3D Multi-parametric Breast MRI Segmentation Using Hierarchical Support Vector Machine with Coil Sensitivity Correction

Rationale and Objectives

The goal of the study is to develop a technique to achieve accurate volumetric breast tissue segmentation using magnetic resonance imaging (MRI) data. This segmentation can be useful to aid in the diagnosis of breast cancers and to assess breast cancer risk based on breast density. Tissue segmentation is also essential for development of acoustic and thermal models used in magnetic resonance guided high-intensity focused ultrasound treatment of breast lesions.

Materials and Methods

In addition to commonly used T1-, T2-, and proton density–weighted images, three-point Dixon water- and fat-only images were also included as part of the multiparametric inputs to a tissue segmentation algorithm using a hierarchical support vector machine (SVM). The effectiveness of a variety of preprocessing schemes was evaluated through two in vivo datasets. The performance of the hierarchical SVM was investigated and compared to the conventional classification algorithms—conventional SVM and fuzzy C-mean (FCM).

Results

The need for co-registration, zero-filled interpolation, coil sensitivity correction, and optimal SNR reconstruction before the final stage classification was demonstrated. The overlap ratios of the hierarchical SVM, conventional SVM and FCM were 93.25%–94.08%, 81.68–92.28%, and 75.96%–91.02%, respectively. Classification outputs from in vivo experiments showed that the presented methodology is consistent and outperforms other algorithms.

Conclusion

The presented hierarchical SVM-based technique showed promising results in automatically segmenting breast tissues into fat, fibroglandular tissue, skin, and lesions. The results provide evidence that both the multiparametric breast MRI inputs and the preprocessing procedures contribute to the high accuracy of tissue classification.

Introduction

Breast magnetic resonance imaging (MRI) has become a very useful imaging modality for breast cancer screening and diagnosis . It has been shown that 17%–34 % of cancer foci visible on breast MRI are not detected by mammography . Not only does breast MRI offer higher sensitivity for detection of breast cancer than x-ray mammography, ultrasound, clinical examination, or any combination of these, it also has a superior ability to delineate fatty and fibroglandular tissue .

Although lesions can be detected by visual inspection of breast MRI images, including dynamic contrast-enhanced (DCE) studies, there is evidence that quantitative measurements of different structures in the breast with and without contrast can assist not only in the detection of abnormal tissues, but also in the discrimination between fibroadenomas, cysts, and various types of malignancies . In an attempt to improve the performance of breast computer-aided diagnosis systems that are designed to supplement visual inspection and interpretation of breast MRI, methods for fully and semiautomatic segmentation of lesion mass based on DCE-MRIs have been developed . Efforts have also been made to discriminate between benign and malignant lesions using quantitative morphological and feature analyses . In addition to automatic lesion detection and discrimination, breast tissue segmentation could also be used to determine the percentage of fibroglandular tissue present in the breast, which is directly linked to breast parenchymal patterns , where the parenchymal pattern characterization parameters are taken as risk factors of developing breast cancer .

Get Radiology Tree app to read full this article<

Materials and methods

Subjects and Image Acquisition

Get Radiology Tree app to read full this article<

Preprocessing

Get Radiology Tree app to read full this article<

Co-registration

Get Radiology Tree app to read full this article<

ZFI

Get Radiology Tree app to read full this article<

Three-point Dixon reconstruction

Get Radiology Tree app to read full this article<

Skin extraction

Get Radiology Tree app to read full this article<

Coil sensitivity estimation

Get Radiology Tree app to read full this article<

Optimal SNR reconstruction

Get Radiology Tree app to read full this article<

Iopt(r)=R(r)Ψ−1ST(r)S(r)Ψ−1ST(r) I

(

)

(

)

−

(

)

(

)

−

(

)

where r denotes the position in the image space; R(r)=[R1(r),R2(r),…,RL(r)] R

(

)

[

(

)

(

)

…

(

)

] is the row vector of coil images; S(r)=[S1(r),S2(r),…,SL(r)] S

(

)

[

(

)

(

)

…

(

)

] is the row vector of coil sensitivities estimated from above step; and Ψ Ψ is an L by L matrix that describes the coupling and noise correlations between the coil elements. The noise matrix was assumed to be an identity matrix in the actual calculation for simplicity, and it was shown by Roemer et al that there was only a 10% SNR loss when assuming there is no noise correlation.

Get Radiology Tree app to read full this article<

Tissue Segmentation Using Hierarchical SVM

Get Radiology Tree app to read full this article<

Statistical Analysis

Get Radiology Tree app to read full this article<

Results

Get Radiology Tree app to read full this article<

Table 1

Overlap Ratios of Various Algorithms with Radiologist’s Manual Classification as Ground Truth

Overlap Ratio h-SVM (Sagittal), % h-SVM (Axial), % c-SVM, % FCM, % h-SVM without CSC, % h-SVM without ZFI, % h-SVM without Co-registration, % Dataset 1 90.94 90.47 86.97 75.96 82.09 90.69 88.81 Dataset 2 94.08 93.92 90.30 80.58 83.42 94.50 93.67 Dataset 3 93.90 93.90 81.68 91.02 95.45 94.41 95.38 Dataset 4 93.25 92.79 92.28 80.06 81.39 92.76 81.12

Hierarchical support vector machine (h-SVM) offers highest overlap ratio than the conventional SVM (c-SVM) and the FCM algorithm. Three preprocessing steps: coil sensitivity correction (CSC), zero-filled interpolation (ZFI), and co-registration are also evaluated.

Get Radiology Tree app to read full this article<

Table 2a

Individual Tissue Type Analysis of Dataset 1 Segmentation: Correctly Classified, Incorrectly Classified, and Overlap Ratios of Hierarchical SVM for Individual Tissue Type, Comparing to the Manual Classification from Radiologist 1

Hierarchical SVM Fat Fibroglandular Skin Lesion Air Overlap Ratio, % Radiologist Fat3877 165 45 4 116 92.16 Fibroglandular 222467 17 2 1 65.87 Skin 8 37180 2 91 56.60 Lesion 0 8 08 0 50.00 Air 6 30 14 03401 98.55