Development of a Breast MRI Computer-Aided Diagnosis System

Institution: University of California, Irvine
Investigator(s): Ke Nie, M.S. -
Award Cycle: 2008 (Cycle 14) Grant #: 14GB-0148 Award: $76,000
Award Type: Dissertation Award
Research Priorities
Detection, Prognosis and Treatment>Imaging, Biomarkers, and Molecular Pathology: improving detection and diagnosis



Initial Award Abstract (2008)

Breast MRI has evolved into an established complimentary clinical tool for detection and diagnosis of breast lesions. In March 2007 the American Cancer Society has issued a guideline recommending annual breast MRI screening to women with greater than 20-25% lifetime risk for breast cancer. Unlike the conventional imaging using mammography or ultrasound, MRI acquires a large number of images, and that poses a great difficulty for the radiologists to evaluate all presented information. As more studies are being performed in small community hospitals, lack of sufficient training of breast imagers in reading MRI becomes an issue. The current commercially available Computer-Aided Diagnosis (CAD) systems developed commercially can extract most important information from a large number of images obtained in a single MR examination to help radiologists’ interpretation. But, this approach offers only a display platform, not a true CAD. In contrast, mammographic CAD is a current standard to provide “a second reader”, which has demonstrated a great success to improve diagnostic sensitivity. Despite the urgency and the great need for MR, there is very limited work to develop “a true CAD” which can provide “an intelligent thinking and final impression”.

Our objective is to develop a systematic breast MRI CAD system which can perform lesion detection and characterization to help radiologists’ interpretation, also to provide a “second reader” to improve their diagnostic accuracy. We believe that a CAD system, which integrates all morphology, texture, and enhancement kinetics information, may be able to achieve a high accuracy in detection and diagnosis of breast cancer. The proposed study will use automated computer algorithms to simulate the entire procedure of radiologists’ interpretation for breast MRI. This computer system will detect and segment suspicious region (possible lesion), then analyze their morphologic and kinetic features. A full panel of features to allow differentiating between malignant and benign lesions will be investigated. All together we have 250 malignant cases and 150 benign cases in our first database collected using a 1.5T MRI scanner, which can be used to find the best diagnostic feature set (i.e. malignant signature) for breast cancer. The performance of each step in breast segmentation and lesion segmentation, as well as the final system performance, will be compared to that of experienced radiologists. Finally, a color-coded quantitative labeling scale will be developed to label the suspicious lesion with a color and a BI-RADS (Breast Imaging Reporting and Data System, a standardized classification for mammographic studies) score between 1-5 from normal to highly suspicious of malignancy. Then the developed CAD will be tested in one additional database collected using a 3.0T MRI scanner.

Computer aided diagnosis system has been proven to improve the diagnostic accuracy in mammography, and given the insufficient training of breast imagers in MRI, it is anticipated that a breast MRI CAD will make a crucial contribution. This research will mostly benefits patients who are recommended to undergo MRI, especially those with dense breasts, particularly in young or Asian women.




Final Report (2010)

The goal of this work is to develop computer-aided analysis methods to extract information from breast MRI to contribute in improved management of breast diseases. With nearly 20 years of research, breast MRI has finally evolved from a research tool to an important clinical imaging modality. The American Cancer Society has issued a guideline recommending women with greater than 20% lifetime risk of developing breast cancer to receive annual screening breast MRI. More and more breast MRI examinations are expected to be performed. Considering the overwhelming 3D information, there is a critical need to develop computer-aided diagnosis (CAD) technology to fully utilize the wealth of information that can be provided. However, there is very limited work in developing "a true MRI CAD" in either commercially available products or in the literature.

To respond to this need, we developed a systematic CAD system that is capable of providing intelligent diagnostic impression of breast lesions shown on MRI. Automated computer algorithms were implemented to simulate the entire procedure of radiologists' interpretation for breast MRI. The steps included: 1) segment the breast region for lesion detection, 2) segment the lesion for diagnosis, 3) extract morphological, texture, and enhancement kinetics features, 3) investigate their roles in classifying malignant and benign tumors using artificial neural networks and logistic regression, and 4) based on likelihood of malignancy to label the lesion. The method was applied to build diagnostic models for mass lesions and non-mass-like enhancement lesions separately, and the diagnostic performance evaluated using the Receiver Operating Characteristics (ROC) curve was shown.

We further extended this work besides the detection and diagnosis to another aspect of breast cancer management: risk prevention using computer-aided techniques. A computer-based algorithm was developed to segment the fibroglandular tissue in the breast. MRI provides 3-dimensional images of the breast with a high tissue contrast, which can be used to analyze the morphological distribution of adipose and dense tissue in addition to the amount of dense tissue. Many studies have demonstrated a strong association between the breast density and the cancer risk, and there is also evidence suggesting that the morphological distribution pattern of dense tissue is related to cancer risk. A consensus has been reached by the Breast Cancer Prevention Collaborative Group (BCPCG) to incorporate quantitative breast density into cancer risk prediction models. The methods presented here can be used to evaluate the role of the density measured by MRI (amount and morphological pattern) for risk prediction or as a surrogate marker predicting the effect of hormone related therapies.

With this grant support, we already have 7 papers and 19 conference proceedings published in a two-year period. We believe this research has very strong potential clinical value that will benefit many patients who need breast MRI examinations.




Symposium Abstract (2010)

Ke Nie (PI), Jeon-Hor Chen, Muqing Lin, Daniel Chang, Yu Hon, Orhan Nalcioglu, Min-Ying Lydia Su (mentor)

In 2007, the American Cancer society has issued a guideline recommending annual breast MRI screening to women with greater than 20-25% lifetime risk for breast cancer. There are well-established indications that will qualify a patient for receiving clinical MRI exams for screening, diagnosis, staging, or therapy evaluation, approved by the insurance company. Many more clinical breast MRI examinations are expected to be performed.  However, the low specificity may lead to great anxiety to patients, and many unnecessary biopsy or over-treatment. As the use of breast MRI increases, how to improve the diagnostic accuracy in breast MRI is becoming a more and more important problem. Furthermore, due to the mature clinical indication for breast MRI, there is a pressure for the small community hospitals or imaging centers to perform breast MRI. They may have well qualified mammographers, but not trained in reading breast MRI. On the other hand, the MRI radiologist may not be trained in breast imaging. The experience and training of radiologists in interpreting MRI raises a critical concern. Thus there is a critical need to develop computer-aided diagnosis system for breast MRI diagnosis.

The current available Computer-Aided Diagnosis (CAD) systems developed commercially can extract most important information from a large number of images obtained in a single MR examination to help radiologists’ interpretation. But, this approach offers only a display platform, not a true CAD which can provide “an intelligent thinking and final impression”. To respond to this great need, we are developing an automated CAD system to aid in detection and diagnosis of breast cancer on MRI.

Our CAD system was built based on 116 cases with both mass type and non-mass enhancement types lesions. The current system could automatically detect and segment suspicious region (possible lesion), then analyze their morphologic and kinetic features. A full panel of all morphology, texture, and enhancement kinetics information could be automatically obtained to allow differentiating between malignant and benign lesions. The selected specific features were trained using both non-linear artificial neural network and linear regression classifiers. A final score between 0-1 will be given to each detected lesion to indicate its malignancy level. Despite of the strong association between breast density and cancer risk, one major problem hampering incorporating of breast density into risk model is the lack of reliable quantitative measurement of density. Compared to mammography, MRI provides 3-dimensional coverage of the entire breast with strong contrast between fatty and fibroglandular tissues, thus has a high potential to provide quantitative density information. Thus, we extend the work from analyzing the lesion alone to quantifying the normal tissue. We further developed a dedicated tool to measure the normal tissue density and its relative distribution to fatty tissue, which would allow new investigations between density and cancer risk.

Overall, with the grant support, we already have 7 papers and over 20 conference proceedings published. We believe the research will finally benefit patients who are recommended to undergo MRI, especially those with dense breasts, particularly in young or Asian women.



Prediction of malignant breast lesions from MRI features: a comparison of artificial neural network and logistic regression techniques.
Periodical:Academic Radiology
Index Medicus: Acad Radiol
Authors: McLaren CE, Chen WP, Nie K, Su MY
Yr: 2009 Vol: 16 Nbr: 7 Abs: Pg:842-51

Development of a quantitative method for analysis of breast density based on three-dimensional breast MRI.
Periodical:Medical Physics
Index Medicus: Med Phys
Authors: Nie K, Chen JH, Chan S, Chau MK, Yu HJ, Bahri S, Tseng T, Nalcioglu O, Su MY
Yr: 2008 Vol: 35 Nbr: 12 Abs: Pg:5253-62

Quantitative analysis of lesion morphology and texture features for diagnostic prediction in breast MRI.
Periodical:Academic Radiology
Index Medicus: Acad Radiol
Authors: Nie K, Chen JH, Yu HJ, Chu Y, Nalcioglu O, Su MY
Yr: 2008 Vol: 15 Nbr: 12 Abs: Pg:1513-25

Selection of diagnostic features on breast MRI to differentiate between malignant and benign lesions using computer-aided diagnosis: differences in lesions presenting as mass and non-mass-like enhancement.
Periodical:European Radiology
Index Medicus: Eur Radiol
Authors: Newell D, Nie K, Chen JH, Hsu CC, Yu HJ, Nalcioglu O, Su MY
Yr: 2010 Vol: 20 Nbr: 4 Abs: Pg:771-81

Quantitative analysis of breast parenchymal patterns using 3D fibroglandular tissues segmented based on MRI.
Periodical:Medical Physics
Index Medicus: Med Phys
Authors: Nie K, Chang D, Chen JH, Hsu CC, Nalcioglu O, Su MY
Yr: 2010 Vol: 37 Nbr: 1 Abs: Pg:217-26

Impact of skin removal on quantitative measurement of breast density using MRI.
Periodical:Medical Physics
Index Medicus: Med Phys
Authors: Nie K, Chang D, Chen JH, Shih TC, Hsu CC, Nalcioglu O, Su MY
Yr: 2010 Vol: 37 Nbr: 1 Abs: Pg:227-33

Age- and race-dependence of the fibroglandular breast density analyzed on 3D MRI.
Periodical:Medical Physics
Index Medicus: Med Phys
Authors: Nie K, Su MY, Chau MK, Chan S, Nguyen H, Tseng T, Huang Y, McLaren CE, Nalcioglu O, Chen J
Yr: 2010 Vol: 37 Nbr: 6 Abs: Pg:2770-6