Quantitative breast mass classification based on the integration of B-mode features and strain features in elastography
Introduction
As an adjunct to mammography, ultrasound (US) is widely used in distinguishing between benign and malignant breast lesions [1]. The Breast Imaging Reporting and Data System (BI-RADS) lexicon for US was defined by American College of Radiology [2] to standardize the descriptive terminology. The sonographic descriptors defined in BI-RADS lexicon such as shape, orientation, margin, lesion boundary, echo pattern, and posterior acoustic features were evaluated to be important in classify breast lesions in the prior studies [3], [4], [5]. With the standardized BI-RADS descriptors, the radiologists improved their agreement and performance in US interpretation across all experience variables in the literature [5].
After quantification, the sonographic descriptors were used in various computer-aided diagnosis (CAD) systems to estimate the likelihood being a carcinoma [6], [7], [8]. The quantitative characteristics extracted from B-mode images such as shape features were used in describing the physical properties of breast tumors [9], [10]. The mechanical property of tumors is also investigated for histological and pathological relevance. For example, the pathological changes of the scirrhous carcinoma correlate with tissue stiffness [11]. As a newly developed sonographic imaging modality [12], [13], [14], elastography estimates the tissue stiffness by calculating the tissue displacement under a certain force. In quasi-static elastography, the tissue displacement is induced by manual pressing on the transducer such as the respiratory movement of patients utilized in this study. As another form of elastography, shear wave elastography [15], automatically radiates shear waves inside tissues and reconstructs the elasticity properties of tissues from the propagated shear wave. However, the simplicity of quasi-static elastography makes it can be implemented using conventional US hardware.
The gradient of the displacement or strain is converted to the pixel value for imaging. Clinically, the size ratio between the strain area in elastographic images and tumor area in B-mode images was evaluated to interpret breast tumor malignancy [16], [17], [18], [19], [20]. The definition is the maximum horizontal length of the tumor measured in the elastographic image divided by the corresponding length measured in the B-mode. A tumor with an elastographic images/B-mode size ratio greater than 1.0 had high likelihood being malignant. The stiffness contrast and normalized shear strain area were also used to diagnose breast tumors [14], [19], [21], [22], [23], [24]. The stiffness contrast is the ratio between the mean strains estimated in the background to that estimated within the breast tumor. The area obtained from the strain area can be normalized via the tumor size measured from the B-mode image. The performances of Az, area under the receiver operating characteristic (ROC) curve achieved 0.84–0.99 in the results of the literature.
However, by observing the elastographic images, the shape and boundary of tumor area are not as clear as those in B-mode images. Radiologists have to manually compare elastographic images and B-mode images to estimate the stiffness of tumor area which is operator-dependent. Burnside et al. indicated that inter-observer variability was an influence factor to the diagnostic performance [17]. Prior CAD approach which extracted quantitative elasticity features from acoustic radiation force impulse imaging achieved an accuracy of 80% in classifying breast tumors [25]. In this study, we proposed an automatic procedure to extract strain features from quasi-static elastography in a CAD system for tumor classification. Tumor segmentation was first applied to the B-mode images and the resulting tumor contour was mapped to the elastographic image. The fuzzy c-means clustering (FCM) [26] was then used to automatically extract stiff tissues with darker values for feature extraction. Compared to other CAD systems [6], [7], [8], [9], [10], the current study explored the performance difference between the quantitative strain features and B-mode features. The complementary advantage of combining both feature sets was also evaluated in the experiment. The flowchart of the diagnostic procedure is shown in Fig. 1.
Section snippets
Patients and data acquisition
This study obtained the approval from our institution review board and the patients׳ informed consent. A total of 90 patients underwent ultrasound examination with an ACUSON S2000 ultrasound system (Siemens Medical Solutions, Mountain View, CA) from November 2011 to June 2012. The equipped linear-array transducer was a 9L4 for the B-mode sonographic examination followed by elastography imaging. The target lesion was displayed in the B-mode sonogram with the corresponding strain image displayed
Results
Table 1 shows the test result of the proposed six strain features. Only one of them was not significant. In Table 2, the B-mode feature set included 15 significant features such as the best-fit ellipse properties and NRL features. The performances of the feature sets using the binary logistic regression model are shown in Table 3. The proposed strain feature set achieved an accuracy of 80% (72/90), a sensitivity of 80% (36/45), a specificity of 80% (36/45), and an Az of 0.84. In the comparisons
Discussion
This study proposed a CAD system to extract quantitative strain features from elastographic images for breast mass classification. With the mapping of tumor contour from the B-mode image, the tumor area in the elastographic image can be specified automatically to reduce user dependence. After FCM clustering, the stiff tissues surrounding the tumor area were extracted to estimate the size of strain area. The one-dimension strain length and two-dimension strain area around a tumor were calculated
Conclusion
This study proposed an automatic procedure to extract strain features from elastographic images with less operator dependence. The quantified strain features can be integrated with the B-mode features in the CAD system to provide a promising diagnosis of breast tumors.
Summary
Conventionally, radiologists have to observe the gray-scale distribution of tissues on the elastographic image to acquire the strain information of tissues. This study proposed a computer-aided diagnosis (CAD) system to extract quantitative strain features from elastographic images to reduce operator-dependence and provide an automatic procedure for breast mass classification. For each breast mass, tumor segmentation was performed on the B-mode image to obtain tumor contour which was then
Conflict of interest statement
None declared.
Acknowledgments
The authors thank the National Science Council (NSC 101-2221-E-002-068-MY3), Ministry of Economic Affairs (100-EC-17-A-19-S1-164), Department of Health (DOH102-TD-C-111-001), and Ministry of Education (AE-00-00-06) of the Republic of China for the financial support.
References (48)
- et al.
Computer-aided diagnosis of breast masses using quantified BI-RADS findings
Comput. Methods Progr. Biomed.
(2013) - et al.
Computer-aided diagnosis based on speckle patterns in ultrasound images
Ultrasound Med. Biol.
(2012) - et al.
Elastography: elasticity imaging using ultrasound with application to muscle and breast in vivo
Ultrason. Imaging
(1993) - et al.
A prospective study to compare the diagnostic performance of breast elastography versus conventional breast ultrasound
Clin. Radiol.
(2010) - et al.
Axial shear strain imaging for differentiating benign and malignant breast masses
Ultrasound Med. Biol.
(2010) - et al.
Axial-shear strain elastography for breast lesion classification: further results from in vivo data
Ultrasound Med. Biol.
(2011) - et al.
Effect of lesion boundary conditions on axial strain elastograms: a parametric study
Ultrasound Med. Biol.
(2007) - et al.
Breast tumor classification using fuzzy clustering for breast elastography
Ultrasound Med. Biol.
(2011) - et al.
Fcm—the fuzzy c-means clustering-algorithm
Comput. Geosci.
(1984) A new segmentation system for brain MR images based on fuzzy techniques
Appl. Soft Comput.
(2008)
Quantitative analysis of lesion morphology and texture features for diagnostic prediction in breast MRI
Acad. Radiol.
Semi-quantitative and qualitative assessment of breast ultrasound elastography in differentiating between malignant and benign lesions
Ultrasound Med. Biol.
Solid breast nodules: use of sonography to distinguish between benign and malignant lesions
Radiology
Breast Imaging Reporting and Data System: ACR BI-RADS—Breast Imaging Atlas
Characterization of solid breast masses use of the sonographic breast imaging reporting and data system lexicon
J. Ultrasound Med.
Probably benign lesions at screening breast us in a population with elevated risk: prevalence and rate of malignancy in the ACRIN 6666 trial
Radiology
Training the ACRIN 6666 investigators and effects of feedback on breast ultrasound interpretive performance and agreement in BI-RADS ultrasound feature analysis
Am. J. Roentgenol.
Computer-aided diagnosis of solid breast nodules: use of an artificial neural network based on multiple sonographic features
IEEE Trans. Med. Imaging
Robustness of computerized lesion detection and classification scheme across different breast US platforms
Radiology
Computer-aided classification of breast masses using speckle features of automated breast ultrasound images
Med. Phys.
Pathology
Evaluation of breast lesions using sonographic elasticity imaging a multicenter trial
J. Ultrasound Med.
Breast lesions: Evaluation with US strain imaging - Clinical experience of multiple observers
Radiology
Supersonic shear imaging: a new technique for soft tissue elasticity mapping
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
Cited by (18)
Automated whole breast segmentation for hand-held ultrasound with position information: Application to breast density estimation
2020, Computer Methods and Programs in BiomedicineCitation Excerpt :The whole breast region was acquired and then the fibroglandular tissue was extracted by FCM classifier. In this study, FCM classifier [40], a well-known extension of the k-means classifier [41] which has been widely used to distinguish fibroglandular tissue from whole breast to compute the breast density [22,42–45], was adopted. It was compared to a FDA approved, commercially available, and fully-automated software for estimating volumetric breast density Quantra and Volpara, and were found to be well-performed [1,46].
An intelligent healthcare system for optimized breast cancer diagnosis using harmony search and simulated annealing (HS-SA) algorithm
2020, Informatics in Medicine UnlockedCitation Excerpt :The analysis of biomedical data (biomedical images like MRI, CT scan, PET, US, mammography, and biomedical signals like EEG, ECG, EMG) has led to the design of automated and smart CAD systems that help to diagnose early, accurate and precise diseases even before symptoms are visible externally. With the development of artificial intelligence techniques, the race of data-driven intelligent classification approaches have been applied for breast cancer diagnosis, such as Naive Bayesian [12–14], Neural Network [15–17], Support Vector Machine (SVM) [18–20], ensemble methods [21–23], K-means [24–26], fuzzy and rough set techniques [27,28], PSO [29–31], semi-supervised techniques [32,33], deep learning [34–37], transfer learning [37–40], Active learning [41–43] or other hybrid algorithms [44–46]. In 2011, B. Zhang et al. [47] proposed a random Subspace cascade with rejection options for the classification of microscopic biopsy images.
Machine learning techniques for breast cancer computer aided diagnosis using different image modalities: A systematic review
2018, Computer Methods and Programs in BiomedicineQuantitative breast density analysis using tomosynthesis and comparison with MRI and digital mammography
2018, Computer Methods and Programs in BiomedicineCitation Excerpt :After that, in order to improve the measurement accuracy of breast volume, further refinement of the breast area was performed on both tomosynthesis and MRI image. With the completion of Breast Area Refinement, tissue classification was carried out by FCM [12,13,27,28] for reconstruct 3-D structure and percent density can be calculated. For detailed steps and methods are as follows:
On the soft tissue ultrasound elastography using FEM based inversion approach
2024, Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in MedicineA Novel Deep-Learning-Based CADx Architecture for Classification of Thyroid Nodules Using Ultrasound Images
2023, Interdisciplinary Sciences – Computational Life Sciences