms. The examination of the seminal samples for different types of populations displays some meaningful facts (Fig.8-a): • Invariance of the pixel histogram along the video sequence for each sample. • Concentration of basic statistical data (mean, standard deviation, mode) for different video samples. where all basic statistics are relative grey level intensity. From the above results we can infer the stability of the static properties of the video images, which is essential to the reliability of the binarization strategy. 4.2 Kinematic contents retrieval Labelling procedures of spermatozoid and bounding of search region for tracking allows identifying events which occur between two consecutive frames. Typical new events correspond to apparition/disapparition of spermatozoa which are treated as insertion/deletion phenomena in the application data structures. The velocity by time unit is obtained by taking a mid size for the head diameter of spermatozoa, and dividing by the distance covered by each mobile spermatozoid. The time unit is 0.04 seconds, because the analysis is now performed for every frame (25 each second), in order to avoid losses or errors in tracking. Inspection of experimental results (see Fig.8-b) provides some relevant consequences which can be read directly from the histogram of the images in gray scale. In the same way as above, we generate similar tables for kinematic histograms which display a reasonable concentration of velocities for the 4 populations of
spermatozoa. The kinematic results represent particle speed statistics measured as pixels per time.
(a) Static histogram
(b) Kinematic histogram
Fig. 8: Static and kinematic histograms of sample #1. 4.3 Evaluation and interpretation of results
Table.1: Motility statistics for the four video samples.
Table 2. Classification according the WHO motility standard. Table 1 shows the obtained statistics of the automatic process for statistics extraction in the examined samples which gives an error lesser than 2% (obtained with a per frame manual counting sequence assisted by the embryologist authors). Furthermore, it is possible to perform a semi-automatic evaluation of the most active populations of spermatozoa by types of populations as percentage as shown in table 2. In this case, type A corresponds to velocity v > 20:0, type B corresponds to 4:0 < b ≤ 20:0, type C to 0:4< v ≤ 4:0, and type D corresponds to v < 0:4. These simple motility statistics provides to the embryology laboratory stuff a quick view of the seminal sample of a patient, making their job agile, assisting them with the manual discrimination of velocity of the sperm samples and finally, but no less important, enforcing the way of doing their daily work in a more systematic and objective way.
5 Conclusions and future work In this work a CASA method system design have been presented for assisting the evaluation of spermatic motility in human fertility analysis. The design of algorithms for the automated kinetic analysis, and an efficient implementation allows to obtain results with
an error lesser than two-per-cent in examined samples of video sequences, with a clear identification of individuals and grouping in populations in four groups or categories of spermatozoa according the WHO standards. Future work concerns mainly to several aspects such as: a) A filter design able of preserving the tail which is lost in binarized images, b) the estimation of optimal shape characteristics for assisted reproduction following spermatic morphology criteria, c) the estimation of typical tail motions for each population, d) the correlation between variable shape and activity levels, e) the evaluation of possible pathologies including shape abnormalities or functionality defects.
Acknowledgements The authors acknowledge to the Embryology Service of the Hospital de Cruces (Bilbao, Spain) for the availability of samples which have been analyzed in this work. Additional acknowledgement is paid also to the MAPA Project for a partial financial support.
References [1] S. Alvarez and J.Finat: “Automated Shape Discrimination for Cells Population”, in Visualization, Imaging and Image processing, M.M.Hamza (ed.), ACTA Press, pp. 395-400, 2001. [2] I.Bankman and I.N.Bankman, eds.: “Handbook of Medical Imaging: Processing and Analysis”, Academic Press, 2000. [3] De Geyter ch, De Geyter M, Schneider HPG, Nieschlag E, Interdependent influence of follicular fluid oestradiol concentration and motility characteristics of spermatozoa on in-vitro fertilization results. Human Reproduction 7:664-670, 1992. [4] R.C..Gonzalez and R.E.Woods: “Digital Image Processing (2d ed.)”, Prentice-Hall, 2002. [5] W.J.Larsen: “Human Embryology (3rd ed)”, Elsevier, 2004. [6] Neuwinger J, Behre HM, Nieschlag E, External quality control in the andrology laboratory: an experimental multicenter trial. Fertil Steril, 54:308-314, 1990. [7] WHO Laboratory Manual for the examination of human semen and sperm - cervical mucus interaction, Cambridge University Press, 4th ed, 1999. [8] Patricia Olds-Clark, Harry M. Baer and Walter L. Gerber: “Human sperm motion analysis by automatic (Hamilton-Thorn motility analyzer) and manual (Image 80) digitization systems” in Journal of Andrology, Vol. 11, No 1, 1990. [9] Irvine DS. Computer assisted semen analysis systems: perm motility assessment. Human Reproduction. 1995; 10(SUPPL. 1): 53-59 [10] Lene Larsen et al.: “Computer-assisted semen analysis parameters as predictors for fertility of men from the general population” in Human Reproduction, 15, No. 7, 1562-1567, July 2000.