Article Computational models for prediction of IVF/ICSI outcomes with surgically retrieved spermatozoa

RBMOnline - Vol 11. No 3. 2005 325-331 Reproductive BioMedicine Online; www.rbmonline.com/article/1800 on web 11 July 2005 Article Computational models for prediction of IVF/ICSI outcomes with surgically retrieved spermatozoa Dr Moshe Wald is a graduate of the Hebrew University and Hadassah School of Medicine, Jerusalem, Israel. Upon completion of his Urology residency at Bnai-Zion Medical Centre, Haifa, Israel, he graduated in the Andrology and Male Infertility fellowship programme at the University of Illinois at Chicago, under the guidance of Dr Craig Niederberger. Dr Wald is currently an Assistant Professor of Male Infertility and Andrology at the University of Iowa, Iowa City, USA. His current research interests include the utilization of surgically retrieved cryopreserved spermatozoa for ICSI, seminal flow cytometry for evaluation of male infertility and testosterone replacement therapy. Dr Moshe Wald Moshe Wald 1,5, Amy ET Sparks 2, Jay Sandlow 3, Brad Van-Voorhis 2, Craig H Syrop 2, Craig S Niederberger 4 1 Department of Urology; 2 Division of Reproductive Endocrinology and Infertility, Department of Obstetrics and Gynecology, University of Iowa Hospitals and Clinics, Iowa City, IA; 3 Department of Urology, Medical College of Wisconsin, Milwaukee, WI; 4 Department of Urology, University of Illinois at Chicago, IL, USA 5 Correspondence: Department of Urology University of Iowa Health Care 200 Hawkins Drive, Iowa City, IA 52242, USA. Tel: +1 319 3568922; Fax: +1 319 3563900; e-mail: moshe-wald@uiowa.edu Abstract IVF/intracytoplasmic sperm injection (ICSI) using surgically retrieved spermatozoa (SRS) is a key option in the treatment of severe male infertility. It was aimed to develop a computational model for the prediction of this modality s outcome. A dataset of 113 exemplars, derived from patients who underwent IVF/ICSI with SRS, was retrospectively analysed. The dataset, containing input features maternal age, sperm retrieval technique, type of spermatozoa used, type of male factor and output intrauterine pregnancy, was randomized into a modelling ( training ) set of 83 and cross-validation ( test ) set of 30. neuron++, a set of C++ programs, was used to model the dataset using linear and quadratic discriminant function analysis, logistic regression, and neural computation. A 4-hidden node neural network was found to have the highest accuracy, with a test set receiver operator characteristic (ROC) curve area of 0.783. Reverse regression of this neural network showed maternal age to be the most significant feature in predicting pregnancy (P = 0.025), followed by sperm type (P = 0.076). Type of male factor (P = 0.47) and sperm retrieval technique (P = 0.88) did not predict outcome. In summary, a neural network of clinical relevance was found to be superior in terms of IVF/ICSI outcome prediction. Future media deployment is planned. Keywords: ICSI, IVF, maternal age, neural network, sperm cryopreservation, sperm retrieval techniques Introduction The introduction of intracytoplasmic sperm injection (ICSI) revolutionized the management of male factor infertility. IVF/ ICSI may be performed with either ejaculated or surgically retrieved spermatozoa. Investigators report acceptable fertilization and pregnancy outcomes in most series of IVF/ICSI cycles from azoospermic men, using either fresh or cryopreserved surgically retrieved spermatozoa, and from men with both obstructive (OA) and non-obstructive (NOA) azoospermia (Nicopoullos et al., 2004). With ejaculated spermatozoa, IVF/ ICSI achieves high fertilization and pregnancy rates, regardless of any of three basic semen parameters (total count, motility and morphology) (Nagy et al., 1995a). However, debate remains regarding certain important issues, including the most suitable site for retrieval, the role of sperm cryopreservation and the effectiveness of IVF/ICSI in patients with OA as compared with those with NOA. While similar IVF/ ICSI outcomes were demonstrated in cycles using epididymal and testicular spermatozoa in OA patients (Hovatta et al., 1995; Nagy et al., 1995b; Silber et al., 1995; Fahmy et al., 1997; Mansour et al., 1997; Rosenlund et al., 1997; Dohle et al., 1998; Palermo et al., 1999), the data comparing ICSI outcomes between patients with OA and NOA are less consistent. Although some authors suggest no difference (Devroey et al., 1996; Windt 325

326 et al., 2002), the majority of reports show significantly impaired fertilization or pregnancy outcome in cycles using testicular spermatozoa from NOA patients compared with testicular spermatozoa from OA patients (Kahraman et al., 1996; Fahmy et al., 1997; Mansour et al., 1997; Palermo et al., 1999; DeCroo et al., 2000; Pasqualotto et al., 2002). The outcome of IVF/ICSI cycles using fresh or cryopreserved retrieved spermatozoa is also a source of much debate. The majority of reports suggest no significant difference in outcome with the use of cryopreserved surgically retrieved spermatozoa (Devroey et al., 1995; Nagy et al., 1995b; Silber et al., 1995; Friedler et al., 1997, 1998; Tournaye et al., 1999; Habermann et al., 2000; Windt et al., 2002). Others, however, have reported a significantly lower fertilization rate (FR) (DeCroo et al., 1998; Wood et al., 2002), clinical pregnancy rate (CPR) (Palermo et al., 1999; Christodoulou et al., 2002) and implantation rate (IR) (DeCroo et al., 1998; Christodoulou et al., 2002) using cryopreserved spermatozoa from NOA and OA patients. This study aimed to use a novel approach to address these controversies. Using several linear and non-linear mathematical models, the relationship of maternal age, type of male factor, sperm retrieval technique and sperm type used with intrauterine pregnancy achieved by IVF/ICSI was investigated. One of these models, neural computation, simulates the physiology of the biological neuron (Niederberger, 2001). A learning algorithm is iteratively presented to this nodal system, in which a transfer function passes information in each node from inputs to output in a defined topology, with a selected optimization algorithm that minimizes the error derived from the network s outcome state and the presented outcomes data, in this case, IVF/ICSI induced intrauterine pregnancies. Maternal age, type of male factor, sperm retrieval technique and sperm type were chosen as input features, and an attempt was made to develop a computational model combining these features to investigate their relationship in IVF/ICSI. Materials and methods Clinical data were collected from 85 patients (22 43 years old, mean age 32.45 years), who had undergone 113 IVF/ICSI cycles, primarily for male factor infertility. The male and female factors are listed in Tables 1 and 2 respectively. Parameters used for this study included maternal age, type of sperm retrieval technique (testicular sperm extraction, testicular sperm aspiration and microscopic epididymal sperm aspiration), type of spermatozoa used (cryopreserved or fresh ), type of male factor infertility, and intrauterine pregnancies achieved through IVF/ICSI. Sperm retrieval, type of male factor and intrauterine pregnancy parameters were assigned either 1, if present for each exemplar, or 0, if not. For sperm type, cryopreserved spermatozoa were assigned 1, while fresh spermatozoa were assigned 0. A total of 113 intrauterine pregnancy scores, as output and the corresponding input variates maternal age, type of surgical retrieval technique, type of spermatozoa used and type of male infertility factor, comprised the dataset for computational modelling using neuron++, a set of C++ programs developed using the Cygwin (Red Hat) GNU C++ port for Windows (Microsoft) distributed across Pentium (Intel) platforms. neuron++ was used to model the dataset using the linear mathematical modelling methods linear and quadratic discriminant function analysis (LDFA, QDFA) and logistic regression (LR), and the non-linear method of neural computation (NNET). The dataset were randomized into a modelling ( training ) set of 83 exemplars, with a separate, completely independent cross-validation, test, set of 30 exemplars. Cross-validation is a frequently used technique for estimating the accuracy of theories learned by machine learning algorithms. This technique repeatedly splits the dataset into two subsets, learns a theory from the first subset and tests the theory ( cross-validates it) by seeing how well it classifies the other subset. Thus, the test set performance serves as a measure to estimate the error rate of the trained model, and as such, is expected to degrade in cross validation analysis. Different architectures with varying hidden nodes have been investigated to identify the model with the greatest accuracy. Model training was considered as complete when the error was observed to be oscillating at a local error minimum. Receiver operator characteristic curve (ROC) area served to assess the model s accuracy, and was computed using the statistical method described by Wickens (2002) and by the traditional trapezoidal method. The best ROC area achieved for each model was then recorded (Table 4, see below). The trapezoidal method uses trapezoids and numerical calculation to calculate the area under the curve. By default, the number of trapezoids is set to the number of data points. Of note, only the Wickens statistical method can generate a P-value. For this method, a smaller P-value is worse (closer to 1 is more linear). Wilk s generalized likelihood ratio test was used to determine which input features were significant to the model s outcome in a reverse regression analysis (Golden, 1996). The model was deployed in the Javascript language for ready availability on the world wide web (Figure 1), and in PalmOS for physicians using handheld computers (Figure 2). The clinician enters the patient s data using the forms-based interface, and the model reports the odds ratio for intrauterine pregnancy. Currently, the model may be found at www.urocomp.net. Results Eighty-five patients were included in this study, with a mean age of 32.45 years (22 43), who had undergone a total of 113 IVF/ ICSI cycles. Cryopreserved and fresh spermatozoa were used in 98 and 15 IVF/ICSI cycles respectively. The data concerning the male infertility factors and sperm retrieval techniques are summarized in Tables 1 and 3 respectively. Intrauterine pregnancies have been achieved in 50 cycles (44.2%). A 1-hidden layer, 4-hidden node (HN) neural network was found most accurately to predict intrauterine pregnancy, with ROC areas of 0.783 and 0.923 for the test and training sets respectively. Results of LDFA, QDFA and logistic regression are shown in Table 4. Discriminant analysis (LDFA, QDFA) This is a multivariate statistical procedure that mathematically defines a special discriminant function to separate a study population by one classification variable. The discriminant function can use several quantitative variables, each of which

Table 1. Male infertility factors. Table 2. Female infertility factors. Male infertility factor Number (%) Female infertility factor Number (%) Previous vasectomy (no reversal) 16 (14.2) Previous vasectomy (reversal performed) 17 (15.0) Congenital bilateral absence of the vas deferens 21 (18.6) Other obstructive conditions 17 (15.0) Varicocoele 4 (3.5) Non-obstructive azoospermia 25 (22.1) Other 13 (11.5) Tubal 7 (8.2) Ovulatory dysfunction 8 (9.4) Endometriosis 6 (7.1) Uterine (fibroids) 2 (2.4) Polycystic ovarian disease 2 (2.4) a b Figure 1. World wide web version of model. (a) Data insertion: forms-based interface. Clinical input features may be easily inserted into the model by typing the maternal age and checking the appropriate box. (b) Model report of odds ratio for pregnancy. This is generated by the model upon clicking the Predict button in the previous screen (see a). a b Figure 2. PalmOS version of model. (a) Data insertion: forms-based interface. Clinical input features may be easily inserted into the model by typing the maternal age, checking the appropriate box and selecting the male factor from a drop-down menu. (b) Model report of odds ratio for pregnancy, which is generated by the model upon clicking the Predict button in the previous screen (see Figure 1b). 327

Table 3. Sperm retrieval techniques. Sperm retrieval technique Number (%) Testicular sperm extraction (TESE) 51 (45.1) Testicular sperm aspiration (TESA) 14 (12.4) Microscopic epididymal sperm aspiration 48 (42.5) Table 4. Model accuracies. ROC = receiver operator characteristic; LDFA = linear discriminant function analysis; QDFA = quadratic discriminant function analysis; HN = hidden nodes. Method ROC area Training set (ROC computation method) Test set (ROC computation method) LDFA 0.576 (trapezoidal method) 0.163 (trapezoidal method) QDFA 0.000 (trapezoidal method) 0.000 (trapezoidal method) Linear regression 0.698 (statistical method, P = 1.0) 0.575 (trapezoidal method) 4-HN neural network 0.923 (trapezoidal method) 0.783 (statistical method, P = 1.0) 328 makes an independent contribution to the overall discrimination. Taking into consideration the effect of all quantitative variables, this discriminant function produces the statistical decision for guessing to which subgroup of classification variable each subject belongs. Assuming a multivariate normal distribution of quantitative variables within each level of classification variable, a parametric method generates either a linear discriminant function (equal within-class covariance) or a quadratic discriminant function (unequal within-class covariance). In either case, the discriminant function is a weighted combination of all quantitative variables. The performance of discriminant analysis can be evaluated by estimating the error rate (probability of misclassification). Wilk s generalized likelihood ratio test (GLRT) was used to assess the significance of the neural network s input parameters. This statistical test was shown to be useful for deciding which of several subsets of artificial neural network system architectures is most appropriate for a certain statistical environment, given that the full model provides a good fit to the observed data. In reverse regression, reduced models are designed to contain the investigated parameters, by eliminating other parameters of the original model. The GLRT provides a formula by which the fitness of a reduced model to the data can be compared with that of the full model, at a certain significance level (Golden, 1996). Statistical analysis of neural computational models is based on the observation that two fully trained neural networks with final errors E(yreduced) ) and E(yfully ) that are sufficiently close may be discriminated by a chi-square probability distribution. In stepwise regression, the two networks yreduced and yfully are chosen to differ by one input variable. Specifically, reverse regression begins with determining the full network s final error. Input variables are then removed one at a time, the subnetworks so generated are trained to completion and their errors recorded, and the variable with the largest P-value (least significant) is then recorded with its P-value. Using the final error of the network which is deficient, the one node with the largest P-value as a starting point, input variables are then removed two at a time, with each pair containing the single variable with the largest P-value and one of the remaining input variables. The pair of variables with the largest P-value (least significant) is then recorded with its P-value, and the process repeats until a chosen threshold P-value is reached. Reverse regression of the neural network based on Wilk s generalized likelihood ratio test revealed maternal age to be most significant (P = 0.025) in predicting pregnancy outcomes, followed by sperm type (P = 0.076). Type of male factor (P = 0.47) and sperm retrieval technique (P = 0.88) were not found to be significant predictors. Discussion The introduction of ICSI and the development of advanced surgical sperm retrieval techniques have revolutionized the treatment of severe male factor infertility. However, the effects of the sperm source, aetiology of azoospermia and sperm cryopreservation on IVF/ICSI outcomes are still controversial. In general, the overall pregnancy rate per patient is influenced not only by the transfer of embryos in the first cycle, but also from the additional contribution provided by surplus embryos that may remain cryopreserved from the same cohort of oocytes initially retrieved. However, as the specific aim of this study was to design a model for the prediction of IVF/ICSI outcomes using surgically retrieved spermatozoa, only ICSI cycles have been included in the modelled dataset, and not treatment cycles in which frozen embryos were used. While inclusion of the embryo transfer cycles is too large an undertaking for this manuscript, these data will be discussed in another manuscript.

Several studies have demonstrated an equal IVF/ICSI pregnancy outcome using epididymal or testicular spermatozoa in azoospermic patients with similar aetiology (Fahmy et al., 1997; Palermo et al., 1999; Nicopoullos, 2004). However, epididymal aspiration has been suggested as the retrieval method of choice for patients with obstructive azoospermia, as it minimizes the potential consequences of testicular retrieval, namely testicular inflammation, hematoma formation and devascularization (Schlegel and Su, 1997). While simple sperm aspiration procedures such as percutaneous epididymal sperm aspiration (PESA) were shown to be as effective as an open microscopic epididymal sperm aspiration (MESA) in terms of retrieval rates and IVF/ICSI outcome (Tournaye et al., 1998), the latter has been recommended as the first-line method of retrieval in men with OA in view of the opportunity it gives for diagnosis, reconstruction where appropriate, and retrieval of greater numbers of spermatozoa for cryopreservation. Furthermore, MESA may be associated with a lesser extent of epididymal injury and subsequent scarring when compared with blind procedures such as PESA. In addition to MESA, testicular sperm extraction and aspiration (TESE and TESA respectively) were included as input parameters to the model, as these methods are recognized as effective and widely used sperm retrieval techniques for IVF/ICSI. Interestingly, the use of an intravenous catheter for testicular aspiration biopsy has recently been reported to significantly improve sperm retrieval compared with FNA in obstructive azoospermia, allowing for cryopreservation of excess tissue. However, the safety of this method is yet to be established (Fahmy et al., 2004). Furthermore, certain advanced sperm retrieval techniques stem from the latter two, including microdissection (Schlegel, 1999) and testis fine needle aspiration mapping (Turek et al., 2000). However, the procedures were categorized as either TESE or TESA, as additional sub-categorization of sperm retrieval techniques is too large an undertaking for this manuscript. Consensus is also lacking with regard to the effect of the aetiology for azoospermia on IVF/ICSI outcomes. Significantly lower fertilization and clinical pregnancy rates were reported using spermatozoa from patients with non-obstructive azoospermia (Kahraman et al., 1996; Fahmy et al., 1997; Mansour et al., 1997). Interestingly, some authors reported impaired fertilization but similar pregnancy outcome in NOA (Palermo et al., 1999; DeCroo et al., 2000), while others noted similar fertilization but impaired pregnancy outcome (Pasqualotto et al., 2002). On the other hand, a recent report showed no statistically significant impairment in any outcome measure in NOA (Windt et al., 2002). Different diagnostic criteria and types of sperm retrieval have been suggested as possible explanations for the inconsistency between the published data (Nicopoulos et al., 2004). The majority of reports suggest no significant worsening in outcome with the use of cryopreserved surgically retrieved spermatozoa (Devroey et al., 1995; Nagy et al., 1995b; Silber et al., 1995; Friedler et al., 1997, 1998; Tournaye et al., 1999; Habermann et al., 2000; Windt et al., 2002). However, most of the studies comparing fresh versus frozen testicular spermatozoa use spermatozoa from a combination of men with OA and NOA without comparing the effect of cryopreservation by aetiology of infertility. Significantly lower fertilization rates have been reported for cryopreserved testicular spermatozoa (DeCroo et al., 1998; Wood et al., 2002; Nicopoullos et al., 2004). Interestingly, no effect of cryopreservation on epididymal spermatozoa was demonstrated in these studies. Pregnancy rates were not significantly impaired after using cryopreserved spermatozoa, but a trend suggestive of impaired implantation following the use of frozen thawed samples was noted which did not reach clinical significance (Nicopoullos et al., 2004). Various factors may affect the treatment outcome for obstructive and non-obstructive azoospermia when spermatozoa from cryopreserved testicular specimens are utilized for ICSI. Interestingly, sperm motility after freeze thawing was reported as a major determinant of the success of ICSI in the setting of non-obstructive azoospermia (Dafopoulos et al., 2005a). However, comparison of fresh and frozen thawed testicular spermatozoa from only men with NOA showed no significant differences in either fertilization or pregnancy outcome (Friedler et al., 1997). In fact, Dafapoulos et al. have looked at ICSI outcomes with surgically retrieved cryopreserved spermatozoa in NOA, and reported a satisfactory cumulative pregnancy rate in this group of patients (Dafopoulos et al., 2005b). Comparative analysis of IVF/ICSI pregnancy and implantation outcome in OA patients only has been limited by the small numbers investigated. Meta-analysis of published data may improve the ability to make clinical decisions based on these findings (Nicopoullos et al., 2003, 2004). Interestingly, an ongoing prospective study investigated the effects of freeze thawing on testicular sperm DNA fragmentation, fertilization rates and pregnancy rates following ICSI with surgically retrieved testicular spermatozoa obtained from men with obstructive azoospermia. Sperm nuclear DNA was found to be significantly damaged by slow freezing followed by fast thawing. While no differences were observed in the fertilization rates between ICSI cycles using either fresh or frozen thawed testicular spermatozoa, the pregnancy rates tended to be higher with fresh spermatozoa (30%) compared with ICSI cycles using cryopreserved spermatozoa (26%), although this difference did not reach statistical significance. However, the authors have acknowledged that a much larger multi-centre trial will be needed to address the possible effect of sperm cryopreservation on ICSI pregnancy rates in this patient population (Thompson- Cree et al., 2003). These controversies are addressed using a different approach in the study reported here. A model has been developed which integrates the clinically relevant data to formulate a numerical prediction of IVF/ICSI outcome. The significance of its particular inputs is then mathematically determined. As a simple, clinically oriented tool was sought, intrauterine pregnancy was used as the model s outcome as it may better reflect the desired end-point for IVF/ICSI. Neural computation is a non-linear modelling technique, adopting features of the physiologic function of the biological neuron to inspire its mathematical models (Niederberger, 2001). Urological problems, such as the detection of prostate cancer and erectile dysfunction are included in the hundreds of biomedical applications of artificial neural networks reported in the scientific literature (Wald et al., 2005). A 4-hidden node network with a modelling ( training ) set of 83 exemplars and a separate cross-validation ( test ) set of 30 was found to have the highest accuracy in predicting IVF/ICSI induced intrauterine 329

330 pregnancies (ROC areas 0.923 and 0.783 for the training and test sets respectively). The 4-hidden node neural network revealing the best goodness of fit was further analysed to determine the individual significance of each of the input features to the model s outcome. Reverse regression based on Wilk s generalized likelihood ratio test revealed maternal age to be most significant (P = 0.025) in predicting pregnancy, followed by sperm type (P = 0.076). Type of male factor (P = 0.47) and sperm retrieval technique (P = 0.88) did not significantly affect pregnancy outcome. These results confirm the clinical significance of the computational model, as they are consistent with the recognized association between maternal age and IVF/ICSI outcome. The analysis of the neural network showed IVF/ICSI outcome to be related to the type of spermatozoa used. It should be noted that in a non-linear mathematical model such as neural computation, a simple statement about superiority of frozen or fresh sperm in pregnancy outcomes cannot be derived. However, clinicians using the model may substitute either fresh or frozen spermatozoa into that variate to predict whether the outcome would be expected to be better or worse for that particular maternal age, retrieval technique and male factor type. Interestingly, the techniques used for sperm retrieval and the male factor type alone were not related to IVF/ICSI outcome. These findings are consistent with studies reporting an equal IVF/ICSI pregnancy outcome using epididymal or testicular spermatozoa in patients with similar aetiology (Fahmy, 1997; Palermo, 1999; Nicopoullos, 2004) and similar outcomes for NOA and OA patients (Windt, 2002) respectively. In summary, the statistical analysis of the neural network, along with its demonstrated clinical correlations support the use of this model as a useful predictive tool when IVF/ICSI is considered in the presence of male factor infertility. References Christodoulou K, Jerkovic S, Geyer J et al. 2002 The effect of TESA cryopreservation on the outcome of ICSI cycles. BFS/ACE abstracts. Human Fertility 5, 86. Dafopoulos K, Griesinger G, A Schultze-Mosgau A et al. 2005a Factors affecting outcome after ICSI with spermatozoa retrieved from cryopreserved testicular tissue in non-obstructive azoospermia. Reproductive BioMedicine Online 10, 455 460. Dafopoulos K, Griesinger G, Schultze-Mosgau A et al. 2005b Cumulative pregnancy rate after ICSI with cryopreserved testicular tissue in non-obstructive azoospermia. Reproductive BioMedicine Online 10, 461 468. DeCroo I, Van der Elst J, Everaert K et al. 2000 Fertilisation, pregnancy, and embryo implantation rates after ICSI in cases of obstructive and non-obstructive azoospermia. Human Reproduction 15, 1381 1388. DeCroo I, Van der Elst J, Everaert K et al. 1998 Fertilisation, pregnancy, and embryo implantation rates after ICSI with fresh or frozen thawed testicular sperm. Human Reproduction 13, 1893 1897. Devroey P, Nagy P, Tournaye H et al. 1996 Outcome of intracytoplasmic sperm injection with testicular spermatozoa in obstructive and nonobstructive azoospermia. Human Reproduction 11, 1015 1018. Devroey P, Silber S, Nagy Z et al. 1995 Ongoing pregnancies and birth after intracytoplasmic sperm injection with frozen thawed epididymal spermatozoa. Human Reproduction 10, 903 906. Dohle GR, Ramos L, Pieters MHEC et al. 1998 Surgical sperm retrieval and intracytoplasmic sperm injection as treatment of obstructive azoospermia. Human Reproduction 13, 620 623. Fahmy I, Kamal A, Aboulghar M et al. 2004 Percutaneous aspiration biopsy using an intravenous catheter for testicular sperm retrieval in patients with obstructive azoospermia. Reproductive BioMedicine Online 9, 102 105. Fahmy I, Mansour R, Aboulghar M et al. 1997 Intracytoplasmic sperm injection using surgically retrieved epididymal and testicular spermatozoa in cases of obstructive and non-obstructive azoospermia. International Journal of Andrology 20, 37 44. Friedler S, Raziel A, Soffer Y et al. 1998 The outcome of intracytoplasmic injection of fresh and cryopreserved epididymal spermatozoa from patients with obstructive azoospermia a comparative study. Human Reproduction 13, 1872 1877. Friedler S, Raziel A, Soffer Y et al. 1997 Intracytoplasmic injection of fresh and cryopreserved testicular spermatozoa in patients with non-obstructive azoospermia a comparative study. Fertility and Sterility 68, 892 897. Golden RM 1996 Mathematical Methods for Neural Network Analysis and Design. MIT Press, Cambridge, MA. Habermann H, Seo R, Cielsak J et al. 2000 In vitro fertilisation outcomes after intracytoplasmic sperm injection with fresh or frozen thawed testicular spermatozoa. Fertility and Sterility 73, 955 960. Hovatta O, Moilanen J, von Smitten K et al. 1995 Testicular needle biopsy, open biopsy, epididymal aspiration and intracytoplasmic sperm injection in obstructive azoospermia. Human Reproduction 10, 2595 2599. Kahraman S, Ozgur S, Alatas C et al. 1996 High implantation rates with testicular sperm extraction and intracytoplasmic sperm injection in obstructive and non-obstructive azoospermia. Human Reproduction 11, 673 676. Mansour RT, Kamal A, Fahmy I et al. 1997 Intracytoplasmic sperm injection in obstructive and non-obstructive azoospermia. Human Reproduction 12, 1974 1979. Nagy Z, Silber S, Liu J et al. 1995a The result of intracytoplasmic sperm injection is not related to any of the three basic sperm parameters. Human Reproduction 10, 1123 1129. Nagy Z, Silber S, Liu J et al. 1995b, Using ejaculated, fresh, and frozen thawed epididymal and testicular spermatozoa gives rise to comparable results after intracytoplasmic sperm injection. Fertility and Sterility 63, 808 815. Nicopoullos JDM, Gilling-Smith C, Almeida PA et al. 2004 The results of 154 ICSI cycles using surgically retrieved sperm from azoospermic men. Human Reproduction 19, 579 585. Nicopoullos JDM, Gilling-Smith C, Almeida PA et al. 2003 Fresh or frozen thawed sperm in azoospermic men? A meta-analysis. Human Reproduction 18, 1 64. Niederberger CS 2001 Neural computation in urology: an orientation. Molecular Urology 5, 133. Palermo GD, Sclegel PN, Hariprashad JJ et al. 1999 Fertilisation and pregnancy outcome with intracytoplasmic sperm injection for azoospermic men. Human Reproduction 14, 741 748. Pasqualotto FF, Rossi-Ferragut LM, Rocha CC et al. 2002 Outcome of in vitro fertilisation and intracytoplasmic injection of epididymal and testicular sperm obtained from patients with obstructive and non-obstructive azoospermia. Journal of Urology 167, 1753 1756. Rosenlund B, Sjoblom P, Dimitrakopoulos A et al. 1997 Epididymal and testicular sperm for intracytoplasmic sperm injection in the treatment of obstructive azoospermia. Acta Obstetrica et Gynecologica Scandinavica 75, 135 139. Schlegel PN. 1999 Testicular sperm extraction: microdissection improves sperm yield with minimal tissue excision. Human Reproduction 14, 131 135. Schlegel PN, Su L 1997 Physiological consequences of testicular sperm extraction. Human Reproduction 12, 1688 1692. Silber SJ, Van Steirteghem AC, Liu J et al. 1995 High fertilisation and pregnancy after intracytoplasmic sperm injection with spermatozoa obtained from testicular biopsy. Human Reproduction 10, 148 152.

Thompson-Cree MEM, McClure N, Donnelly ET et al. 2003 Effects of cryopreservation on testicular sperm nuclear DNA fragmentation and its relationship with assisted conception outcome following intracytoplasmic sperm injection. Reproductive BioMedicine Online 7, 449 455. Tournaye H, Merdad T, Silber S et al. 1999 No differences in outcome after intracytoplasmic sperm injection with fresh or frozen thawed epididymal spermatozoa. Human Reproduction 14, 90 95. Tournaye H, Clasen K, Aytoz A et al. 1998 Fine needle aspiration versus open biopsy for testicular sperm recovery: a controlled study in azoospermic patients with normal spermatogenesis. Human Reproduction 13, 901 904. Turek PJ, Ljung BM, Cha I et al. 2000 Diagnostic findings from testis fine needle aspiration mapping in obstructed and nonobstructed azoospermic men. Journal of Urology 163, 1709 1716. Wald M, Seftel AD, Ross LS et al. 2005 Computational models for detection of erectile dysfunction. Journal of Urology 173, 167 170. Wickens TD 2002 Elementary Signal Detection Theory. Oxford University Press, New York. Windt ML, Coetzee K, Kruger TF et al. 2002 Intracytoplasmic sperm injection with testicular spermatozoa in men with azoospermia. Journal of Assisted Reproduction and Genetics 19, 53 59. Wood S, Thomas K, Schnauffer K et al. 2002 Reproductive potential of fresh and cryopreserved epididymal and testicular spermatozoa in consecutive intracytoplasmic sperm injection cycles in the same patients. Fertility and Sterility 77, 1162 1166. Received 29 March 2005; revised and resubmitted 9 May 2005; accepted 9 June 2005. 331