Original contribution
Estimation of the content of fat and parenchyma in breast tissue using MRI T1 histograms and phantoms

https://doi.org/10.1016/j.mri.2005.02.006Get rights and content

Abstract

Mammographic breast density has been correlated with breast cancer risk. Estimation of the volumetric composition of breast tissue using three-dimensional MRI has been proposed, but accuracy depends upon the estimation methods employed. The use of segmentation based on T1 relaxation rates allows quantitative estimates of fat and parenchyma volume, but is limited by partial volume effects. An investigation employing phantom breast tissue composed of various combinations of chicken breast (to represent parenchyma) and cooking fats was carried out to elucidate the factors that influence MRI T1 histograms. Using the phantoms, T1 histograms and their known fat and parenchyma composition, a logistic distribution function was derived to describe the apportioning of the T1 histogram to fat and parenchyma. This function and T1 histograms were then used to predict the fat and parenchyma content of breasts from 14 women. Using this method, the composition of the breast tissue in the study population was as follows: fat 69.9±22.9% and parenchyma 30.1±22.9%.

Introduction

A mammographically dense breast refers to a breast which appears to more intensely attenuate the X-ray beam and is regarded as depicting mostly parenchyma, while a breast which is not dense appears radiolucent and is regarded as being composed of mostly fatty tissue [1]. In 1976, Wolfe [2] found a relationship between mammographic breast density and breast cancer risk. Since then, there have been a substantial number of reports that have shown that the odds ratio for developing breast cancer for the most dense compared with the least dense breast tissue categories ranges from 1.8 to 6.0, with most studies yielding an odds ratio of 4.0 or greater [3]. Despite the undoubted importance of breast density in the etiology and the detection of breast cancer, it has been described as “perhaps the most undervalued and underutilized risk factor in studies investigating the causes of breast cancer” [4]. Technical reasons are probably partly responsible for the limited utilization of breast density in cancer research.

Mammographic density is commonly estimated by delineating the radiographically dense areas on the mammogram from the entire breast area, and the percentage breast density is calculated as the area of high radiographic density divided by the total breast area [5], [6]. Although digital segmentation with either interactive [7] or automated [8] thresholding has been employed to differentiate between fat and parenchyma, the entire notion of a threshold means that the differentiation is somewhat subjective. Another limitation of the current mammography methodology is that pixels are identified as either fat or parenchyma, and no account is taken of the actual depth of the pixel being imaged. Variation in the thickness of breasts during compression, variations in positioning of the breast, nonlinear characteristics of film and film digitizers, and variations in X-ray flux have also been suggested as factors that could influence the apparent density of a mammogram [9]. Given these confounding factors, it is not surprising that recent research has demonstrated the high degree of inaccuracy of mammography for estimating the fat and parenchyma content of breasts [10]. Despite these problems, there is now evidence that there is large variation in the parenchyma/fat tissue in women [11] and that this ratio declines with age [1], [12] and can be expected to decline as the compressed thickness of the breast increases [13]. Breast composition also undergoes cyclic changes during the menstrual cycle [14] and may be influenced by diet (intakes of polyunsaturated fat, vitamins C, E and B12, and by intake of alcohol) [15] and by hormone replacement therapy [16].

Despite the importance of breast composition to the etiology and diagnosis of breast cancer, there have been only a few studies that have made direct chemical measurements of parenchyma and fat in breast tissue, and these involved only a small number of fresh mastectomy specimens [17], [18].

An alternative modality for estimating the fat and parenchyma content of tissues in animals and humans involves MRI [19], [20], [21]. Moreover, a significant correlation has already been shown between X-ray mammography percent density and two MRI parameters: mean T2 relaxation time and relative water content [22]. A number of different methods have been used to estimate fat and parenchyma content of tissues from MRI data. Kover et al. [19], when estimating the fat content of live chickens, used a mouse to manually draw around areas of adipose tissue and thus segment the MR image into fat and parenchyma. Baulain [20] used both gray value distribution (histogram) of the pixels and also cluster analysis methods to estimate body composition of live pigs and sheep. In human breast tissue, fat is often diffusely dispersed throughout parenchymal tissue, and it may not necessarily be practical or possible to manually outline in a MR image distinct regions of adipose tissue. Therefore, a variety of methods have been used to segment MR images and to classify the tissue in the different segments [23]. There are two distinct approaches that rely on unique NMR properties of fat relative to the fibroglandular tissues: chemical shift and T1. Although chemical shift techniques may be more specific in identifying fat, T1-based techniques are easier to implement over a wide variety of scanner platforms and field strengths. This is a technique currently being employed in a study at the University of Pennsylvania.

T1-based techniques take advantage of the fact that inversion recovery images with different T1 times, or T1 weighted images with different TR values, will have a gray scale such that the signal in each voxel is related to the T1 relaxation time of the voxel (or cube of tissue) that the pixel depicts. A T1 image or map can be calculated from a series of images produced using an inversion recovery pulse sequence. The gray scale of a specific pixel is related to the T1 relaxation time of the voxel (or cube of tissue) which that pixel depicts. T1 histograms depict the number of voxels that have specific T1 relaxation times. Fig. 1A–D depicts a variety of commonly observed T1 histograms of women's breasts, calculated from an inversion recovery experiment with T1=1600, 800, 400, 200 and 100 ms. Fig. 1A shows a T1 histogram obtained from the left breast of a woman. The Gaussian-type curve centered on a T1 time of approximately 280 ms corresponds to the signal obtained from adipose dominant tissue. This histogram is typical of that observed in many older women with large, fatty or “glandular” breasts. The T1 histogram shown in Fig. 1B comes from a breast containing small amounts of both adipose and parenchymal tissue. The Gaussian-type curve center centered on approximately 940 ms corresponds to the signal obtained from parenchyma dominant tissue. The bridging area between the two peaks is thought to represent tissue that contains both adipose and parenchymal tissue. This type of histogram is typically obtained from young women athletes who have small, “muscular” breasts. Fig. 1C shows a T1 histogram with a substantial “fat” peak, but in the region of the histogram normally associated with parenchyma, the T1 curve is so flat that it is not possible to accurately determine the time of the parenchyma peak. This type of histogram is typical of histograms obtained from many young women. In Fig. 1D, there is only one substantially skewed peak and in the region of the T1 histogram normally associated with parenchyma, the histogram declines monotonically. This histogram is very unusual and, by visual inspection of the MR image, resulted from a breast containing a substantial quantity of parenchyma finely dispersed throughout the adipose tissue.

Some researchers have analyzed T1 histograms by fitting data to a sum of exponentials [24]. However, in our experience, the diversity of T1 histograms as demonstrated in Fig. 1A–D makes this approach problematic. Segmentation of the type of histogram shown in Fig. 1B–D to provide an estimate of the fat and parenchyma content of a breast is bedeviled by the problem of partial voluming. This phenomenon arises since volumes of breast tissue, even those as small as a voxel, can contain mixtures of fat and parenchyma. In the past, many researchers have examined each voxel in each slice of a histogram and, on the basis of a threshold T1 value, have designated each voxel as being either fat or parenchyma and then aggregated these specific voxel types to obtain an overall estimate of the total content of fat and parenchyma [19], [25]. A primary difficulty with this approach is determining the value of the threshold [23]. A second problem with this approach is that it denies the fact that many voxels are composed of combinations of tissue types (partial volume effects). In an example of the types of problems that arise from the use of threshold algorithms, Laidlaw et al. [26] presented a brain MR image in which both stair step artifacts and layers of misclassified voxels are clearly evident. In 1996, Graham et al. [22], in an attempt to improve the estimation of breast composition, correlated the mean T2 value of a MRI histogram (calculated as the first moment of the estimated continuous T2 distribution) with the relative volumetric water content of breasts (also estimated by MRI). This approach may allow a global estimation of fat and parenchymal content but may not be appropriate for small regions of interest or individual voxels.

Another approach to overcome the difficulties caused by the partial volume phenomenon is the use of Bayesian probability theory to estimate the highest probability combination of materials within each voxel-sized region [26]. The possible tissue types within each voxel are identified, and continuous “basis” functions p(x) are assumed/developed to represent the probability that a particular voxel contains a particular tissue type. Linear combinations of these functions are fitted to each voxel, and the most likely combination of materials is chosen probabilistically.

This contribution describes the use of breast phantoms to elucidate the effect of various combinations of fat and parenchyma on the pattern of T1 histograms, the nature of the p(x) basis functions and the use of basis functions to predict the percentage of fat and parenchyma in breasts from the whole-breast T1 histogram.

Section snippets

Phantoms

The phantoms were made of a plastic-capped, 50-ml polyethylene (1-mm-thick) sampling tube containing various quantities and proportions of ground lean chicken breast (representing parenchyma) and cooking fat (Crisco) or canola oil representing adipose tissue (see Table 1). The volume of chicken breast added to each phantom was determined by water displacement, while the volumes for liquefied cooking fat or canola oil were measured. In order to simulate the heterogeneity of fat and parenchyma

Phantom histograms

T1 histograms of cooking fat, canola oil, chicken breast and mixtures of chicken breast and fat are shown in Fig. 4A–D. The histogram of cooking fat was Gaussian in form and closely resembled the fat peak that occurs in typical breast T1 histograms (compare with Fig. 1A). The histogram from the canola oil had a much smaller peak and was more platykurtotic than the histogram obtained from the cooking fat. The histograms produced from phantoms containing mixtures of fat and chicken had reduced

Conclusion

The empirical logistic model presented here has the flexibility to accurately and plausibly segment all of the MRI T1 histograms that we have encountered.

References (35)

  • C. Byrne

    Studying mammographic density: implications for understanding breast cancer

    J Natl Cancer Inst

    (1997)
  • J.A. Shepherd et al.

    Measurement of breast density with dual X-ray absorptiometry: feasibility

    Radiology

    (2002)
  • G. Ursin et al.

    Mammographic density changes during the menstrual cycle

    Cancer Epidemiol Biomark Prev

    (2001)
  • J.W. Byng et al.

    Analysis of mammographic density and breast cancer risk from digitized mammograms

    Radiographics

    (1998)
  • X.H. Wang et al.

    Automated assessment of the composition of breast tissue revealed on tissue-thickness-corrected mammography

    AJR Am J Roentgenol

    (2003)
  • N.A. Lee et al.

    Fatty and fibroglandular tissue volumes in the breasts of women 20–83 years old: comparison of X-ray mammography and computer-assisted MR imaging

    AJR Am J Roentgenol

    (1997)
  • M.E. Costanza

    Breast cancer screening in older women

    Cancer Suppl

    (1992)
  • Cited by (37)

    • Analysing breast tissue composition with MRI using currently available short, simple sequences

      2016, Clinical Radiology
      Citation Excerpt :

      Direct comparison of these three studies is, however, challenging, because of the differing field strengths used. Many studies have compared different clustering methods to quantify breast tissue composition on MRI images.8,9,11,12,15,16 Because the focus of this study was to determine the most robust MRI sequence to use, a basic clustering method was employed to identify the trough in the signal intensity histogram.

    • Volatile Biomarkers in Breath Associated With Liver Cirrhosis - Comparisons of Pre- and Post-liver Transplant Breath Samples

      2015, EBioMedicine
      Citation Excerpt :

      A study involving women with early-stage breast cancer taking a high oral dose of limonene (2 g/daily for 2–6 weeks before surgery) found that mean limonene concentration in breast tissue was 41.3 ± 49.9 μg/g which is much higher than that found in a control group (0.08 ± 0.13 μg/g) (Miller et al., 2013). Breast tissue is primarily composed of fat (Boston et al., 2005). This supports our hypothesis that unmetabolised limonene accumulates in fat tissue.

    • Women In Steady Exercise Research (WISER) Sister: Study design and methods

      2015, Contemporary Clinical Trials
      Citation Excerpt :

      Bilateral contrast MRI examinations were performed between days 6 and 10 of the menstrual cycle at months one and seven. Images were acquired on a 1.5-Tesla Siemens scanner according to published methods [36]. Specifically, three-dimensional bilateral slab interleaved spoiled gradient echo inversion recovery sequence was used to acquire data (TR/TE 17.6/1.4 ms, flip angle = 55°, ± 64 kHz sampling bandwidth, 64 slices).

    • Breast anatomy and physiology

      2009, Hughes, Mansel & Webster's Benign Disorders and Diseases of the Breast: Third Edition
    • High Mammographic Breast Density Is Independent Predictor of Local But Not Distant Recurrence After Lumpectomy and Radiotherapy for Invasive Breast Cancer

      2009, International Journal of Radiation Oncology Biology Physics
      Citation Excerpt :

      Normal breast tissue is composed of an epithelial cell compartment separated from fibroadipose stroma and ECM regulated by a complex hormonal milieu. Breast tissue that represents mammographic density is thought to be nonadipose fibroglandular tissue whose nature and specific composition is an area of active investigation (3–5). Mammographic breast density (MBD) is one of the strongest known risk factors for the development of breast cancer (6), and the percentage of MBD correlates strongly with risk.

    View all citing articles on Scopus

    This work was supported by grants from the NIH (CA82707-01, CA090699-0182) and by a grant from the Susan G. Komen Breast Cancer Foundation (IMG 2000-224).

    View full text