Classification of Bovine Reproductive Cycle Phase using Ultrasound-Detected Features

Transcription

1 Classification of Bovine Reproductive Cycle Phase using Ultrasound-Detected Features Idalia Maldonado-Castillo Department of Computer Science 1 Science Place Saskatoon, SK, CANADA, S7N C9 idm9@mail.usask.ca Mark G. Eramian Department of Computer Science 1 Science Place Saskatoon, SK, CANADA, S7N C9 eramian@cs.usask.ca Roger A. Pierson Obstetrics, Gynecology, and Reproductive Sciences 3 Hospital Drive Saskatoon, SK, CANADA, S7N W8 pierson@erato.usask.ca Jaswant Singh Department of Veterinary Biomedical Sciences Western College of Veterinary Medicine Campus Drive Saskatoon, SK, CANADA, S7N B jaswant.singh@usask.ca Gregg P. Adams Department of Veterinary Biomedical Sciences Western College of Veterinary Medicine Campus Drive Saskatoon, SK, CANADA, S7N B gregg.adams@usask.ca Abstract Studies of ovarian development in female mammals have shown a relationship between the day in the estrous cycle and the size of the main structures and physiological status of the ovary. This paper presents an algorithm for the automatic classification of bovine ovaries into temporal categories using information extracted from ultrasound images. The temporal classes corresponded roughly to the metestrus, diestrus, and proestrus phases of the bovine reproductive cycle. Features based on the sizes of ovarian structures formed the patterns on which the classification was performed. A Naïve Bayes classifier was able to correctly classify the stage of the estrous cycle for 8.3% of the test patterns. A decision tree classified % of the test patterns correctly. The decision tree inference algorithm used to build the classifier constructed a tree that used only two of the five available features indicating that they form a sufficiently rich set of features for robust classification. 1 Introduction The reproductive cycles of humans and large mammals can be monitored by imaging the ovaries using ultrasonography [1,, 3,, 7, 1, 13, 1, 17,, ]. The reproductive cycle, or estrous cycle, of domestic animals consists of four phases: metestrus, diestrus, proestrus and estrus. In nonpregnant cows, ovulation occurs at approximately or 1 day intervals and usually happens near the end of (or shortly after) the estrus phase. The ovarian structures visible in an ultrasound image provide clues to the animal s current reproductive phase. The objective of this study was to develop an imagebased classification system which can automatically determine an animal s current stage in its reproductive cycle based on a single day s ultrasonographic examination of the ovaries. The algorithm was developed and tested using images of bovine ovaries which were to be classified as being in either the early (metestrus), middle (diestrus), or late (proestrus) phase of the estrous cycle.

2 Corpus Luteum Subordinate Follicles Oocytes Subordinate Follicles Corpus Luteum Figure 1. Schematic representation of a mammalian ovary. The major structures are follicles, in which oocytes (eggs) develop, and the corpus luteum (CL). The largest follicle of a wave is termed the dominant follicle and the remaining follicles are termed subordinate. 1.1 Mammalian Ovarian Cycle The ovaries are part of the reproductive system in female mammals. The major ovarian structures are follicles and corpora lutea; these structures are illustrated schematically in Figure 1 and an ultrasound image showing examples of these structures is shown in Figure. Ovarian follicles are roughly spherical, fluid-filled structures which contain developing oocytes (eggs). The reproductive cycle culminates in the rupture of a follicle and release of its egg. The corpus lutem (CL) is a gland that is formed from the remains of the ruptured follicle following ovulation. Ovarian follicular development is a dynamic process which occurs in a wave-like pattern [1,, 3, 13, 1]. A group of follicles begin growing simultaneously as a cohort at a diameter of to mm (wave emergence). The group of follicles continues growing for to 3 days. At this time, all follicles in this cohort, save one, begin to regress and degenerate while the dominant follicle continues preferential development. It has been shown that cattle and women may have either two or three waves of follicular activity per cycle. In both - and 3-wave growth patterns, the dominant follicle of the final wave ovulates, while dominant follicles of earlier waves ultimately regress and degenerate in a process known as atresia. All subordinate follicles in each follicular wave ultimately become atretic. Follicles within a wave either ovulate, or degenerate; follicles do not regress Figure. Ultrasound image of a bovine ovary with major structures identified. at the end of one wave, and then re-emerge in a subsequent wave. [1,, 3, 1, 1]. Figure 3 illustrates normal follicle growth and decline over a single wave. In two-wave cycles, the first wave typically emerges on the day of ovulation, denoted day. The dominant follicle is selected around day 3 and continues to develop and grow, while the remaining subordinate follicles in the cohort regress and degenerate. The dominant follicle reaches its maximal diameter around day, remains static for a time, and begins to degenerate around day 9 or. At this time, a second wave of new follicles emerges. The dominant follicle from this second cohort ovulates at the end of the cycle at around day 1 and the subordinate follicles regress and degenerate, as in the first wave. A new reproductive cycle begins if pregnancy is not established. This process is illustrated in Figure and is described in more detail in [8, 9]. Due to the wave-like follicular growth, the size of the dominant follicle, the size of the two largest subordinate follicles, and the total number of subordinate follicles will be useful features for distinguishing between the metestrus, diestrus and proestrus phases. Following ovulation, the wall of the ruptured follicle forms a structure called the corpus luteum. The CL passes through a period of initial growth (metestrus), followed by a period of maximal size and function (diestrus), and then enters a period of regression and ultimate demise (proestrus/estrus) preceding the next ovulation [13, 3]. The mean diameter of the CL [] over the estrous cycle is shown in Figure. Thus, the size of the CL is also a useful feature for reproductive phase discrimination. The relationship between the day in the estrous cycle and the sizes of the main structures of the ovary has been studied in detail

3 1 Wave Pattern Diameter (mm) Follicle diameter (mm) ovulation 1 8 Corpus Luteum 1 regression 8 Subordinate follicles Day of Follicular Wave 1 Figure 3. A schematic illustration of wavelike follicle growth. A group of follicles start growing simultaneously. After or 3 days, all subordinate follicles cease growing and begin to regress. The dominant follicle continues growing until it either ovulates (final wave behavior) or starts regressing (intermediate wave behavior). ovulation metestrus [8, 9,, 1]. In this study, an algorithm was developed for classification of bovine ovaries as belonging to either metestrus, diestrus, or proestrus phases based on a single day s examination. The estrus phase is not considered due to its short duration and lack of available data diestrus D>17 proestrus 1 Days in the cycle Figure. Schematic representation of follicle wave growth in a -wave cycle. Variation in diameter of the corpus luteum through the cycle is also shown. Motivation ovaries can be used to classify ovaries by phase of the estrous cycle. The decision tree algorithm is a procedure that partitions a decision into a series of simpler decisions at each node. This method does not necessarily use all of the available features. At each tree node, different questions about certain features may be evaluated. It has been shown that decision trees generalize well and can be used to solve a wide range of problems and efficiently classify new samples [, 1, 7, 8, 9]. The Bayes Decision Classifier is a statistical pattern recognition technique that assigns an input pattern to the class to which it most likely belongs. This classifier is based on Bayes rule of conditional probabilities [,, 1, 7]. Bayes classification requires knowledge about the classes, since it makes use of prior probabilities and classconditional probabilities. In our experiments, we used a Naı ve Bayes Classifier [, 1]. Monitoring reproductive cycles in mammals can be used to detect follicle wave emergence and to determine whether an animal is pre- or post-selection of a dominant preovulatory follicle. The ability to determine the current reproductive phase of an animal automatically from a single day s examination would enable rapid decisions to be made whether to begin monitoring the animal daily in order to determine dominant follicle selection, facilitate the division of a livestock herd into reproductively active or unresponsive factions, aid in the determination of optimal timing for initiation of drug treatments to obtain gametes for use in assisted reproduction protocols and to determine the effects of other drugs on reproductive function. In addition, excised bovine ovaries obtained from abbatoirs are routinely used for in vitro fertilization and embryo production. Automated classification of these ovaries into different estrous cycle phases will help in obtaining uniform groups of oocytes for commercial and scientific purposes. 1.3 Subordinate follicles Subordinate follicles 8 1. regression 1 Classification Methods.1 Materials and Methods Image Data Sets The image data set described herein was obtained from previous studies by Singh et. al. [, ]. For each an- Two pattern recognition methods were used to test the hypothesis that ultrasound-detected features of bovine 3

4 imal, both the left and right ovaries were ovariectomized (surgically removed). Three dimensional ultrasonographic volumes of ovaries were imaged in vitro in parallel planes at.mm increments with a broad-band, convex-array, ultrasound transducer interfaced with a ATL Ultra Mark 9 HDI ultrasound machine (Advanced Technology Laboratories, Brothell, WA). All images were 8 pixel 8-bit grayscale. The ovaries used in this study were collected and imaged on day 3 (day = day of previous ovulation) of wave 1 (), day 1 of wave () and after onset of proestrus (D 17) corresponding to the metestrus, diestrus, and proestrus phases respectively []. A total of pairs of ovaries (left and right ovaries of one animal) were used in the current study, and were divided into training and testing data sets. The training data set (data set A) consisted of 3 pairs of ovaries ovariectomized during metestrus (n = 8, ), diestrus (n = 7, ), and proestrus (n = 8, D 17). These animals exhibited both and 3 wave patterns; 19 exhibited a -wave pattern and exhibited a 3-wave pattern. The testing data set (data set B) consisted of a different group of pairs of ovaries ovariectomized during metestrus (n = 8, ), diestrus, (n =, ) and proestrus (n = 8, D 17). of these animals exhibited a -wave pattern and exhibited a 3-wave pattern. Both data sets were accompanied by a full set of schematic diagrams containing all of the information about the main structures inside both ovaries (size and location of the dominant and subordinate follicles and CL) which was recorded daily from ultrasound examinations that commenced from the day of the last ovulation (day ) until the day of the ovariectomy and complete in vitro imaging.. Feature Selection and Extraction The discussion of the wave theory of folliculogenesis in Section 1 illustrated why dominant follicle size, subordinate follicle sizes, and CL size, are key features in distinguishing between phases of the estrous cycle. Therefore, the features chosen to describe the current estrous phase of the ovary were size of the largest follicle (over both left and right ovaries), sizes of the largest and second-largest subordinate follicles (over both left and right ovaries), size of the CL, and total number of subordinate follicles mm. The sizes of follicles and CL were defined as the mean of the lengths of their major and minor axes. Features were extracted from data set A by manually measuring the diameter of the largest follicle in the ultrasound images. The slice containing the largest follicle area was located and the lengths of the major and minor axes of the follicle within that slice measured. The diameter (D) was taken as the mean of these two values. The remaining features, diameter of first subordinate follicle (S1), diameter of the second subordinate follicle (S), diameter of the CL (CL) and number of follicles (NF) were obtained from the schematic diagrams. For data set B, all features were obtained from the schematic diagrams. The automated extraction of these features from images has been the subject of past and current work. At present, it is not possible to extract all of these features automatically, reliably, and accurately. Follicle segmentation algorithms have been proposed [11, 18, 19, ], the best of which [19] can recognize follicles with mean sensitivity of.78, but includes a significant number of false positives (mean specificity was.71). There are no known segmentation algorithms for the CL, however, this is a topic of current research. A robust automatic segmentation algorithm for ovarian structures would enable the classification problem considered herein to be solved without human intervention..3 Classifier Design Two classifiers were implemented and their performances compared. A decision tree classifier was constructed using the C. decision tree generation algorithm (an improved divide and conquer approach to decision tree induction) which is based on the ID3 algorithm []. This method uses the information gain criterion to select the best feature on which to base a decision at a given node. The split value was calculated using the Gini coefficient, which is a measure of inequality [9]. A Naïve Bayes classifier was implemented to use the minimum error rate decision rule []. Uniform prior probabilities were used and the class-conditional probabilities were assumed to have a normal distribution [, 1, 8].. Experiment 1 For experiment 1, both the decision tree and Naïve Bayes classifiers were trained using data set A and tested using data set B. The mean feature values from data sets A and B are shown in Figure. The graph expresses the mean diameter in millimeters of D, S1, S, and CL and their standard deviations. In the case of the NF (number of follicles) feature, the value is dimensionless.. Experiment For experiment, the patterns for animals that exhibited a 3-wave pattern were removed from data sets A and B to form data sets A and B respectively. For data set A, samples in the D 17 class were removed leaving 8 patterns from, 7 patterns from and patterns from D 17; a total of 19 patterns. For data set B, samples in class D 17 exhibiting a 3-wave pattern were removed,

5 Data Set A Experiment 1 Data Set A Experiment Diameter (mm) Diameter (mm) 1 1 Data Set B D!17 Diameter (mm) Diameter (mm) 1 1 Data Set B D!17 D!17 D!17 Figure. Mean feature values and standard deviations for data sets A (top) and B (bottom) for Experiment 1. All features are measured in millimeters, except for NF which is dimensionless. Figure. Mean feature values and standard deviations for data sets A (top) and B (bottom) for Experiment. All features are measured in millimeters, except for NF which is dimensionless. leaving 8 patterns from, patterns from and patterns from D 17; a total of patterns. The decision tree and Naïve Bayes classifiers were then trained using the data set A, and tested using data set B. The mean feature values of this experiment are expressed in Figure for the data set A and data set B. The graph expresses the mean diameter in millimeters of D, S1, S, and CL. In the case of the NF feature, the value is dimensionless. From the figures, there is not a notable difference between data set A and data set A and data set B and data set B respectively due to the small number of patterns that presented a 3-wave pattern and the fact that the 3-wave animals belong only to the D 17 classes. There were a number of unavoidable potential sources of error when extracting the features used in this study from only single pair of images of an animal s ovaries. The sizes of the follicles were considered regardless of whether they were in their growing or regressing process since one cannot differentiate a growing dominant follicle from a regressing dominant follicle based on a single day s examination. From Figures and, it can be seen that the mean size of the first subordinate follicle (S1) in class has a large value compared to the expected value illustrated in Figure, which suggests that some of the S1 measurements for class may in fact have been measurements of the future dominant follicle of wave. Similarly some of the S values recorded for class could have resulted from the future first subordinate follicle of wave. For the D 17 class, the size of the mean for the S1 feature suggests that some values may have resulted from the regressing dominant follicle of wave 1. Similarly, some of the S measurements could be the size of the first subordinate follicle of wave. Section 3 shows that our classifier performed reasonably well, despite the potential sources of error arising from collecting features from only a single pair of images. Moreover, it is necessary to design a classifier that is robust to these errors since they will be unavoidable in practice; a single snapshot of an ovary in time may contain both regressing follicles from one wave and growing follicles from the subsequent wave. 3 Results 3.1 Experiment 1 For experiment 1, Both classifiers were trained using data set A and tested with data set B.

6 Table 1. Experiment 1: Confusion Matrix resulting from the classification of data set B by the decision tree classifier in Figure 7. The classification rate was %. Experiment 1: Decision Tree CL "18 >18 Classified as: D 17 Total 8 8 D D "1 > Decision Tree Classifier D!17 The decision tree derived from the C. algorithm is illustrated in Figure 7. Elliptical nodes are internal nodes which represent nodes where a choice of following the left or right branch is made based on the threshold of a particular feature in the pattern. For example, the root node indicates that the left branch is taken if the diameter of the CL is less than or equal to 18mm. The classification proceeds from top to bottom, starting from the root node (CL). Left or right branches are made at each decision node depending on the value of the feature in the given pattern. Rectangular nodes are leaf nodes. When a leaf node is encountered, the pattern is assigned to the class corresponding to that leaf node. It is notable that the inference algorithm determined that only two features from the training patterns were needed. This implies that the most discriminating of the features chosen are diameter of the CL and diameter of the dominant follicle. The confusion matrix for the classification of data set B is shown in Table 1. The rows and columns of the confusion matrix are indexed by the three classes, and D 17. The value of entry (i, j) in the matrix indicates the number of samples of class i that were classified as class j; thus, the desired result is a diagonal matrix. The last column represents the total number of tested instances per class. The classification rate is defined as the percentage of patterns that were classified correctly. The classification rate of the Decision Tree Classifier for Experiment 1 was % ( of patterns were classified correctly). This is an excellent result as it suggests that extremely high classification rates can be achieved through a decision tree that makes only two comparisons in the worst case, and requires only two features to be extracted from the input images Naïve Bayes Classifier For the Bayes Classifier, the features were considered to be conditionally independent, also known as the Naïve Bayes Rule. The class-conditional probabilities were assumed to be normal. The probability density function for a normal Figure 7. Experiment 1: The decision tree inferred from data set A. distribution with mean µ and standard deviation σ is expressed in equation 1 [, 1, 8]. f(x) = 1 ] (x µ) exp [ πσ σ (1) The mean µ and standard deviation σ of each feature were estimated from the set of values of the feature occuring in the training vectors. Prior probabilities were assumed to be uniform, although, in reality, the diestrus phase is typically longer than the others a fact that might warrant the use of non-uniform prior probabilities if this system were used to classify animals randomly chosen from a herd. The performance of the Naïve Bayes classifier was evaluated using data set B. The resulting confusion matrix is shown in Table. The matrix shows that 19 instances were classified correctly and 3 classified incorrectly for a classification rate of 8.3%. All instances of the class were classified correctly, however, the classifier misclassified two instances, and one D 17 instance. 3. Experiment For experiment, both classifiers were trained with data set A and tested with data set B from which samples arising from animals with a 3-wave follicular growth pattern were eliminated Decision Tree Classifier The decision tree inferred from data set A was identical to the decision tree of Experiment 1 (see Figure 7). The Decision Tree Classifier performance was evaluated using the

7 Table. Experiment 1: Confusion Matrix resulting from the classification of data set B by a Naïve Bayes classifier with normal classconditional probability distributions. The classification rate was 8.3%. Classified as: D 17 Total D Table 3. Experiment : Confusion matrix for the Naïve Bayes classifier with normal classconditional probability distributions using data Set B as the testing set. The classification rate was 9%. Classified as: D 17 Total D 17 data set B and % of the patterns were classified correctly. 3.. Naïve Bayes Classifier The class-conditional probability distributions were estimated using data set A with the same procedure used in Experiment 1. The performance of the classifier was evaluated using data set B. The resulting confusion matrix is shown in Table 3. The matrix shows that 9% (18 of ) patterns were classified correctly. All instances for D 17 class were classified correctly while and had one misclassification each and were classified as D 17 and respectively. Interestingly, the 3-wave patterns that were eliminated for this experiment were in fact classified correctly in experiment 1, which suggests the -wave and 3-wave patterns were not confused during the classification and presented similar characteristics. Summary and Conclusion The results of the present study support the hypothesis that the estrous phase of an animal of the bovine species can be automatically and robustly determined from ultrasounddetected features based on a single day s examination. The performance of the decision tree classifier in both experiments indicates that the D and CL features alone comprise sufficient information for accurate classification. Experiment 1 was trained with the data set A and tested with the data set B using animals with and 3 wave patterns. The decision tree classifier performed perfectly, classifying all instances correctly. The decision tree used only the CL and D features for the classification. The size of the CL is generally larger in the (diestrus) class ( 3.mm for data set A and mm for data set B) and smaller in classs (metestrus, 1.mm for data set A and 1.88mm for data set B) and D 17 (proestrus, 1mm for data set A and 1.1mm for data set B). The Naïve Bayes classifier classified 8.3% of the instances correctly (n=19) and 13.% incorrectly (n=3). Experiment was trained with the data set A and tested with the data set B which included only animals that exhibited a -wave pattern. The decision tree classifier classified % of the instances correctly. The Naïve Bayes classifier exhibited a similar performance compared with experiment 1, classifying 9% of the instances correctly. This experiment suggested that the extraction of 3-wave patterns did not eliminate any error or improve the performance significantly. Moreover in experiment 1, the 3-wave patterns were classified correctly, which suggests that both growth patterns patterns presented similar characteristics on the days of ovariectomy. Evaluation of the classifiers using a larger data set is required to fully demonstrate their insensitivity to two- and three-wave patterns of follicular growth. This would achieve an important level of robustness since it is not currently possible to determine whether an animal exhibits a or 3 wave pattern without daily examination. The success of the decision tree classifier based on only two features is somewhat surprising, given the errors that can arise in feature extraction due to the potential presence of follicles belonging to different waves in a single image. The largest follicle from an image acquired during the late metestrus phase (see Figure ) may in fact be the dominant follicle of an emerging second wave since there is significant variance in the actual day of wave emergence. That such a simple decision tree solves such an apparently complicated classification problem so well is rather astonishing and offers the potential for extremely fast, reliable, and consistent automatic decision making. The work herein constitutes the third stage of what could become a fully automated system for determining the current reproductive phase of mammals on the basis of a single ultrasound examination. The first stage of such a system would be the segmentation of the relevant ovarian structures. If the size of the dominant follicle and size of the CL are a sufficiently rich feature set, then the follicle segmentation problem can be solved fairly easily. Potočnik reported that his algorithm correctly segments nearly % of 7

8 large follicles greater than mm [19]. Segmentation of the CL is the subject of current research. For the second step, one need only recognize the largest follicle, and measure its diameter. Thus, if future work can achieve a robust segmentation algorithm for the CL, the entire process could be fully automated. Acknowledgements This study was supported by grants from the Natural Sciences and Engineering Research Council of Canada, the Canadian Institutes of Health Research, and the University of Saskatchewan. References [1] G. Adams and R. Pierson. Bovine model for study of ovarian follicular dynamics in humans. Theriogenology, 3:113 1, 199. [] A. Baerwald, G. Adams, and R. Pierson. Characterization of ovarian follicular wave dynamics in women. Biology of Reproduction, 9:3 31, 3. [3] A. Baerwald, G. Adams, and R. Pierson. A new model for ovarian follicular development during the human menstrual cycle. Fertility and Sterility, 8(1):11 1, July 3. [] P. A. Devijver and J. Kittler. Pattern Recognition: A Statistical Approach. Prentice Hall International, 198. [] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley & Sons, second edition, 1. [] O. Ginther. Ultrasonic Imaging and Animal Reproduction: Fundamentals (Book 1). Equiservices Publishing, 199. [7] O. Ginther. Ultrasonic Imaging and Animal Reproduction: Horses (Book ). Equiservices Publishing, 199. [8] O. Ginther, J. Kastelic, and L. Knopf. Composition and characteristics of follicular waves during the bovine estrous cycle. Animal Reproduction Science, :187, [9] O. Ginther, J. Kastelic, and L. Knopf. Intraovarian relationships among dominant and subordinate follicles and the corpus luteum in heifers. Theriogenology, 3():787 79, [] O. Ginther, L. Knopf, and J. Kastelic. Temporal associations among ovarian events in cattle during oestrous cycle with two and three follicular waves. Reproduction Fertility, 87:3 3, [11] A. Krivanek, W. Liang, G. Sarty, R. Pierson, and M. Sonka. Automated follicle analysis in ovarian ultrasound. Medical Imaging, Proceedings of SPIE, 3338:88 9, [1] R. Pierson and G. Adams. Computer-assisted image analysis, diagnostic ultrasonography and ovulation induction: Strange bedfellows. Theriogenology, 3: 11, 199. [13] R. Pierson and O. Ginther. Ultrasonography of the bovine ovary. Theriogenology, 1:9, 198. [1] R. Pierson and O. Ginther. Follicular populations during the estrous cycle in heifers: Part I. influence of day. Animal Reproduction Science, 1:1 17, [1] R. Pierson and O. Ginther. Reliability of diagnostic ultrasonography for identification and measurement of follicles and detecting corpus luteum in heifers. Theriogenology, 9:1 37, [1] R. Pierson and O. Ginther. Follicular populations during the estrous cycle in heifers: Part III. time of selection of ovulatory follicle. Animal Reproduction Science, 1:81 9, [17] R. Pierson and O. Ginther. Ultrasonic imaging of the ovaries and uterus in cattle. Theriogenology, 9:1 37, [18] B. Potočnik. Automated ovarian follicle segmentation using region growing. In First Int l workshop on Image and Signal Processing and Analysis, pages 17 1,. [19] B. Potočnik and D. Zazula. Automated analysis of a sequence of ovarian ultrasound images. Part I: segmentation of single D images. Image and Vision Computing, :17,. [] J. R. Quinlan. Induction of Decision Trees. Morgan Kaufmann, 199. Originally published in Machine Learning 1:81, 198. [1] S. Russell and P. Norvig. Artificial Intelligence A Modern Approach. Prentice Hall, second edition, 3. [] G. Sarty, W. Liang, M. Sonka, and R. Pierson. Semiautomated segmentation of ovarian follicular ultrasound images using knowledge-based algorithm. Ultrasound in Medicine and Biology, (1):7, [3] J. Singh. Bovine Ovary: Morphologic and Biochemical Kinetics. PhD thesis,, [] J. Singh, R. Pierson, and G. Adams. Ultrasound image attributes of bovine corpus luteum: Structural and functional correlates. Journal of Reproduction and Fertility, 9:3, [] J. Singh, R. Pierson, and G. Adams. Ultrasound image attributes of bovine ovarian follicles and endocrine and functional correlates. Journal of Reproduction and Fertility, 11:19 9, [] J. Tom, R. Pierson, and G. Adams. Quantitative echotexture analysis of bovine corpora lutea. Theriogenology, 9:13 13, [7] A. R. Webb. Statistical Pattern Recognition. John Wiley & Sons, second edition,. [8] I. H. Witten and E. Frank. Data Mining, Practical Machine Learning Tools and Techniques. Morgan Kaufmann Publishers, second edition,. [9] N. Ye. The Handbook of Data Mining. Lawrence Erlbaum Associates Publishers, 3. 8