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


     

This Article
Right arrow Abstract Freely available
Right arrow Figures Only
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 Chen, D.-R.
Right arrow Articles by Huang, Y.-L.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Chen, D.-R.
Right arrow Articles by Huang, Y.-L.
Related Collections
Right arrowRelated Article
(Radiology. 1999;213:407-412.)
© RSNA, 1999

Computer Applications

Computer-aided Diagnosis Applied to US of Solid Breast Nodules by Using Neural Networks1

Dar-Ren Chen, MD, Ruey-Feng Chang, PhD and Yu-Len Huang, PhD

1 From the Department of General Surgery, China Medical College and Hospital, 2 Yer-Der Rd, Taichung, Taiwan, Republic of China (D.R.C.), and the Department of Computer Science and Information Engineering, National Chung Cheng University, Chiayi, Taiwan, Republic of China (R.F.C., Y.L.H.). Received May 16, 1998; revision requested July 30; final revision received April 22, 1999; accepted June 8. Address reprint requests to D.R.C. (e-mail: dlchen88@ms13.hinet.net).


   Abstract
TOP
 Abstract
Introduction
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 References
 
PURPOSE: To increase the capabilities of ultrasonographic (US) technology for the differential diagnosis of solid breast tumors by using a neural network.

MATERIALS AND METHODS: One hundred forty US images of solid breast nodules were evaluated. When a sonogram was obtained, an analog video signal from the VCR output of the scanner was transmitted to a notebook computer. A frame grabber connected to the printer port of the computer was then used to digitize the data. The suspicious tumor region on the digitized US image was manually selected. The texture information of the subimage was extracted, and a neural network classifier with autocorrelation features was used to classify the tumor as benign or malignant. In this experiment, 140 pathologically proved tumors (52 malignant and 88 benign tumors) were sampled with k-fold cross-validation (k = 10) to evaluate the performance with receiver operating characteristic curves.

RESULTS: The accuracy of neural networks for classifying malignancies was 95.0% (133 of 140 tumors), the sensitivity was 98% (51 of 52), the specificity was 93% (82 of 88), the positive predictive value was 89% (51 of 57), and the negative predictive value was 99% (82 of 83).

CONCLUSION: This system differentiated solid breast nodules with relatively high accuracy and helped inexperienced operators to avoid misdiagnoses. Because the neural network is trainable, it could be optimized if a larger set of tumor images is supplied.

Index terms: Breast neoplasms, diagnosis, 00.30 • Breast neoplasms, US, 00.1298 • Computers, diagnostic aid • Images, analysis • Images, interpretation • Ultrasound (US), tissue characterization, 00.1298


   Introduction
TOP
 Abstract
 Introduction
MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 References
 
Breast cancer ranks first in the causes of cancer deaths among women in developed countries and is second to cervical cancer in developing countries (1). The best way to reduce deaths due to breast cancer is to treat the disease at an earlier stage. Earlier treatment requires early diagnosis, and early diagnosis requires an accurate and reliable diagnostic procedure that allows physicians to differentiate benign breast tumors from malignant ones. Current procedures for detecting and diagnosing breast cancers, however, illustrate the difficulty in maximizing both sensitivity and specificity.

Although the smaller hospitals use conventional ultrasonographic (US) equipment to obtain medical images quickly, estimates of the accuracy of US diagnostic methods are controversial, and the role of US in current practice is not yet defined (2,3). One of the important reasons for this controversy is the considerable overlap of benign and malignant findings on US images; interpretation is subjective and depends on the operator.

The US examination described by Stavros et al (4) is much more extensive than the usual examinations performed at most breast imaging centers. In their study, Stavros et al (4) found that the sensitivity of breast US for malignancy was 98.4% (123 of 125 findings), the specificity was 67.8% (424 of 625), the overall accuracy was 72.9% (547 of 750), the positive predictive value was 38.0% (123 of 324), and the negative predictive value was 99.5% (424 of 426). However, these improved diagnostic results were achieved by experienced radiologists. In practice, many invasive diagnostic procedures are still required in most cases. Most of these procedures are avoidable because the rate of positive findings at biopsy for cancer is low—between 10% and 31% (5–7). To overcome the aforementioned shortcomings of US and to fully expand the capability of US, only an efficient computerized model offers objective evidence and avoids interobserver variations; therefore, an optimal and stable high diagnostic rate can be achieved by using such a model.

Both mammography (8–11) and US (12) can be used to detect and classify breast tumors. Although mammography can be used to visualize nonpalpable and small tumors, US is a convenient and safe tool to use in the classification of tumors, especially palpable tumors. Garra et al (13) suggest that analysis of US image texture is a simple means of markedly reducing the number of biopsies performed for benign lesions. Garra et al (13) used a co-occurrence matrix and the linear Bayesian classifier. Neural network techniques have been applied to detect microcalcifications and have even been applied to distinguish benign and malignant microcalcifications on digital mammographic images (9–11). Moreover, Choong et al (14) used neural networks to make prognoses for women with breast cancer.

The multilayer feed-forward neural network can be used to extract higher-order statistics by adding one or more hidden layers. This model has become extremely popular in terms of classification and prediction. The powerful error back-propagation algorithm proposed by Rumelhart el al (15) and Hirose et al (16) is the most widely used algorithm for multilayer feed-forward neural networks. We used this neural network model as a classifier to determine whether the breast tumors were benign or malignant. The diagnostic model proposed herein can exploit the nonlinear property of the learning algorithm of the neural network to classify the US images of the breast more accurately. The multilayer feed-forward neural network is a reliable choice for the proposed model because it trains well and computes efficiently.


   MATERIALS AND METHODS
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 References
 
In general, a physician can easily identify a tumor on a US image by the shape of the tumor and by the contrast of internal echoes, but the computer cannot easily segment a tumor from a digitized image. However, if a physician has already identified the tumor, a computer can be used to diagnose the tumor with information about intensity variation and texture on the image. In this study, a physician (D.R.C.) first extracted the subimage of the region of interest, and then the computer was used to analyze the subimage by using the information about intensity variation and texture to make a differential diagnosis.

Data Acquisition
The US image database contained 140 images of pathologically proved tumors: benign breast tumors from 88 patients and carcinomas from 52 patients (tumor size, >1 cm in all patients). The database contained only one image from each patient. The US images that depicted the largest diameter of the tumor were captured. The images were collected from January 1, 1997, to May 31, 1998, and the patients' ages ranged from 18 to 62 years. US was performed by using an Aloka SSD 1200 scanner (Tokyo, Japan) and a 7.5-MHz linear transducer with freeze-frame capability. No acoustic standoff pad was used with any of the patients. The US gain setting remained unchanged throughout the entire study, except for changes made to obtain the best view.

The following description explains how digital US images were obtained. When a sonogram was obtained, an analog video signal was transmitted from the VCR output of the scanner to a notebook computer. The data were then digitized by means of a Video CATcher frame grabber (Top Solution Technology; Taipei, Taiwan, Republic of China) that was connected to the printer port of the computer. The capturing resolutions of the portable computer and the external frame grabber was 736 x 566 pixels for a National Television Systems Committee video-screen picture. We used the ProImage software package from Prolab (version 2.0; Taipei, Taiwan, Republic of China), which was bundled with the frame grabber, to capture the real-time digital image. The monochrome US image was quantized to 8 bits (ie, 256 gray levels). The subimage of the region of interest was manually selected by using the ProImage (Prolab) package. That is, the software package was used to capture the full image from the US scanner and to select the region-of-interest subimage manually. The region-of-interest subimage was then saved as a file for later analysis with the neural network program. Figure 1 illustrates a real-time digitized monochrome US image of a malignant tumor, with 58 x 58 pixels in a 1 x 1-cm rectangle (Fig 1, A) and an image of a tumor (approximately 3.1 x 1.5 cm) that was captured with a resolution of 179 x 88 pixels (Fig 1, B).



View larger version (48K):
[in this window]
[in a new window]
  		Figure 1. A, Transverse digital US image depicts a malignant tumor, with a resolution of 736 x 556 pixels. Note that there are 58 x 58 = 3,364 pixels in a 1 x 1-cm rectangle. B, The region-of-interest rectangle is approximately 3.1 x 1.5 cm and is captured with a resolution of 179 x 88 pixels. ROI = region of interest.

 
Image Analysis
A US image consists of many points with different values of gray-level intensity. Different tissues have markedly different textures. Benign tumors are classically described as regular masses with homogeneous internal echoes, but carcinomas are described as masses with fuzzy borders and heterogeneous internal echoes. Figure 2 illustrates digital images of a benign tumor and a malignant tumor.



View larger version (88K):
[in this window]
[in a new window]
  		Figure 2a. Typical transverse US region-of-interest subimages in the breast depict (a) a benign lesion (arrows) and (b) a malignant lesion (arrows).

 


View larger version (97K):
[in this window]
[in a new window]
  		Figure 2b. Typical transverse US region-of-interest subimages in the breast depict (a) a benign lesion (arrows) and (b) a malignant lesion (arrows).

 
We used the correlation between neighboring pixels on the images as classifying features of the tumor. The normalized autocorrelation coefficients (17) can be used to reflect the interpixel correlation on an image. The two-dimensional normalized autocorrelation coefficient {gamma} between pixel (i, j) and pixel (i + {Delta}m, j + {Delta}n) on an image with size m x n can be defined as

where

Moreover, the two-dimensional autocorrelation coefficients were further modified to a mean-removed version to generate the similar autocorrelation features for images with different brightness but with similar texture. This modified version was expressed as

where was the mean value of f(x, y). The absolute value was used in Equation (3) because a negative value may be produced when the mean is subtracted from the gray level of a pixel. We determined these two-dimensional autocorrelation coefficients for each US image of a breast tumor and used these coefficients as the interpixel features to distinguish between benign and malignant tumors.

Neural Network Classification
A multilayer feed-forward neural network contains one or more hidden layers. The function of neurons in the hidden layer is to arbitrate between the input and the output of the neural network. The input feature vector is fed into the source nodes in the input layer of the neural network. The neurons of the input layer constitute the input signals applied to the neurons in the hidden layer. The output signals of the hidden layer can be used as inputs to the next hidden or output layer. Finally, the output layer produces the output result and terminates the neural computing process.

Among the algorithms used to design the multilayer feed-forward neural networks, the back-propagation algorithm is the most popular. In general, there are two different phases in the back-propagation algorithm: the forward phase and the backward phase. In the forward phase, the input signals are computed and passed through the neural network, layer by layer. Then, the neurons in the output layer produce the output signals of the neural network. During this phase, comparing the output response of the neural network with the desired response can generate error signals.

During the backward phase of the back-propagation algorithm, some free parameters can be adjusted by referring to the error signals. This work can be used to minimize the distortion of the neural network. The multilayer feed-forward neural network has a high capability for learning and for computational efficiency. We iteratively executed the back-propagation learning algorithm for the training set and then produced the synaptic weight vectors that were applied to the neural network. We classified the benign and malignant tumors in the diagnostic model by applying the final synaptic weight vectors to the multilayer feed-forward neural network.

We used the modified version of the two-dimensional normalized autocorrelation matrix for the input of the neural network. The dimension of the matrix can be fixed for an image of any size. In this study, both {Delta}m and {Delta}n were 5, so processing a US image produced a 5 x 5 autocorrelation matrix (ie, 25 autocorrelation coefficients). The value of {gamma}(0, 0) was always 1 for a normalized autocorrelation matrix. Thus, except for the element {gamma}(0, 0), other autocorrelation coefficients were formed as a 24-dimensional image feature vector.

We used a multilayer feed-forward neural network with 25 input nodes, 10 hidden nodes, and one output node, as illustrated in Figure 3. The 24-dimensional image feature vector and a predefined threshold of the input layer were used as the input signals for the neural network. Moreover, the value produced by the output node was used to decide whether a tumor was benign or malignant. The output value of the neural network was either 0 or 1. When the output value of a US breast image was near enough to 1, the system classified the tumor on the image as malignant. Conversely, when the output value was close to 0, the system classified the tumor as benign. Figure 4 illustrates the structure of the neural network model for tissue classification. Figure 5 illustrates the program of the neural network diagnostic model. In this program, the output of the neural network was multiplied by 100.



View larger version (18K):
[in this window]
[in a new window]
  		Figure 3. Structural graph depicts the multilayer feed-forward neural network used in this study. Numbers indicate the number of nodes.

 


View larger version (22K):
[in this window]
[in a new window]
  		Figure 4. Diagram depicts the structure of the neural network model for tissue classification. ROI = region of interest, 2D = two-dimensional.

 


View larger version (55K):
[in this window]
[in a new window]
  		Figure 5. Illustration of the program for the neural network diagnostic model. The weights of the neural network are loaded, and then the region-of-interest subimage file is downloaded. If the output of neural network is larger than a predefined threshold of 20, the result will be "probably malignant." If the output is less than the threshold, the result will be "probably benign." Note that the output of the neural network is multiplied by 100. NN = neural network, ROI = region of interest.

 
Training and Testing
The receiver operating characteristic (ROC) curve and the k-fold cross-validation method (18) were used to estimate the performance of the neural network. In this study, the software package LABROC1 developed by Professor C. E. Metz, University of Chicago, Ill, was used to fit the ROC curve. With the k-fold cross-validation method, the 140 images in the database were randomly divided into k groups. The first group was set aside, and the remaining (k - 1) groups were used to train the neural network. The training procedure was stopped when the improvement of error distortion was smaller than 1 x 10-6. Moreover, the maximal number of iterations was limited to 10,000. Once trained, the network was then tested with the group that was set aside. The second group was then removed, the remaining (k - 1) groups were trained, and the network was tested with the excluded group. This process was repeated until all k groups were used, in turn, as the group that was set aside and used for testing. In the simulations, k was 10, and each group had 14 images.


   RESULTS
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
 RESULTS
DISCUSSION
 References
 
Table 1 lists the mean and SD of the autocorrelation coefficients for malignant and benign tumors and the mean difference between the two groups. The substantial mean differences of the two groups were used to prove whether autocorrelation coefficients were good features for distinguishing malignant and benign tumors. Moreover, with larger {Delta}m and {Delta}n, the mean difference was larger, as illustrated in Figure 6. Table 2 lists the number of training iterations and error distortions for each training set. The error distortion was defined as the absolute difference between the desired output and the actual output of the neural network. Figure 7 illustrates the ROC curve for the neural network in the classification of malignant and benign tumors. The overall performance of the neural network was evaluated by examining the ROC area index AZ over the testing output values. Our method had a high AZ value of 0.9560 ± 0.0183 (SD). Table 3 lists the performance for different threshold values. With a threshold of 0.2, the network correctly identified 51 of 52 malignant tumors and 82 of 88 benign tumors. Table 4 lists the number of tumors that were misdiagnosed in each test set by using the neural network at a threshold of 0.2. The accuracy of the neural network for detecting malignant tumors was 95.0% (133 of 140), the sensitivity was 98% (51 of 52), the specificity was 93% (82 of 88), the positive predictive value was 89% (51 of 57), and the negative predictive value was 99% (82 of 83), as illustrated in Table 5. Table 6 lists the number of tumors of various specific types in this study. From the limited data in this study, the tumor type seemed to have no relationship with the ability of the network to differentiate benign and malignant tumors.


View this table:
[in this window]
[in a new window]
  		TABLE 1. Mean, SD, and Mean Difference between Autocorrelation Coefficients for Malignant and Benign Tumors
 


View larger version (14K):
[in this window]
[in a new window]
  		Figure 6. Plot depicts the mean differences between the autocorrelation coefficients for the malignant and benign tumors. The mean differences are increasing for larger {Delta}m and {Delta}n.

 

View this table:
[in this window]
[in a new window]
  		TABLE 2. Number of Malignant and Benign Tumors, Iterations, and Error Distortions for Each Training Set
 


View larger version (7K):
[in this window]
[in a new window]
  		Figure 7. Plot depicts the ROC curve for the neural network in the classification of malignant and benign tumors. The AZ value of the ROC curve is 0.9560 ± 0.0183. FPF = false-positive fraction, TPF = true-positive fraction.

 

View this table:
[in this window]
[in a new window]
  		TABLE 3. Performance of the Neural Network for Different Threshold Values
 

View this table:
[in this window]
[in a new window]
  		TABLE 4. Number of Misdiagnosed Tumors by Using the Neural Network at a Threshold of 0.2
 

View this table:
[in this window]
[in a new window]
  		TABLE 5. Performance of the Neural Network at a Threshold of 0.2
 

View this table:
[in this window]
[in a new window]
  		TABLE 6. Types of Tumors
 

   DISCUSSION
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
References
 
We proposed an efficient neural network diagnostic system in which interpixel correlations on the US image were used to differentiate benign and malignant tumors. The proposed system was used to diagnose breast tumors. The co-occurrence features and the linear classifier (AZ = 0.91) proposed by Garra et al (13) was replaced with the autocorrelation features and the neural network (AZ = 0.96) to obtain a better result. This system adopted the multilayer feed-forward neural network to classify US tumor images to distinguish benign and malignant tumors more accurately. The back-propagation learning algorithm was used to construct the synaptic weights for the neural network classifier in our diagnostic system.

The architecture of the neural network scheme was simple, redressed easily, and was appropriate for hardware design. Our scheme used only one type of neural network model, that is, the multilayer feed-forward neural network. Moreover, when the performance is suboptimal for new US images, these images can be added to the original training set to produce a new set of synaptic weight vectors by adjusting the free parameters. A new synaptic weight vector for the neural network can control the performance of the system. Notably, the neural network modules can redress synaptic weight vectors without modifying the other functions.

According to the experimental results, the performance of our diagnostic model in making differential diagnoses was very good. These results indicate that benign and malignant tumors can be distinguished by using interpixel correlations on digital US images. From the highly satisfactory specificity and sensitivity of the results, our system is expected to be a useful computer-aided diagnostic tool in the classification of benign and malignant tumors on sonograms. It can provide a second reading to help reduce misdiagnoses. Further studies of larger test sets of tumor images are underway.


   Footnotes
 
See also the editorial by Tourassi (pp 317–320 ) in this issue.

Abbreviation: ROC = receiver operating characteristic

Author contributions: Guarantors of integrity of entire study, D.R.C., R.F.C., Y.L.H.; study concepts, D.R.C.; study design, D.R.C., R.F.C.; definition of intellectual content, D.R.C., R.F.C., Y.L.H.; literature research, D.R.C., R.F.C., Y.L.H.; clinical studies, D.R.C.; experimental studies, Y.L.H., R.F.C.; data acquisition, D.R.C., R.F.C., Y.L.H.; data analysis, Y.L.H., R.F.C.; statistical analysis, D.R.C.; manuscript preparation and review, D.R.C., R.F.C., Y.L.H.; manuscript editing, D.R.C., R.F.C.


   References
TOP
 Abstract
 Introduction
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 References
 

  1. Pisani P, Parkin DM, Ferlay J. Estimates of the worldwide mortality from eighteen major cancers in 1985: implications for prevention and projections of future burden. Int J Cancer 1993; 55:891-903.[Medline]
  2. Blickstein I, Goldchmit R, Strano SD, Goldman RD, Barzili N. Echogenicity of fibroadenoma and carcinoma of the breast: quantitative comparison using gain-assisted densitometric evaluation of sonograms. J Ultrasound Med 1995; 14:661-664/LAST-PAGE>.[Abstract]
  3. Blickstein I, Goldman R. The relation between different system gain settings and the accuracy of densitometric assessment of echogenicity. J Ultrasound Med 1994; 13:675-678.[Abstract]
  4. Stavros AT, Thickman D, Rapp CL, Dennis MA, Parker SH, Sisney GA. Solid breast nodules: use of sonography to distinguish between benign and malignant lesions. Radiology 1995; 196:123-134/LAST-PAGE>.[Abstract/Free Full Text]
  5. Gisvold JJ, Martin JK, Jr. Prebiopsy localization of nonpalpable breast lesions. AJR 1984; 143:477-481.[Abstract/Free Full Text]
  6. Rosenberg AL, Schwartz GF, Feig SA, Patchefsky AS. Clinically occult breast lesions: localization and significance. Radiology 1987; 162:167-170.[Abstract/Free Full Text]
  7. Bassett LW, Liu TH, Giuliano AE, Gold RH. The prevalence of carcinoma in palpable vs impalpable, mammographically detected lesions. AJR 1991; 157:21-24.[Abstract/Free Full Text]
  8. Petrick N, Chan HP, Sahiner B, Wei D. An adaptive density-weighted contrast enhancement filter for mammographic breast mass detection. IEEE Trans Med Imaging 1996; 15:59-67.
  9. Dhawan AP, Chitre Y, Kaiser-Bonasso C, Moskowitz M. Analysis of mammographic microcalcifications using gray-level image structure features. IEEE Trans Med Imaging 1996; 15:246-259.[Medline]
  10. Zheng B, Qian W, Clarke LP. Digital mammography: mixed feature neural network with spectral entropy decision for detection of microcalcifications. IEEE Trans Med Imaging 1996; 15:589-597.[Medline]
  11. Sahiner B, Chan HP, Petrick N, Wei D, Helvi MA, Adler DD, Goodsitt MM. Classification of mass and normal breast tissue: a convolution neural network classifier with spatial domain and texture images. IEEE Trans Med Imaging 1996; 15:598-610.[Medline]
  12. Shankar PM, Reid JM, Ortega H, Piccoli CW, Goldberg BB. Use of non-Rayleigh statistics for identification of tumors in ultrasonic B-scans of the breast. IEEE Trans Med Imaging 1993; 12:687-692.[Medline]
  13. Garra BS, Krasner BH, Horii SC, Ascher S, Muk SK, Zeman RK. Improving the distinction between benign and malignant breast lesions: the value of sonographic texture analysis. Ultrasound Imaging 1993; 15:267-285.[Medline]
  14. Choong PL, deSilva CJS, Dawkins HJS, Sterrett GF. Entropy maximization networks: an application to breast cancer prognosis. Neural Networks 1996; 7:568-577.
  15. Rumelhart DE, Hinton GE, Williams RJ. Learning representation by back-propagation errors. Nature 1986; 323:533-536.
  16. Hirose Y, Yamashita K, Hijiva S. Back-propagation algorithm which varies the number of hidden units. Neural Networks 1991; 4:61-66.
  17. Gonzalez RC, Woods RE. Image compression In: Digital image processing. Reading, Mass: Wesley, 1992; 312-315.
  18. Weiss SM, Kapouleas I. An empirical comparison of pattern recognition neural nets and machine learning classification methods In: Proceedings of the 11th International Joint Conference on Artificial Intelligence. Detroit, Mich: IJCAI, 1989; 234-237.

Related Article

Journey toward Computer-aided Diagnosis: Role of Image Texture Analysis
Georgia D. Tourassi
Radiology 1999 213: 317-320. [Full Text] [PDF]



This article has been cited by other articles:


Home page
 RadiologyHome page
B. Sahiner, H.-P. Chan, M. A. Roubidoux, L. M. Hadjiiski, M. A. Helvie, C. Paramagul, J. Bailey, A. V. Nees, and C. Blane
Malignant and Benign Breast Masses on 3D US Volumetric Images: Effect of Computer-aided Diagnosis on Radiologist Accuracy
Radiology, March 1, 2007; 242(3): 716 - 724.
[Abstract] [Full Text] [PDF]

Home page
 J Ultrasound MedHome page
S.-J. Chen, K.-S. Cheng, Y.-C. Dai, Y.-N. Sun, Y.-T. Chen, K.-Y. Chang, S.-N. Yu, T.-W. Chang, H.-M. Tsai, and C.-C. Hsien
Quantitatively Characterizing the Textural Features of Sonographic Images for Breast Cancer With Histopathologic Correlation
J. Ultrasound Med., May 1, 2005; 24(5): 651 - 661.
[Abstract] [Full Text] [PDF]

Home page
 Arch SurgHome page
D.-R. Chen, R.-F. Chang, W.-M. Chen, and W.-K. Moon
Computer-Aided Diagnosis for 3-Dimensional Breast Ultrasonography
Arch Surg, March 1, 2003; 138(3): 296 - 302.
[Abstract] [Full Text] [PDF]

Home page
 RadiologyHome page
C.-M. Chen, Y.-H. Chou, K.-C. Han, G.-S. Hung, C.-M. Tiu, H.-J. Chiou, and S.-Y. Chiou
Breast Lesions on Sonograms: Computer-aided Diagnosis with Nearly Setting-Independent Features and Artificial Neural Networks
Radiology, February 1, 2003; 226(2): 504 - 514.
[Abstract] [Full Text] [PDF]

Home page
 RadiologyHome page
G. D. Tourassi
Journey toward Computer-aided Diagnosis: Role of Image Texture Analysis
Radiology, November 1, 1999; 213(2): 317 - 320.
[Full Text]

This Article
Right arrow Abstract Freely available
Right arrow Figures Only
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 Chen, D.-R.
Right arrow Articles by Huang, Y.-L.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Chen, D.-R.
Right arrow Articles by Huang, Y.-L.
Related Collections
Right arrowRelated Article

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