Radiology
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Biagiotti, R.
Right arrow Articles by Gacci, G.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Biagiotti, R.
Right arrow Articles by Gacci, G.
(Radiology. 1999;210:399-403.)
© RSNA, 1999

Computer Applications

Predicting Ovarian Malignancy: Application of Artificial Neural Networks to Transvaginal and Color Doppler Flow US

Roberto Biagiotti, MD1,2, Cristina Desii, MD1, Ermanno Vanzi, MD2,3 and Guido Gacci, MD1

1 Division of Obstetrics and Gynecology, Santa Maria Annunziata Hospital (R.B., C.D., G.G.)
2 Departments of Obstetrics and Gynecology (R.B., E.V.)
3 Radiology (E.V.), Università di Firenze, Florence, Italy.


   Abstract
TOP
 Abstract
Introduction
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 References
 
PURPOSE: To compare the performance of artificial neural networks (ANNs) with that of multiple logistic regression (MLR) models for predicting ovarian malignancy in patients with adnexal masses by using transvaginal B-mode and color Doppler flow ultrasonography (US).

MATERIALS AND METHODS: A total of 226 adnexal masses were examined before surgery: Fifty-one were malignant and 175 were benign. The data were divided into training and testing subsets by using a "leave n out method." The training subsets were used to compute the optimum MLR equations and to train the ANNs. The cross-validation subsets were used to estimate the performance of each of the two models in predicting ovarian malignancy.

RESULTS: At testing, three-layer back-propagation networks, based on the same input variables selected by using MLR (ie, women's ages, papillary projections, random echogenicity, peak systolic velocity, and resistance index), had a significantly higher sensitivity than did MLR (96% vs 84%; McNemar test, P = .04). The Brier scores for ANNs were significantly lower than those calculated for MLR (Student t test for paired samples, P = .004).

CONCLUSION: ANNs might have potential for categorizing adnexal masses as either malignant or benign on the basis of multiple variables related to demographic and US features.

Index terms: Computers, diagnostic aid • Computers, neural network • Ovary, neoplasms, 852.30 • Ovary, US, 852.1298, 852.12983 • Ultrasound (US), Doppler studies, 852.12983


   Introduction
TOP
 Abstract
 Introduction
MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 References
 
The introduction of high-frequency transvaginal ultrasonography (US) has allowed detailed morphologic evaluation of ovarian lesions. However, when morphologic features alone are applied to the prediction of ovarian malignancy, there is a tendency to overdiagnose malignant tumors because of a substantial overlap between malignant and benign-appearing masses (1–3). Several studies have reported that adding transvaginal color Doppler imaging with pulsed Doppler spectral analysis might improve the characterization of ovarian masses by means of quantitative blood flow measurements obtained from tumor vessels (4–9).

To predict a dichotomous outcome (eg, presence or absence of a disease) from a given set of variables that bear on that outcome, a multivariate logistic regression model can be used. Recently, this approach has been applied to estimate the probability of malignancy in an adnexal mass on the basis of multiple parameters related to demographic and US characteristics (10). An alternative to conventional statistical multivariate methods for classifying data is represented by artificial neural networks (ANNs). They consist of a variable number of processing elements (artificial neurons or nodes) connected together in hierarchical layers: an input layer, one or more hidden layers, and an output layer. Each node, with the exception of the input neurons, receives multiple weighted inputs and produces an output that is usually a nonlinear function of the inputs. Unlike rule-based methods, ANNs are capable of learning from examples and generalizing; they have the power of pattern recognition and discrimination tasks. The most widely used classes of ANNs are back-propagation networks (11), which have been used in various clinical diagnostic problems (12,13). One of the areas to which the networks have been extensively applied is imaging interpretation in radiology (14–18).

The aim of the present study was to compare the diagnostic accuracy of ANNs with that of multiple logistic regression (MLR) analysis for predicting ovarian malignancy in patients with adnexal masses by using B-mode and color Doppler transvaginal US.


   MATERIALS AND METHODS
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 References
 
The patient population comprised 207 women who underwent transvaginal US examinations before surgery for 226 adnexal masses. One hundred forty-six women were premenopausal (median age, 36 years) and 61 were postmenopausal (median age, 57 years). Nineteen patients had bilateral adnexal masses. None of the women had a benign tumor on one side and a malignant one on the other. All the adnexal masses were included in the study. Transvaginal US was performed with an Asynchronous AU 590 (Hitachi, Tokyo, Japan) or an AU 4 Idea (Esaote, Genoa, Italy) equipped with a 6.5-MHz intravaginal transducer and a multifrequency (5–7.5-MHz) intravaginal transducer, respectively. In women in whom vaginal scanning could not image the whole tumor, transabdominal scanning was also performed. In the color mode, the lowest possible measurable velocity is 1 cm/sec. The high-pass filter was set at 50–100 Hz. All the scanning examinations were performed by the same operator (R.B.) within 1 month before surgery. Tumor size and morphology were initially assessed by means of conventional real-time B-mode scanning. The mean diameter of each mass was derived from the diameters in the three perpendicular planes. The morphologic parameters recorded for all the masses were represented by locularity (unilocular or multilocular); presence of papillary projections; and overall echogenicity of the lesion, that is, anechoic or uniform or random echogenicity. A careful survey of the vessels was then performed on the entire mass by using color Doppler imaging, and a pulsed Doppler cursor was placed on the areas of increased flow, where the colored hot spots were noted. The arterial waveform was analyzed spectrally, and from the envelope of the Doppler waveform, the peak systolic velocity (19) and the resistance index (20), defined as the difference between the peak systolic and end-diastolic flow velocities divided by the peak systolic velocity, were calculated. The resistance indexes derived from the flow velocity waveforms with the highest peak systolic velocities were used to characterize the tumor. The Doppler examinations were performed in the proliferative phase of the menstrual cycle in premenopausal patients to avoid the presence of the corpus luteum.

The distributions of morphologic and Doppler-related characteristics in benign and malignant tumor groups were compared by using the Mann-Whitney U test for continuous variables and the {chi}2 test for categoric variables.

The data were divided into training and testing sets by using a "leave n out method" (21). In this method, 80% of all the cases were randomly chosen for generating a training set; the 20% that was left out was then used for cross-validation. This procedure was repeated five times so that each case appeared once in a testing subset. The training subsets were used to compute the optimum regression equations and to train the ANNs. The cross-validation subsets were used to estimate the performance of each of the two models to predict ovarian malignancy when they were applied to a different sample from the one on which the method was trained.

MLR Analysis
The seven predictor variables used for MLR analysis were women's ages, mean diameter of each mass, presence of multilocularity, presence of papillary projections, presence of random echogenicity, peak systolic velocity, and resistance index. Qualitative variables were coded in binary. If the variable was absent, it was assigned a value of 0; if it was present, it was assigned a value of 1. When the blood flow was not detectable, the peak systolic velocity and resistance index were set to 6 cm/sec and 0.90, respectively, for inclusion in multivariate analysis. These values were derived from the extremes of the distributions for benign tumors, because no malignancies were encountered in this group. Histologic diagnosis of malignancy in adnexal masses was entered as a dependent variable in the MLR model and was coded as 0 for absent and 1 for present. A forward stepwise procedure was used to select the independent variables that should have been included in the model to predict the dependent variable. A variable was entered into the model if the probability of its score statistics was less than .05. A variable was removed if its likelihood ratio statistic had a P value of greater than or equal to .10. The reason why the significance level to remove a variable from the equation must be slightly higher than that to include a variable is that this makes it possible to avoid infinite loops. The regression equation derived from each of the training subsets was applied to the just-removed cross-validation subset to estimate the individual probability of having a malignant ovarian tumor. This probability (Q) was calculated as follows (22):

where e is the base of natural logarithms and Z is the linear combination calculated as Z = {alpha} + ß1x1 + ß2x2 + . . . + ßnxn, in which {alpha}, which is the intercept, and ßn are coefficients estimated from the data, and xn is the predictor variable included in the model. In particular, the exponential form of ßi (ie, eßi) was the factor by which the odds of malignancy changed when the ith predictor variable increased by one unit. The probability Q estimated by using the logistic model always ranges between 0 and 1, regardless of the value of Z. In this study, if Q was greater than 0.5, the case was classified as malignant. The converse applied when the estimated probability was less than 0.5.

ANN Topology
ANNs were constructed by using the commercially available software BRAINMAKER (version 3.0; California Scientific Software, Nevada City, Calif). We used networks with three layers and a back-propagation algorithm. A logistic function was used as a transfer function for each of the nodes in the hidden layer and the output layer of the networks. This is a nonlinear function expressed as:

where Tf({sum}wiai) is the output activity of the neuron i and wiai is the net input to the neuron, which was derived by multiplying the connection weight (w) and the incoming activation (a) of the unit. The input layer of the networks consisted of the same variables selected by using stepwise logistic regression analysis. A hidden layer of three neurons was connected to the input layer. No more than three hidden neurons were used because this would determine an artificially high classification rate, taking into account the sample size of our study population (23). The hidden layer was connected to an output layer of a single neuron producing a normalized value in the range of 0 to 1, which could be viewed as representing the likelihood of ovarian malignancy. The training parameters of learning rate and momentum were set at 0.25 and 0.90, respectively. A training error tolerance of 0.10 was used; this means that if the networks' output was not within 0.1 of the target output, it was considered incorrect. This error value was used by the back-propagation algorithm to adjust the network weights to minimize the output error during the training. We took care not to overfit the training data, which happens when the network fits the training data too closely and it cannot generalize new data well. In brief, to prevent this problem, three methods may be applied: first, limiting the number of hidden neurons; second, constraining the network not to use large weights; or, third, stopping the training at an earlier cycle than that corresponding to the minimum training error.

Evaluation of Performance
Overall rates for sensitivity and specificity obtained from logistic regression models and ANNs at a cutoff value for outputs greater than 0.50 were calculated as the average of the individual cross-validation rate for each testing subset. These rates were compared by means of the McNemar test (24), a nonparametric test for two related dichotomous variables. The performances of the two classification procedures were compared statistically by using the Brier score (25), that is, the mean of the squared differences between the probabilities estimated by using a probability model and the actual value. The lower the score, the better the performance. The Brier score is independent of the classification threshold used, and, as previously reported (26,27), it represents a better test of predictive accuracy of a classifier than the receiver operating characteristic, or ROC, curve analysis. The scores calculated in each testing subset, by using MLR and ANNs, were compared by means of the Student t test for paired samples.


   RESULTS
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
 RESULTS
DISCUSSION
 References
 
A total of 226 adnexal masses in 207 patients were analyzed: Fifty-one were malignant and 175 were benign tumors at histopathologic examination (Table 1). The median age of the patients with benign tumors was 36 years, which is significantly lower than that of the patients with malignant tumors, whose median age was 51 years (P < .001). Morphologic findings and color Doppler flow findings in benign tumors and malignant tumors are shown in Table 2, along with the two-tailed probabilities for evaluation with the univariate tests (Mann-Whitney U test or {chi}2 test) of the null hypothesis of no difference between the two groups with respect to each parameter. The median diameter was 53.3 mm in benign lesions and 64.7 mm in malignant lesions (P < .001). Malignancy in the adnexal masses was significantly associated with the presence of papillary formations, random echogenicity, and loculations. Color flow examinations depicted intratumoral blood vessels in all malignant masses and in 134 of the 175 benign tumors (76.6%). There were significant differences in the Doppler values between the benign cases and the malignant cases.


View this table:
[in this window]
[in a new window]
  		TABLE 1. Histopathologic Diagnosis of the 226 Adnexal Masses
 

View this table:
[in this window]
[in a new window]
  		TABLE 2. Characteristics of Ovarian Tumors
 
Stepwise logistic regression analysis of the five training subsets resulted in selection of five of the seven predictor variables as the most discriminating; the mean diameter of each mass and multilocularity were not included in any model. From the selected variables, no significant bivariate interaction terms were identified.

The ANNs we used included five input neurons, each of which corresponded to a predictor variable selected by using MLR; three hidden neurons; and one output unit. The networks had 18 connection weights that consisted of 15 weights toward the hidden layer and three weights toward the output neuron.

When logistic regression models were validated against the testing subsets, an overall sensitivity of 84% (43 of 51; 95% CI: 71.4%, 92.9%) and specificity of 96.6% (169 of 175; 95% CI: 92.7%, 98.7%) were obtained with a cutoff value greater than a 50% probability. At the same threshold, the networks had a significantly higher sensitivity of 96% (49 of 51; 95% CI: 86.5%, 99.5%; McNemar test, P = .04), with a specificity of 97.7% (171 of 175; 95% CI: 94.2%, 99.4%; P = .48).

The Brier scores for ANNs in the testing subsets were significantly lower than those calculated for MLR (mean values, 0.031 vs 0.051; two-tailed Student t test, P = .004), thus showing that ANNs classified significantly better than use of MLR did.


   DISCUSSION
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
References
 
Our results suggest that an ANN classifier has potential for categorizing adnexal masses as either malignant or benign tumors on the basis of multiple variables related to demographic and US features.

The distinction between benign and malignant ovarian lesions is important in planning the correct treatment by identifying those patients who require extensive surgery versus those in whom conservative management or minimally invasive surgery can be safely performed. The use of US morphologic criteria alone to differentiate between benign and malignant ovarian neoplasm is hindered by a high false-positive rate, with a consequent high number of invasive procedures required to confirm the abnormal US findings (1–3). The more recent introduction of color Doppler imaging can contribute to the differential diagnosis of adnexal masses (4–9), particularly in postmenopausal women. The most widely used method of interpretation of Doppler parameters (eg, resistance index or peak systolic velocity) is the concept of cutoff values, which are intended to help discriminate ovarian neoplasms by assigning a fixed limit of abnormality to the parameters. Thus, continuous variables are reduced to binary variables, with a potential lack of information contained in the data. This might explain the substantial variation in results reported by different authors who have evaluated the diagnostic accuracy of Doppler indexes alone in differentiating benign from malignant ovarian tumors (5,6,9,28). On the whole, it seems clear that an improvement in predictive accuracy can be achieved only by taking into account multiple parameters. A standard computational method for predicting a diagnosis from multivariate parameters, as well as identifying the variables useful in making the prediction, is represented by MLR. Furthermore, when continuous variables such as age or resistance index are included in this model, any change in these variables will be reflected in the final prediction, without a fixed cutoff of abnormality being needed. By using this approach, we obtained an overall sensitivity of 84% and a specificity of 96.6% on the basis of multiple variables related to demographic and US characteristics. These findings are in line with those previously reported by Tailor et al (10). A back-propagation ANN presents a potential method for finding a relationship between different input variables and binomial output variables. In the current study, we used simple three-layer networks to predict the presence of malignancy in an adnexal mass by using the same input variables selected by means of MLR. The ANNs correctly diagnosed 96.1% of malignant and 97.7% of benign lesions at testing. These results compare favorably with the corresponding rates obtained with MLR models. However, these rates are only a crude indicator of the quality of a classifier, whose predictive accuracy can be better quantified by using a test such as the Brier index, which summarizes how close each patient's prediction is to the true outcome (26,27). Evaluation of these scores in our testing subsets suggests that the networks were able to categorize adnexal masses more accurately than logistic regression models. Thus, ANNs seem to be able to extract information from the data that are not apparent to the logistic regression analysis. In general, ANNs have an inherent advantage over conventional statistical models, which is related to their capability of generating decision boundaries between overlapping diagnostic categories, when the relation between data items is determined by means of a complex multidimensional nonlinear function. The data set studied here was relatively simple, so it is conceivable that the advantage of ANN would become more apparent if other predictor variables were incorporated into the model. However, to evaluate whether a better predictive accuracy of the network might be achieved by adding other variables, it will be necessary to gather an adequate sample of data. Indeed, data sets with small sample size in relation to the number of input categories increase the possibility of a high classification rate by chance (29).

We acknowledge that ANNs are not the panacea for all complex data analysis, and several criticisms may be leveled at these techniques. These include the empirical nature of choosing network parameters and mainly the inability of the networks to explain themselves. However, in our opinion, an ANN that is nothing other than a nonlinear regression model can be used as a decision-support device, provided that its predictive ability is clearly demonstrated independently of a thorough understanding of its inner workings.

The purpose of this preliminary study was not to obtain a trained ANN for clinical application but to evaluate the feasibility of this approach. Although our results seem encouraging, further studies are needed to assess prospectively the performance of an ANN in predicting malignancy in adnexal masses as compared with the performance of traditional multivariate statistical methods. If such a study is successful, the possibility for this computational method to be applied to the clinical practice could be explored.


   Footnotes
 
Address reprint requests to R.B., Via Ghibellina 61, 50122 Florence, Italy.

Abbreviations: ANN = artificial neural network MLR = multiple logistic regression

Author contributions: Guarantors of integrity of entire study, R.B., G.G.; study concepts, R.B., G.G.; study design, R.B., E.V.; definition of intellectual content, R.B., G.G.; literature research, E.V.; data acquisition and analysis, R.B., C.D.; statistical analysis, R.B.; manuscript preparation, R.B., E.V.; manuscript editing and review, E.V.

Received March 18, 1998; revision requested May 22, 1998; revision received July 6, 1998; accepted September 8, 1998.
   References
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 References
 

  1. Sassone AM, Timor-Tritsch IE, Artner A, Westhoff C, Warren W. Transvaginal sonographic characterization of ovarian disease. Obstet Gynecol 1991; 78:70-76.[Medline]
  2. De Priest PD, Shenson D, Fried A, et al. A morphology index based on sonographic findings in ovarian cancer. Gynecol Oncol 1993; 51:7-11.[Medline]
  3. Van Nagell JR, De Priest PD, Puls LE, et al. Ovarian cancer screening in asymptomatic postmenopausal women by transvaginal sonography. Cancer 1991; 68:458-462.[Medline]
  4. Bourne TH, Campbell S, Steer C, Whitehead MI, Collins WP. Transvaginal colour flow imaging: a possible new screening technique for ovarian cancer. BMJ 1989; 299:1367-1370.
  5. Kurjak A, Zalud I, Alfirevic Z. Evaluation of adnexal masses with transvaginal color ultrasound. J Ultrasound Med 1991; 10:295-299.[Medline]
  6. Fleischer AC, Rogers WH, Rao BK, Kepple DM, Jones HW. Assessment of ovarian tumor vascularity with transvaginal color Doppler sonography. J Ultrasound Med 1991; 10:563-568.[Abstract]
  7. Timor-Tritsch IE, Lerner JP, Monteagudo A, Santos R. Transvaginal ultrasonographic characterization of ovarian masses by means of color flow-directed Doppler measurements and a morphologic scoring system. Am J Obstet Gynecol 1993; 168:909-913.[Medline]
  8. Bromley B, Goodman H, Benacerraf BR. Comparison between sonographic morphology and Doppler waveform for the diagnosis of ovarian malignancy. Obstet Gynecol 1994; 83:434-437.[Medline]
  9. Hata K, Hata T, Kitao M. Intratumoral peak velocity as a new possible predictor for detection of adnexal malignancy. Am J Obstet Gynecol 1995; 172:1496-1500.[Medline]
  10. Tailor A, Jurkovic D, Bourne TH, Collins WP, Campbell S. Sonographic prediction of malignancy in adnexal masses using multivariate logistic regression analysis. Ultrasound Obstet Gynecol 1997; 10:41-47.[Medline]
  11. Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating errors. Nature 1986; 323:533-536.
  12. Baxt WG. Application of artificial neural networks to clinical medicine. Lancet 1995; 346:1135-1138.[Medline]
  13. Dybowski R, Gant V. Artificial neural networks in pathology and medical laboratories. Lancet 1995; 346:1203-1207.[Medline]
  14. Asada N, Doi K, MacMahon H, et al. Potential usefulness of an artificial neural network for differential diagnosis of interstitial lung disease: pilot study. Radiology 1990; 177:857-860.[Abstract/Free Full Text]
  15. Boone JM, Gross GW, Greco-Hunt V. Neural networks in radiologic diagnosis. I. Introduction and illustration. Invest Radiol 1990; 25:1012-1016.[Medline]
  16. Wu Y, Giger ML, Doi K, Vyborny CJ, Schmidt RA, Metz CE. Artificial neural networks in mammography: application to decision making in the diagnosis of breast cancer. Radiology 1993; 187:81-87.[Abstract/Free Full Text]
  17. Lo JY, Baker JA, Kornguth PJ, Iglehart JD, Floyd CE, Jr. Predicting breast cancer invasion with artificial neural networks on the basis of mammographic features. Radiology 1997; 203:159-163.[Abstract/Free Full Text]
  18. Da Ponte JS, Sherman P. Classification of ultrasonic image texture by statistical discrimination analysis of neural networks. Comput Med Imaging Graph 1991; 15:3-9.[Medline]
  19. Gosling RG, King DH. Ultrasound angiology. In: Macus AW, Adamson L, eds. Arteries and veins. Edinburgh, Scotland: Churchill-Livingstone, 1975; 61-98.
  20. Pourcelot L. Applications cliniques de l'examen Doppler transcutané. In: Peronneau P, eds. Velocimetrie ultrasonore Doppler. Paris, France: Seminare Institut National de la Santé et de la Recherche Médicale, 1975; 213-240.
  21. Lachenbruch P. An almost unbiased method for obtaining confidence intervals for the probability of misclassification in discriminant analysis. Biometrics 1967; 23:639-645.[Medline]
  22. Hosmer DW, Lemeshow S. Applied logistic regression New York, NY: Wiley, 1989.
  23. Baum EB, Haussler D. What size net gives valid generalization? In: Touretzky DS, ed Advances in neural information processing systems. Vol 1. San Mateo, Calif: Kauffman, 1989; 81-90.
  24. Siegel S, Castellan NJ, Jr. Nonparametric statistics for the behavioral sciences New York, NY: McGraw-Hill, 1988; 63-67.
  25. Winkler RL, Murphy AH. "Good" probability assessors. J Appl Meteorol 1968; 7:751-758.
  26. Redelmaier DA, Bloch DA, Hickam DH. Assessing predictive accuracy: how to compare Brier scores. J Clin Epidemiol 1991; 44:1141-1146.[Medline]
  27. Gurney JW. Neural networks at the crossroads: caution ahead. Radiology 1994; 193:27-30.[Free Full Text]
  28. Tekay A, Jouppila P. Validity of pulsatility and resistance indices in classification of adnexal tumors with transvaginal color Doppler ultrasound. Ultrasound Obstet Gynecol 1992; 2:338-344.[Medline]
  29. Fletcher JM, Rice WJ, Ray RM. Linear discriminant function analysis in neuropsychological research: some uses and abuses. Cortex 1978; 14:564-577.[Medline]



This article has been cited by other articles:


Home page
 J Ultrasound MedHome page
P. Yoruk, O. Dundar, B. Yildizhan, L. Tutuncu, and T. Pekin
Comparison of the Risk of Malignancy Index and Self-Constructed Logistic Regression Models in Preoperative Evaluation of Adnexal Masses
J. Ultrasound Med., October 1, 2008; 27(10): 1469 - 1477.
[Abstract] [Full Text] [PDF]

Home page
 Clin. Cancer Res.Home page
C. Van Holsbeke, B. Van Calster, L. Valentin, A. C. Testa, E. Ferrazzi, I. Dimou, C. Lu, P. Moerman, S. Van Huffel, I. Vergote, et al.
External Validation of Mathematical Models to Distinguish Between Benign and Malignant Adnexal Tumors: A Multicenter Study by the International Ovarian Tumor Analysis Group
Clin. Cancer Res., August 1, 2007; 13(15): 4440 - 4447.
[Abstract] [Full Text] [PDF]

Home page
 RadiologyHome page
S. H. Kim, J. M. Lee, J. H. Kim, K. G. Kim, J. K. Han, K. H. Lee, S. H. Park, N.-J. Yi, K.-S. Suh, S. K. An, et al.
Appropriateness of a Donor Liver with Respect to Macrosteatosis: Application of Artificial Neural Networks to US Images--Initial Experience
Radiology, March 1, 2005; 234(3): 793 - 803.
[Abstract] [Full Text] [PDF]

Home page
 Ann. Surg. Oncol.Home page
T. Kan, Y. Shimada, F. Sato, T. Ito, K. Kondo, G. Watanabe, M. Maeda, S. Yamasaki, S. J. Meltzer, and M. Imamura
Prediction of Lymph Node Metastasis with Use of Artificial Neural Networks Based on Gene Expression Profiles in Esophageal Squamous Cell Carcinoma
Ann. Surg. Oncol., December 1, 2004; 11(12): 1070 - 1078.
[Abstract] [Full Text] [PDF]

Home page
 Cancer Res.Home page
Y. Xu, F. M. Selaru, J. Yin, T. T. Zou, V. Shustova, Y. Mori, F. Sato, T. C. Liu, A. Olaru, S. Wang, et al.
Artificial Neural Networks and Gene Filtering Distinguish Between Global Gene Expression Profiles of Barrett's Esophagus and Esophageal Cancer
Cancer Res., June 1, 2002; 62(12): 3493 - 3497.
[Abstract] [Full Text] [PDF]

Home page
 RadiologyHome page
K. Kinkel, H. Hricak, Y. Lu, K. Tsuda, and R. A. Filly
US Characterization of Ovarian Masses: A Meta-Analysis
Radiology, December 1, 2000; 217(3): 803 - 811.
[Abstract] [Full Text]

Home page
 Arch SurgHome page
R.-F. Chang, W.-J. Kuo, D.-R. Chen, Y.-L. Huang, J.-H. Lee, and Y.-H. Chou
Computer-Aided Diagnosis for Surgical Office-Based Breast Ultrasound
Arch Surg, June 1, 2000; 135(6): 696 - 699.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Biagiotti, R.
Right arrow Articles by Gacci, G.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Biagiotti, R.
Right arrow Articles by Gacci, G.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
RADIOLOGY RADIOGRAPHICS RSNA JOURNALS ONLINE