Early-life body size has been consistently associated with breast cancer risk. The direction of the association changes over time, with high birth weight, smaller adolescent body size, and adult weight gain all increasing breast cancer risk. There is also a clear positive association between larger body size and increased breast adipose tissue measured by mammograms, but less is known about how body size changes across life stages affect stromal and epithelial breast tissue. Using breast tissue slides from women with benign breast disease, Oh and colleagues applied machine learning methods to evaluate body size across the life course and adipose, epithelial, and stromal tissue concentrations in adulthood. They found consistent patterns for higher adipose and lower stromal tissue concentrations with larger childhood and adult body size at age 18 years. They reported lower levels of epithelial tissue with larger body size at 18 years, but not at other time periods. Additional studies examining how body size at different life stages may affect breast tissue composition will be important. Noninvasive methods that can provide measures of breast tissue composition may offer potential ways forward to ensure generalizability, and repeated measurements by life stage.

See related article by Oh et al., p. 608

There is consistent epidemiologic evidence linking body size over the life course with breast cancer risk. The direction of the association changes over time, however, with high birth weight, smaller adolescent and early adult (e.g., at age 18) body size, and adult weight gain all increasing breast cancer risk (1). A hypothesis is that the relationship between anthropometry and breast cancer risk may be mediated through breast tissue density. Although previous studies support an association between body size across the life course and mammographic breast density (MBD), the relative amount of breast area, which is dense (fibroglandular) and nondense (adipose) tissue as seen on a mammogram (2–5), does not disentangle associations with breast tissue types beyond what is captured on a 2D mammogram image. The study conducted by Oh and colleagues is the first to examine early-life and adult body size in relation to the specific components of breast density, including the two types of fibroglandular tissue (epithelial and fibrous stromal) and adipose tissue, in a subcohort of pre- and postmenopausal women with benign breast disease (BBD) from a nested case–control study in the Nurses' Health Study (NHS; ref. 6). Using a novel deep learning approach to quantify the proportion of different tissue types, the authors build on growing literature that early-life body fatness may have a long-term impact on breast tissue composition independent of key risk factors. This study provides new insights into the specific tissue components that may partially mediate the relationship between body size and breast cancer.

What is clear from Oh and colleagues' study is that childhood body fatness tracks with adipose breast tissue. A novel finding, however, comes from their ability to decompose measures of dense area into epithelial and fibrous stromal tissue. Early-life body fatness was significantly associated with lower percentage of fibrous stromal tissue and exhibited differential effects on epithelial tissue based on menopausal status, where there was a negative correlation in premenopausal women and a positive relationship in postmenopausal women. Body mass index (BMI) at age 18 was the only anthropometric factor that significantly influenced epithelial tissue overall, showing evidence of a negative relationship. A negative relationship between current adult BMI and change in BMI since age 18 with epithelial tissue was only seen in premenopausal women. Both childhood and adult body size measures were negatively associated with fibrous stromal tissue, with stronger effects observed for adult body size measures.

The authors used pictograms to measure average body fatness between ages 5 and 10 years, previously validated by others with fair to good correlation (7). Evidence of negative associations between early-life body size and breast cancer is consistently seen regardless of measurement type (absolute vs. relative body size; ref. 1), lending support for the use of pictograms and their observed findings. Oh and colleagues also analyzed recalled weight and height at age 18, collected at baseline before breast biopsy procedures. The consistency of the associations between body size across the life course and breast tissue may stem from the fact that all body measures were recalled at the same time by baseline questionnaire.

There is strong evidence that adult height is a risk factor for breast cancer (8); although there is less consistency with MBD (9–12). The authors did not find evidence of an association between adult height and breast tissue types. These results suggest that height as risk factor for breast cancer may act via a pathway independent of breast tissue composition and be more related to systemic growth factors. However, a previous NHS analysis (3) found an association between adult height and percent MBD in premenopausal (but not postmenopausal) women. Oh and colleagues acknowledge that discrepancies may be attributed to differences between study populations (e.g., BBD vs. healthy women with mammograms). Future studies that address whether selection issues may affect these conflicting results should help in elucidating any associations between height, height growth, and breast tissue composition.

Machine learning techniques, including deep learning algorithms, are becoming a useful tool in improving diagnostic methods in breast and pathology image analysis, like tumor identification and classification and metastasis detection (13–15). Oh and colleagues applied deep learning computational pathology methods, including convolutional neural networks, a class of deep neural networks most commonly used in image analyses, to analyze benign breast biopsy slides. The advantages of using this approach for quantifying percentage of breast tissue types on a microscopic level is that deep learning methods show strong discrimination across heterogeneous pathology images (16) and may have greater accuracy in classification than conventional pathology methods (17). Deep learning methods for mammogram and/or pathology image analysis also have the capability to extract additional features beyond fibroglandular and adipose tissue that may be informative for understanding breast tissue composition and breast cancer risk. Furthermore, these methods can perform high-throughput analysis on large-scale datasets and the use of trained algorithms may reduce potential reader measurement error or variability. Oh and colleagues' methods highlight the potential utility of using deep learning algorithms to aid in discriminating breast tissue types in future studies.

There are several challenges in measuring the complex trajectories of growth and breast tissue changes. Regarding the outcome, a key challenge is related to whether there are repeated measures of tissue types across life stage. With biopsy samples, this is a limitation, which we will discuss below. There are some statistical and other methodologic aspects of this study that should be considered. First, the use of conventional multivariable regression models might be insufficient for evaluating time-varying exposures that are correlated over time. Understanding the complex interplay between body size across the life course and breast tissue composition likely requires more robust causal methods, such as marginal structural models (18). Furthermore, several model covariates occurred after early-life body size, some of which might have contributed to weight change from age 18 to baseline (e.g., smoking and physical activity at baseline), and thus, estimates might be biased by overadjustment. In addition to the statistical considerations, external validity may be affected by the sampling for the study, including the lack of ethnic diversity in the study. By using biopsy samples only from women known not to go on to be diagnosed with breast cancer may also affect the generalizability of the findings, a concept known as “deleting the susceptible” (19). We agree with the authors that examining breast tissue after cancer diagnosis may be reverse causality, but if they were able to study prior benign biopsies in the women who went on to be diagnosed with breast cancer, might be a more appropriate comparison with the group they did examine.

Despite these methodologic limitations, the idea of using novel methods to examine epithelial, stromal, and adipose tissues separately for etiologic purposes in epidemiology may hold promise similar to how studying molecular tumor markers have advanced our understanding of etiology. Breast tissue biopsies from healthy women are not readily available or easy to collect, and thus, alternative approaches are needed to study change in normal breast tissue composition over time. Standard screening methods, including mammography and MRI without background parenchymal enhancement, are limited by the fact that they only provide 2D images of the breast and do not capture the metabolic activity in breast tissue. Furthermore, mammography involves ionizing radiation, and, is thus, not recommended for women under 40 years of age. There are alternative, nonscreening and nondiagnostic, methods that can be used to evaluate structural and metabolic changes in the breast tissue. For example, optical spectroscopy is a noninvasive measurement tool that provides information about bulk tissue properties by capturing the unique red and near-infrared light absorption patterns of breast tissue chromophores (20–22), including chromophores that map to breast density (water/lipid ratio), as well as collagen, total hemoglobin, and saturated hemoglobin. Optical spectroscopy does not involve radiation or breast compression, and can, thus, be used repeatedly starting at a young age to study normal breast tissue composition over the life course. Optical spectroscopy and other novel approaches to measuring breast tissue composition thus offer promise for future breast cancer prevention research and risk assessment.

The different directions in the association between body size and breast cancer risk may seem to make public health messages more complicated as other modifiable factors for breast cancer risk reduction, such as reducing alcohol consumption and increasing physical activity, have a more consistent relationship with breast cancer risk across the life course. However, we have previously estimated that even in a high-risk cohort where breast cancers are more likely to occur at younger ages, maintaining a healthy body size across the life course has a much larger impact on lifetime risk than body size in early life, as most breast cancers, even in an enriched cohort, are diagnosed later in life. The novel methods and findings presented by Oh and colleague should challenge us to continue to investigate how changes in adolescent and early-adult breast tissue structure, including the relationship between adipose, stromal, and epithelial tissue, may influence breast biology, BBD risk, and ultimately breast cancer risk (22).

All authors report grants from NIH outside the submitted work.

1.
Terry
MB
. 
Consistency, now what?
Breast Cancer Res
2017
;
19
:
85
.
2.
Yochum
L
,
Tamimi
RM
,
Hankinson
SE
. 
Birthweight, early life body size and adult mammographic density: a review of epidemiologic studies
.
Cancer Causes Control
2014
;
25
:
1247
59
.
3.
Oh
H
,
Rice
MS
,
Warner
ET
,
Bertrand
KA
,
Fowler
EE
,
Eliassen
AH
, et al
Early-life and adult anthropometrics in relation to mammographic image intensity variation in the Nurses' Health Studies
.
Cancer Epidemiol Biomarkers Prev
2020
;
29
:
343
51
.
4.
Schoemaker
MJ
,
Jones
ME
,
Allen
S
,
Hoare
J
,
Ashworth
A
,
Dowsett
M
, et al
Childhood body size and pubertal timing in relation to adult mammographic density phenotype
.
Breast Cancer Res
2017
;
19
:
13
.
5.
Tseng
M
,
Byrne
C
. 
Adiposity, adult weight gain and mammographic breast density in US Chinese women
.
Int J Cancer
2011
;
128
:
418
25
.
6.
Oh
H
,
Yaghjyan
L
,
Austin-Datta
RJ
,
Heng
YJ
,
Baker
GM
,
Sirinukunwattana
K
, et al
Early-life and adult adiposity, adult height, and benign breast tissue composition
.
Cancer Epidemiol Biomarkers Prev
2021
;
30
:
608
15
.
7.
Must
A
,
Willett
WC
,
Dietz
WH
. 
Remote recall of childhood height, weight, and body build by elderly subjects
.
Am J Epidemiol
1993
;
138
:
56
64
.
8.
Zhang
B
,
Shu
XO
,
Delahanty
RJ
,
Zeng
C
,
Michailidou
K
,
Bolla
MK
, et al
Height and breast cancer risk: evidence from prospective studies and mendelian randomization
.
J Natl Cancer Inst
2015
;
107
:
djv219
.
9.
Brisson
J
,
Morrison
AS
,
Kopans
DB
,
Sadowsky
NL
,
Kalisher
L
,
Twaddle
JA
, et al
Height and weight, mammographic features of breast tissue, and breast cancer risk
.
Am J Epidemiol
1984
;
119
:
371
81
.
10.
Sellers
TA
,
Vachon
CM
,
Pankratz
VS
,
Janney
CA
,
Fredericksen
Z
,
Brandt
KR
, et al
Association of childhood and adolescent anthropometric factors, physical activity, and diet with adult mammographic breast density
.
Am J Epidemiol
2007
;
166
:
456
64
.
11.
Dorgan
JF
,
Klifa
C
,
Shepherd
JA
,
Egleston
BL
,
Kwiterovich
PO
,
Himes
JH
, et al
Height, adiposity and body fat distribution and breast density in young women
.
Breast Cancer Res
2012
;
14
:
R107
.
12.
Jeffreys
M
,
Warren
R
,
Gunnell
D
,
McCarron
P
,
Smith
GD
. 
Life course breast cancer risk factors and adult breast density (United Kingdom)
.
Cancer Causes Control
2004
;
15
:
947
55
.
13.
Geras
KJ
,
Mann
RM
,
Moy
L
. 
Artificial intelligence for mammography and digital breast tomosynthesis: current concepts and future perspectives
.
Radiology
2019
;
293
:
246
59
.
14.
Hamidinekoo
A
,
Denton
E
,
Rampun
A
,
Honnor
K
,
Zwiggelaar
R
. 
Deep learning in mammography and breast histology, an overview and future trends
.
Med Image Anal
2018
;
47
:
45
67
.
15.
Acs
B
,
Rantalainen
M
,
Hartman
J
. 
Artificial intelligence as the next step towards precision pathology
.
J Intern Med
2020
;
288
:
62
81
.
16.
Khosravi
P
,
Kazemi
E
,
Imielinski
M
,
Elemento
O
,
Hajirasouliha
I
. 
Deep convolutional neural networks enable discrimination of heterogeneous digital pathology images
.
EBioMedicine
2018
;
27
:
317
28
.
17.
Brinker
TJ
,
Hekler
A
,
Enk
AH
,
Berking
C
,
Haferkamp
S
,
Hauschild
A
, et al
Deep neural networks are superior to dermatologists in melanoma image classification
.
Eur J Cancer
2019
;
119
:
11
7
.
18.
Robins
JM
,
Hernan
MA
,
Brumback
B
. 
Marginal structural models and causal inference in epidemiology
.
Epidemiology
2000
;
11
:
550
60
.
19.
Stovitz
SD
,
Banack
HR
,
Kaufman
JS
. ‘ 
Depletion of the susceptibles' taught through a story, a table and basic arithmetic
.
BMJ Evid Based Med
2018
;
23
:
199
.
20.
Walter
EJ
,
Knight
JA
,
Lilge
L
. 
A multi-wavelength, laser-based optical spectroscopy device for breast density and breast cancer risk pre-screening
.
J Biophotonics
2017
;
10
:
565
76
.
21.
Blackmore
KM
,
Knight
JA
,
Jong
R
,
Lilge
L
. 
Assessing breast tissue density by transillumination breast spectroscopy (TIBS): an intermediate indicator of cancer risk
.
Br J Radiol
2007
;
80
:
545
56
.
22.
Lilge
L
,
Terry
MB
,
Walter
J
,
Pinnaduwage
D
,
Glendon
G
,
Hanna
D
, et al
Non-invasive optical spectroscopic monitoring of breast development during puberty
.
Breast Cancer Res
2017
;
19
:
12
.