Computer assistance for MR based diagnosis of breast cancer: Present and future challenges
Introduction
Breast cancer is the second most prevalent cancer and the most common malignant disease in women in the industrialized countries. Besides common imaging techniques such as X-ray mammography and sonography used for screening and diagnosis of breast cancer, MR based methods such as dynamic contrast-enhanced MRI (DCE MRI or, more generally, breast MRI) or MR spectroscopy (MRS) have emerged as powerful diagnostic tools due to their high sensitivity. Other imaging techniques such as positron emission tomography or scintigraphy are used complementary in selected cases to further enhance specificity.
The use of breast MRI has a fundamental impact on the detection and diagnosis of breast cancer. Although X-ray mammography is the commonly used technique for screening and diagnosis in clinical practice, breast MRI has advantages due to its higher sensitivity and the absence of marring radiation.
Because of its ability to image dense breasts, MRI has a high potential for screening of high-risk cases. However, the reported specificities have displayed a very high range, varying from fairly low 20% to 100% [1], [2], [3]. Furthermore, microcalcifications are not visible on MR mammograms.
In addition, breast MRI, as is generally known, is associated with higher costs than X-ray mammography. Nevertheless, for women at high-familial risk of breast cancer, breast MRI might be a cost-effective screening modality [4].
The use of breast MRI as a screening modality is currently under investigation in several centers. In 1997 the MARIBS (magnetic resonance imaging in breast screening) study was initiated in the UK to evaluate breast MRI as a method of screening of premenopausal women with high risk of developing breast cancer [5]. Publications referring to the MARIBS study term that acceptable values of sensitivity and specificity were reached for MRI in these high-risk cases. A higher sensitivity of 77% was reported for MRI compared with 40% for X-ray [3]. Both methods combined resulted in a sensitivity of 94%. Specificity achieved only 81% for MR and 93% for X-ray mammography.
Breast MRI, however, is applied in selected cases. Added to standard imaging methods in cases of highly suspicious findings, breast MRI has shown to result in a change of treatment in many cases due to the assessment of further lesions in the contralateral breast or the assessment of multicentricity and multifocality [6], [7]. In combination with a better ability in the prognosis of tumor extent [7], [8], breast MRI is recommended for preoperative staging. Moreover, breast MRI should be applied in women with lymph node metastases and unknown primary cancer, as postoperative follow-up and for the examination of breast implants in patients suffering from symptoms, e.g. palpable findings or pain. Another important application of breast MRI is the monitoring of anti-cancer therapies for the assessment of drug effects on size of tumors and angiogenetic properties [9].
For the analysis of breast MRI data, both the importance of morphologic and the importance of kinetic parameter assessment was emphasized [10]. This has led to either a recommendation for a high spatial or a high temporal resolution of the MR sequences used. Recently, several studies have shown the importance of a combination for the discrimination of malignant from benign breast lesions [11], [12]. Moreover, the combination of morphologic and dynamic parameter assessment has proven to result in changes in recommended patient management and better treatment planning in many cases [6].
Combining different criteria, interpretation models (tree-shaped decision aids) have been evaluated in Refs. [12], [13] trying to differentiate malignant from benign disease. Furthermore, automated classification systems have been developed by several groups based on combinations of dynamic and morphologic features [14]. These systems as an addition to radiological reading were found to increase specificity and help in excluding benign lesions from further workup [15].
With the increasing amount of data in the acquired MR images, manual processing and interpretation becomes almost infeasible in a limited amount of time. There is an apparent need for computer assistance in preprocessing the data to extract and emphasize relevant information and present it to the radiologist. In addition, as the complexity of information increases, automated feature extraction for diagnostic purposes will become essential to improve the radiologists’ performance. Computer assistance for breast MRI has the potential to increase the specificity of the method. Deurloo et al. have stated that combining radiological reading with computer analysis has the potential to accurately exclude benign lesions from further workup [7]. Wiener et al. found that the implementation of computer-aided diagnosis for breast MRI will prove helpful [6]. An early system for computer assistance in DCE MRI in the breast was published by Meyer et al. [16].
In contrast to computerized analysis of X-ray mammograms, only a few investigations have pursued both the automated feature extraction and automated lesion classification. Most methods are based on the rating of features by radiologists followed by the merging of these ratings by an automated classifier [13]. Gilhuijs et al. developed a system for computerized delineation, rating and classification of breast lesions [14]. Chen et al. [17] evaluated a computerized interpretation of breast MRI using morphologic and kinetic features. Deurloo et al. [15] showed an improvement in lesion characterization rated indeterminate or suspicious by radiologists with a trained computer analysis system using a model that combined the results of the computerized analysis with those of clinical reading. Subramanian et al. [18] presented a system to identify, process, visualize and quantify breast lesions from MRI data.
The need for CAD is apparent when it comes to accurate diagnosis and regaining efficiency for radiologists. At present, there are two different meanings of “CAD” in its narrower sense: computer-aided detection and computer-aided diagnosis. The former denotes the detection of suspicious tissue and the latter denotes the differentiation of suspicious tissue into benign and malignant tissue. In computer-aided diagnosis, a computerized analysis serves as a second opinion for radiologists in the diagnosis of radiologic images. The goal is an estimate of the likelihood of malignancy based on computer-extracted features [19]. In computer-aided detection, a computer-based tool highlights suspicious areas to alert a human reader, thus trying to reduce the oversight error. In this article both aspects of CAD will be addressed.
Computer assistance in its wider sense additionally comprises automated or semi-automated procedures such as image preprocessing, image registration, image segmentation, color-overlay, signal processing for spectroscopic datasets and 3D rendering techniques.
In this paper the application of computerized methods for the evaluation of dynamic contrast-enhanced MRI of the breast will be discussed in more detail in Section 2 with an emphasis on registration methods for the reduction of motion artifacts, segmentation issues, as well as morphologic and dynamic lesion analyis. In the following sections the potentials of MR spectroscopy (Section 3) will be outlined and the application of MR guided biopsies (Section 4) is presented.
Section snippets
Dynamic contrast-enhanced MRI
In dynamic contrast-enhanced MRI, a certain dose of contrast agent (e.g. Gd-DTPA) is injected intravenously. One or several precontrast datasets are recorded before the injection of the contrast agent and several postcontrast datasets after injection. In order to extract both dynamic and morphologic information, a variety of different contrast agents can be used. The distribution of such contrast agents over time shows an increasing spread into tumorous tissue due to the increased permeability
MR spectroscopy
In vivo magnetic resonance spectroscopy (MRS) yields information about the chemical content of a distinct lesion within the breast. This information can be used for several clinical applications, such as monitoring response to cancer therapies or improving the accuracy of lesion diagnosis. Initial studies have already shown promising results (see Table 1 for details). In this section we will give a short overview of the current state of research in this field and discuss the future challenges
MRI guided interventions
Because of the high sensitivity and the moderate specificity of breast MRI, it is common to get histological clarification of an enhancing lesion. The standard procedures for this are either percutaneous biopsy or placement of hookwires for later surgical excision. Lesions that can be displayed using conventional imaging methods can undergo these procedures under control of ultrasound or X-ray, but for occult lesions which only show up on MRI, an MRI guided method is required.
It has been shown
Conclusion and future challenges
The most important issue for further research in the area of computer assistance in breast MRI is the implementation of robust and standardized post-processing techniques. The lack of standardized methods in the acquisition of dynamic contrast-enhanced MRI data and the employed analysis techniques complicate the comparison of results from different sites [67]. Different types of sequences, spatial and temporal resolution and the use of varying contrast agent doses strongly affect the relative
Acknowledgements
The authors would like to thank Dr. M. Rominger, University Hospital Marburg, Germany and Dr. J. Wiener and Dr. K. Schilling, Boca Raton Community Hospital, Florida for providing the breast MR data, and Dr. S. Kohle for indispensable help and contributions.
Sarah Behrens was born in Bremen, Germany, in 1978. She received her diploma in computer science from the University of Bremen, Germany, in 2003. Since 2003 she is working at MeVis Research, Bremen, as a scientist in medical image processing. Her research interest focuses on the analysis of MR images in oncology for diagnosis and therapy planning and breast MR imaging in particular.
References (91)
- et al.
Rationale for a national multi-centre study of magnetic resonance imaging screening in women at genetic risk of breast cancer
Breast
(2000) - et al.
Additional breast lesions in patients eligible for breast-conserving therapy by MRI: impact on preoperative management and potential benefit of computerised analysis
Eur J Cancer
(2005) - et al.
Interactive detection and visualization of breast lesions from dynamic contrast enhanced MRI volumes
Comp Med Imaging Graphics
(2004) - et al.
Improving breast cancer diagnosis with computer-aided diagnosis
Acad Radiol
(1999) - et al.
A non-rigid registration algorithm for dynamic breast MR images
Artif Intell
(1999) - et al.
A fuzzy c-means (FCM)-based approach for computerized segmentation of breast lesions in dynamic contrast-enhanced MR images
Acad Radiol
(2006) - et al.
Neural network-based segmentation of dynamic MR mammographic images
Magn Reson Imaging
(2002) - et al.
Analysis of dynamic MR breast images using a model of contrast enhancement
Med Image Anal
(1997) - et al.
Dynamic contrast-enhanced imaging and analysis at high spatial resolution of MCF7 human breast tumors
J Magn Reson
(1997) - et al.
In vivo 1H-MRS evaluation of malignant and benign breast diseases
Breast
(2003)
MR imaging of the breast for the detection, diagnosis, and staging of breast cancer
Radiology
International investigation of breast MRI: results of a multicentre study (11 sites) concerning diagnostic parameters for contrast-enhanced MRI based on 519 histopathologically correlated lesions
Eur Radiol
Reading protocol for dynamic contrast-enhanced MR images of the breast: sensitivity and specificity analysis
Radiology
Cost-effectiveness of screening with contrast enhanced magnetic resonance imaging vs. X-ray mammography of women at a high familial risk of breast cancer
Br J Cancer
Assessment of suspected breast cancer by MRI: a prospective clinical trial using a kinetic and morphologic analysis
Am J Roentgenol
Efficacy of MR mammography (MRM) in providing preoperative locoregional information on breast cancer: correlation between MRM and histological findings
Magn Reson Med Sci
Prediction of clinicopathologic response of breast cancer to primary chemotherapy at contrast-enhanced MR imaging: initial clinical results
Radiology
Dynamic breast MR imaging: are signal intensity time course data useful for differential diagnosis of enhancing lesions?
Radiology
Diagnostic architectural and dynamic features at breast MR imaging: multicenter study
Radiology
High-spatial-resolution MR imaging of focal breast masses: interpretation model based on kinetic and morphological parameters
Radiat Med
Update of breast MR imaging architectural interpretation model
Radiology
Breast MR imaging in women at increased lifetime risk of breast cancer: clinical system for computerized assessment of breast lesions initial results
Radiology
Clinically and mammographically occult breast lesions on MR images: potential effect of computerized assessment on clinical reading
Radiology
Comput Assisted Radiol Surg
Computerized interpretation of breast MRI: investigation of enhancement-variance dynamics
Med Phys
Dynamic contrast-enhanced magnetic resonance imaging in oncology
Magn Reson Imaging
Dynamic contrast-enhanced magnetic resonance imaging in oncology
Biomechanically based elastic breast registration using mass tensor simulation
Nonrigid registration using free-form deformations: application to breast MR images
IEEE Trans Med Imaging
Volume-preserving nonrigid registration of MR breast images using free-form deformation with an incompressibility constraint
IEEE Trans Med Imaging
Elastic matching of dynamic MR mammographic images
Magn Reson Med
Numerical methods for image registration
Towards real-time multi-modality 3-d medical image registration
Alignment by maximization of mutual information
Int J Comput Vision
Challenges for image processing in DCE MR mammography
Eur Radiol Suppl MRM
Characterization of lesions of the breast with proton MR spectroscopy: comparison of carcinomas, benign lesions, and phyllodes tumors
Am J Roentgenol
A novel projection based approach for medical image registration
GPU accelerated registration in two and three dimensions
Automatic segmentation of relevant structures in DCE MR mammograms
Automatic cropping of breast regions for registration in MR mammography
Dynamic contrast-enhanced MRI in the differentiation of breast tumors: user-defined versus semi-automated region-of-interest analysis
J Magn Reson
Automatic identification and classification of characteristic kinetic curves of breast lesions on DCE-MRI
Med Phys
Cited by (66)
Three-dimensional breast tumor segmentation on DCE-MRI with a multilabel attention-guided joint-phase-learning network
2021, Computerized Medical Imaging and GraphicsCitation Excerpt :A self-supervised learning model trained on large-scale image datasets without annotation from radiologists may further promote the learning performance. Second, for the analysis of breast tumors, future work will focus on the automatic classification of malignant and benign breast tumors using an end-to-end neural network to better aid clinical diagnosis (Behrens et al., 2007). Finally, for practical applications, the segmentation and classification from DCE-MRI could be realized in real practice to provide the comprehensive analysis of the breast tumor to assist radiologists.
Pre and post-hoc diagnosis and interpretation of malignancy from breast DCE-MRI
2019, Medical Image AnalysisCitation Excerpt :DCE-MRI is known to increase the sensitivity, compared to mammography, especially in young patients that have denser breasts (Kriege et al., 2004). However, the diagnosis and interpretation of DCE-MRI is a challenging and time consuming task that involves the interpretation of large amounts of data (Behrens et al., 2007) and is prone to high inter-observer variability (Grimm et al., 2015; Lehman et al., 2013). Computer-aided diagnosis (CAD) systems are designed to reduce the analysis time (Gubern-Mérida et al., 2016; Wood, 2005), increase sensitivity (Vreemann et al., 2018) and specificity (Meinel et al., 2007), and serve as a second (automated) reader (Shimauchi et al., 2011).
Patch-based system for Classification of Breast Histology images using deep learning
2019, Computerized Medical Imaging and GraphicsSpatiotemporal features of DCE-MRI for breast cancer diagnosis
2018, Computer Methods and Programs in BiomedicineSimplification of breast deformation modelling to support breast cancer treatment planning
2016, Biocybernetics and Biomedical Engineering
Sarah Behrens was born in Bremen, Germany, in 1978. She received her diploma in computer science from the University of Bremen, Germany, in 2003. Since 2003 she is working at MeVis Research, Bremen, as a scientist in medical image processing. Her research interest focuses on the analysis of MR images in oncology for diagnosis and therapy planning and breast MR imaging in particular.
Hendrik Laue was born in Bremen, Germany, in 1971. He received the diploma in physics from the University of Bremen, Germany, in 1999 and a PhD in 2004. Since 2005, he works at MeVis Research, Bremen, as physicist and software engineer. His main research topics are contrast agent dynamics for magnetic resonance imaging and its applications in tumor diagnostics.
Matthias Althaus was born in Stadtlohn, Germany, in 1974. He received his diploma in experimental physics from the University of Dortmund, Germany, in 2001 and a PhD in 2004 from the University of Bremen, Germany. From 2005–2007 he has been with MeVis Research, Bremen. His research interests include MR imaging in general and MR spectroscopy in particular.
Tobias Boehler was born in Oberhausen, Germany, in 1977. He received his diploma in computer science from the University of Dortmund in 2005 and works at MeVis Research since then. His main research interests are registration methods for motion correction of dynamic images and intersequence registration with a special focus on MR mammography.
Bernd Kuemmerlen was born in Heilbronn, Germany, 1970. He received a diploma in physics from the University of Heidelberg in 1998. Since 1998, he has been working with MeVis Research, Bremen, Germany, on various topics, including the analysis of dynamic MR images, vessel segmentation and MR guided intervention planning. Since 2006 he is working for MeVis Technology AG, Bremen, Germany, as R&D and Product Manager.
Horst Karl Hahn was born in 1972 in Freiburg, Germany. He received his diploma in physics in Heidelberg, Germany in 1998 and a PhD at the University of Bremen in 2005. His current activities are in the field of quantitative medical image analysis and multimodal visualization and are driven by clinical applications for diagnosis, therapy planning and monitoring. In 1999, he set up the Neuroimaging group at MeVis Research, which focuses on neurosurgical risk analysis and on neurological diagnosis and therapy monitoring. Hahn is currently director of research and vice president at MeVis Research, Bremen, Germany.
Heinz-Otto Peitgen (born 1945 in Bruch near Cologne) is a German mathematician. Peitgen studied mathematics, physics and economy from 1965 until 1971 in Bonn, later working for six years at the Institute for Applied Mathematics at the University of Bonn, where he received his PhD in 1973. After receiving his habilitation in 1977, he first taught as private docent in Bonn before obtaining a professorship for mathematics at the University of Bremen. Peitgen is director of the Centre for Complex Systems and Visualization at the University of Bremen. His research work emphasizes on dynamical systems, numerical analysis, image analysis, and data analysis, as well as the use of computers in image-based medical diagnostics. Dr. Peitgen is currently a professor of Mathematics and Biomedical Sciences at Florida Atlantic University and the University of Bremen in Germany. Dr. Peitgen is also the founder and president of MeVis Research – Center for Medical Image Computing, GmbH in Bremen, Germany.