Keywords breast cancer - breast diagnostics - fibroadenoma - ultrasound - elastography
Introduction
Breast cancer is the most commonly diagnosed cancer and the second leading cause of
cancer death among women [1 ]. Breast ultrasound has the capability to differentiate between cystic and solid
masses and can highlight characteristics of solid masses that raise suspicion, necessitating
a biopsy [2 ]. Fibroadenomas are the most common solid benign breast lesion and are typically
encountered in premenopausal women under the age of 40 years but can occur at any
age [3 ]. Clinically they are characterized by a solid, mobile, and well-defined mass that
is often painless [4 ]. If ultrasound findings are in keeping with the diagnosis of a fibroadenoma and
there is no clinical suspicion of a malignant disease, it is proposed that a biopsy
is not required, and the imaging findings are classified as probably benign (BI-RADS
III) with necessary follow-up examinations to demonstrate lack of growth or development
of malignant features [5 ].
Nonetheless, several breast lesions may resemble fibroadenomas upon first examination,
with only discrete distinctive features on ultrasound [6 ]. Phyllodes tumors and other relatively rare fibroepithelial tumors, yet also some
high-degree tumors such as triple negative breast cancer (TNBC) can resemble benign
fibroadenomas during ultrasound with benign morphological characteristics on ultrasound,
thus the differentiation between fibroadenoma and breast cancer is sometimes difficult
[7 ]
[8 ]
[9 ]. Such lesions are initially interpreted as suspicious, which could lead up to ultrasound-guided
core needle biopsy. Between 55% and 85% of these biopsies ultimately yield benign
breast lesions, leading to unwarranted treatments, heightened patient anxiety, and
increased healthcare costs [10 ].
To avoid unnecessary biopsy of benign lesions and to select those lesions that require
biopsy, an accurate characterization of breast lesions with imaging is crucial. In
breast ultrasound, imaging features such as a circumscribed margin, homogeneous echo
texture, parallel orientation, and gently lobulated shape indicate benignity [9 ]
[10 ]. Still, overlapping imaging findings in fibroadenoma-like tumors without these typical
features may mimic several other types of breast masses [11 ].
The objective of this retrospective study is to assess the reliability of ultrasound
for characterizing features of fibroadenomas and malignant tumors of the breast mimicking
fibroadenomas (MTMF) and to identify imaging features that might indicate malignancy.
Methods
Study design
The study was conducted in accordance with the Declaration of Helsinki (as revised
in 2013). Study approval was granted by the local ethical review board on October
21st, 2021. This is a retrospective single-center study of patients who underwent
breast ultrasound of fibroadenomas and MTMF with available histopathological results
between May 2011 and September 2021.
Screening procedure
A full-text-based search was carried out within the department’s electronic medical
records database (Centricity RIS-i, GE Healthcare, Chicago, United States) for all
breast biopsy examinations between May 2011 and September 2021 including the full-text
terms “fibroadenoma” and “phylloides tumor” in the referral diagnosis if the examination
was sent for an ultrasound assessment, or in the radiological report if it was conducted
in-house. Search results were then manually checked for fulfilment of inclusion and
exclusion criteria (Supplementary Table 1). Of 623 screened cases, 421 cases were
included. 202 patients had to be excluded for several reasons ([Fig. 1 ]).
Fig. 1 Flowchart of participants and screening procedure.
Sonographic features evaluation
The ultrasound images of the included patients were reviewed. For an overview of the
examined sonographic features, please refer to [Table 1 ]. In cases with a previous depiction of the relevant lesion, the volume doubling
time (VDT) was calculated following the modified Schwartz formula: VDT = [ln(2) ×
ΔT]/[ln(V2 /V1 )] with T being the time interval between current and prior examination, V1 being the prior and V2 the current tumor volume.
Table 1 List of examined lesion features.
Group
Feature
Unit/level
Localization
Side
Left/right
Skin surface distance
mm
Nipple distance
mm
Sector
1–11
Geometry
Size (horizontal, vertical, sagittal)
mm
Volume
cm3
Height-to-width
–
Growth rate: Schwartz volume doubling time
Days
B-Mode features
Surrounding tissue
Fat/mixed/fibroglandular
Pectoral contact or impression
Yes/no
Echogenicity
Anechoic/hypoechoic/isoechoic/
hyperechoic
Tissue homogeneity
Yes/no
Border sharpness
Sharp/partial/diffuse
Contour
Oval/lobulated/irregular
Hyperechoic rim
Yes/no
Cystic areas
Yes/no
Calcifications
Yes/no
Acoustic shadowing
Hypoechoic/none/hyperechoic
Doppler evaluation
Peritumoral vessel count
N
Intratumoral vessel count
N
Intratumoral vessel density
n/cm3
Feeding vessel
Yes/no
Vascular architecture
None/solitary vessel/tree/chaotic
Vascular distribution
None/peripheral/central/whole Lesion
Strain elastography
Tissue stiffness (qualitative)
Soft/intermediate/hard
Strain pattern
Homogeneous/heterogeneous
Strain halo depth
mm
Histologic data
Histologic data was extracted from the in-house patient data archival system (KIS
Powerchart; Cerner Corporation, North Kansas City, USA) and include specific diagnosis
and histopathologic classification (B2 = benign; B3 = intermediate malignant potential
or B5 = malignant). If available, full resection results were used.
Statistics
All data were stored in Microsoft Excel 16.66.1 (Microsoft; Redmond, USA). The statistical
software that was used was SPSS 27.0 for Windows (SPSS; Chicago, USA), GraphPad Prism
10.0.1 (GraphPad Software LLC; La Jolla, USA) and WEKA (University of Waikato; Waikato,
New Zealand). Results include mean ± standard deviation (SD) and ranges in brackets
or relative frequency (absolute values in brackets). Categorical variables were compared
via pairwise Fisher’s exact test (in case of 2 × 2 tables) or a χ2 test. Contingency tables were used to calculate sensitivity, specificity, positive
(PPV) and negative predictive value (NPV), and likelihood ratio (LR). Results include
95% confidence intervals (CI). Statistical significance was considered for p-values
˂ 0.05. Descriptive statistics for all patients include demographic (age, familial
history of breast cancer, prior breast cancer) and disease-related factors, frequency
of benign, intermediate, and malignant tumors (including subtypes) as well as the
BI-RADS classification after radiological examination.
To determine the ability to correctly classify cases as a) tumors with intermediate
differentiation/malignant tumors (B3 or B5) vs. benign tumors (B2) and b) malignant
(B5) vs. non-malignant tumors (B2 or B3), a J48 decision tree algorithm with 100-fold
cross validation was employed. Results include model classification rate, true positive
(TP) and false positive (FP) rates, precision, and receiver-operating characteristic
(ROC) area under the curve (AUC). For a quantitative ranking of predictors, a gain
ratio merit ranking algorithm with 100-fold cross validation was used. Results include
rank and gain ratio merit as well as odds ratios (OR) and p-values calculated from
a Fisher’s exact test with a cut-off determination following Youden’s J method in
case of continuous variables.
Results
Patient characteristics
Overall, 421 female patients were retrospectively included from May 2011 to September
2021. The average age was 44.7 ± 12.8 years (range: 18.1 to 88.0 years). 21.4% (n
= 90) of cases were histologically verified after lesion follow-up (average observation
period from first diagnosis: 20.4 ± 22.3 months). The rest of the cases were sampled
at first diagnosis. 4.3% (n = 18) of participants had breast cancer more than 5 years
prior to the examination and 5.9% (n = 25) had a familial history of breast cancer.
Examination setting and reasons for biopsy
65.8% (n = 277) of participants were scheduled for routine mammography, while 30.6%
(n = 129) were referred for assessment from an extramural radiology practice and 3.6%
(n = 15) were referred by a gynecologist or general practitioner due to breast-related
symptoms.
28.9% (n = 121) of all lesions were radiologically classified as BI-RADS III, 66.4%
(n = 278) as BI-RADS IV, and 4.8% (n = 20) as BI-RADS V before biopsy ([Fig. 2 ]
a ). The leading reasons for biopsy were a newly developed tumor (51.8%, n = 218), tumor
growth (18.1%, n = 76), suspicious morphology (10.7%, n = 45), or a high-risk constellation
due to known genetic breast-cancer-associated mutations or familial history of breast
cancer (5.9%, n = 25) ([Fig. 2 ]
b ). 78 cases (18.5%) underwent biopsy after an initial decision to follow-up the lesion
due to lesion growth or the development of suspicious imaging characteristics. The
average follow-up interval was 620.6 ± 680.0 days.
Fig. 2 Overview of BI-RADS distribution (a ) and reasons for biopsy (b ).
Histopathologic results
A majority of tumors was benign (86.2%, n = 363) with a minority showing intermediate
(4.3%, n = 18) or malignant (9.5%, n = 40) differentiation. The most common benign
tumors were fibroadenoma (90.9%, n = 319), fibrous-cystic mastopathy (7.1%, n = 25),
and adenosis (2.9%, n = 10). Among the intermediate tumors the cases included adenomyoepitheliomas
(33.3%, n = 6), phyllodes tumors (27.8%, n = 5), papillomatous neoplasia (22.2% n
= 4), atypical ductal hyperplasia (11.1%, n = 2), and sclerosing adenosis (5.6%, n
= 1). Malignant tumors comprised carcinoma of no specific type (NST) (67.5%, n = 27),
invasive mucinous carcinoma (7.5%, n = 3), ductal carcinoma in situ (DCIS) (10.0%,
n = 4), and other subtypes (15.0%, n = 6).
General imaging findings
48.9% of all lesions were found on the right side. Viewed by sector, most tumors were
found within the upper outer sectors of the breast (Supplementary Figure 1). Lesions
were on average located 6.0 ± 4.6 mm below the skin surface and 60.9 ± 27.5 mm from
the nipple.
The average size was 17.5 ± 12.8 mm, 15.2 ± 10.7 mm, and 9.9 ± 6.8 mm, in horizontal,
vertical, and sagittal dimensions, respectively, averaging a volume of 4.4 ± 22.7
cm3 (range 0.1 to 418.1 cm3 ). The average height-to-width ratio was 0.6 ± 0.2.
Radiological assessment of probably benign lesions (BI-RADS III) and lesions suspicious
for malignancy (BI-RADS IV and V) with correlation to histopathological results
Initial radiologic assessment of BI-RADS class III or higher was correlated to the
final histopathological diagnosis with 93.5% of radiological BI-RADS III lesions being
benign, 4.1% intermediate, and 2.4% malignant. Tumors initially assessed as BI-RADS
IV or higher were benign in 79.2% of cases, intermediate in 8.4%, and malignant in
12.4% (p ˂ 0.0001).
Accordingly, for the detection of malignancy in these lesions, the sensitivity was
92.5% (95% CI 80.1 to 97.4%), the specificity was 31.5% (27.0 to 36.3%), the positive
predictive value was 12.4% (9.1 to 16.7%), and the negative predictive value was 97.6%
(93.1 to 99.3%) at a likelihood ratio of 1.4.
Algorithm-based identification of tumors with intermediate differentiation/malignant
tumors (B3 or B5) vs. benign tumors (B2)
Using a J48 decision tree algorithm to identify intermediate and malignant tumors,
the overall correct classification rate was 83.1% (TP rate 83.1%, FP rate 66.3%, precision
79.7%, ROC AUC 59.8%). Accordingly, the sensitivity was 21.4% (95% CI 13.4 to 32.4%),
the specificity was 95.4% (92.7 to 97.2%), the positive predictive value was 48.4%
(32.0 to 65.2%), and the negative predictive value was 85.9% (82.1 to 89.0%) with
likelihood ratio of 4.7.
Algorithm-based identification of malignant (B5) vs. non-malignant tumors (B2 or B3)
Again, using a J48 decision tree algorithm to identify intermediate and malignant
tumors, the overall correct classification rate was 90.0% (TP rate 90.0%, FP rate
70.4%, precision 87.8%, ROC AUC 53.4%). Accordingly, the sensitivity was 22.5% (95%
CI 12.3% to 37.5%), the specificity was 97.1% (94.9 to 98.9%), the positive predictive
value was 45.0% (25.8% to 65.8%), the negative predictive value was 92.3% (89.2 to
94.5%) with a likelihood ratio of 7.8.
Ranking of diagnostic properties among demographic, general, and sonographic tumor
properties
Using a gain-ratio-merit-based ranking method, the predictors contributing most to
correct classification of malignant tumors were presence of a hyperechoic rim, diffuse
lesion border, higher strain elastography halo depth, pectoral contact (i.e., deep
location), an irregular lesion shape, low echogenicity, and chaotic vascular architecture
([Table 2 ]).
Table 2 Highest ranked predictors for correct tumor classification as “malignant”.
Predictor
Rank
Gain ratio merit
Odds ratio (OR)
p-value§
§ Fisher’s exact test
Hyperechoic rim [yes]
1.0 ± 0.0
0.135 ± 0.004
12.08 (5.86 to 24.01)
˂ 0.0001
Border sharpness [irregular]
2.1 ± 0.32
0.057 ± 0.002
7.12 (3.49 to 14.41)
˂ 0.0001
Strain elastography halo depth [mm] (cut-off 2.0 mm)
3.1 ± 0.49
0.054 ± 0.002
7.26 (3.01 to 16.54)
˂ 0.0001
Pectoral contact [yes]
3.8 ± 0.46
0.051 ± 0.003
5.15 (2.63 to 10.05)
˂ 0.0001
Irregular shape
[yes]
5.2 ± 0.41
0.029 ± 0.001
14.17 (2.70 to 74.31)
0.0070
Feeding vessel [yes]
5.8 ± 0.41
0.027 ± 0.002
3.48 (1.63 to 7.42)
0.0020
Echogenicity [low]
7 ± 0.22
0.018 ± 0.001
2.73 (0.91 to 8.18)
0.1106
Vascular architecture [chaotic]
8 ± 0.24
0.015 ± 0.001
6.85 (1.31 to 35.92)
0.0764
Sector [1 ]
[2 ]
[3 ]
[4 ]
9.1 ± 0.26
0.009 ± 0.0
1.30 (0.69 to 2.42)
0.5041
Cystic areas [no]
10.1 ± 0.52
0.008 ± 0.001
2.39 (0.96 to 5.77)
0.0808
Of note, patient age, size (in all dimensions), volume, height-to-width ratio, and
volume doubling time did not affect classification ([Table 3 ]).
Table 3 Lowest ranked predictors for correct tumor classification as “malignant”.
Predictor
Rank
Gain ratio merit
Odds ratio (OR)
p-value§
§ Fisher’s exact test
Schwartz volume doubling time [days] (cut-off: ˂ 277 days)
27.6 ± 1.96
0.0 ± 0.0
2.11 (0.56 to 9.17)
0.4057
Depth [mm] (cut-off: > 7.3 mm)
27.8 ± 1.91
0.0 ± 0.0
2.23 (1.12 to 4.43)
0.0232
Height-to-width ratio (cut-off: > 0.59)
28.3 ± 3.16
0.0 ± 0.0
1.29 (0.40 to 4.15)
> 0.9999
Nipple distance [mm] (cut-off: ˂ 84.7 mm)
30.1 ± 2.2
0.0 ± 0.0
0.24 (0.10 to 0.54)
0.0003
Volume [cm3 ] (cut-off: > 0.8cm3 )
30.9 ± 1.35
0.0 ± 0.0
2.71 (1.31 to 5.64)
0.0072
Size Z [mm] (cut-off: > 12.4 mm)
31.6 ± 1.3
0.0 ± 0.0
1.57 (0.68 to 3.82)
0.3386
Size Y [mm] (cut-off: > 12.3 mm)
32.5 ± 1.16
0.0 ± 0.0
1.90 (0.94 to 3.87)
0.0767
Size X [mm] (cut-off: > 13.3 mm)
33.0 ± 1.87
0.0 ± 0.0
1.86 (0.92 to 3.64)
0.0939
Age [years] (cut-off: > 50.6 years)
35.0 ± 0.0
0.0 ± 0.0
3.01 (1.58 to 5.72)
0.0013
Discussion
This study focused on the reliability of ultrasound for the differentiation of fibroadenomas
and malignant tumors of the breast mimicking fibroadenomas (MTMF) in 421 histologically
verified cases and to determine ultrasound features which are predictive for malignancy.
In our study the average age of the patients was 44.7 ± 12.8 years (range: 11.2 to
88.0 years), which is above the age in which fibroadenomas are most commonly found
[4 ]. Typically, the size of fibroadenomas is 2 cm to 3 cm, although the range might
be between 1 cm to over 10 cm [4 ]. In our study the average size was 1.7 cm ± 1.28 cm, and we demonstrated that the
size of the lesion, height-to-width ratio as well as depth and distance to nipple
were not predictive factors for malignancy. Tumors were most commonly found in the
upper outer quadrant, which is the most frequent breast cancer site [12 ]. On the other hand, we showed that pectoral contact – i.e., deep lesion location
– is a predictor for malignancy.
In our study histopathologically malignant subtypes mimicking fibroadenomas were most
likely hypoechogenic and often showed at least partially diffuse border sharpness
with an irregular shape. These features are typically more pronounced in malignant
lesions, which often exhibit an irregular shape, diffuse margins, hypoechogenicity,
posterior acoustic shadowing, and architectural distortion [13 ]
[14 ]. Regardless, some malignant entities may exhibit overlapping ultrasound features
with fibroadenomas like TNBC, which can present with well-circumscribed margins, indicating
a rapidly proliferating tumor without a significant stromal reaction [15 ]. In our study we showed that the highest ranked feature that indicates malignancy
in such cases is a hyperechoic rim which is associated with malignant tumor cells
infiltrating into adjacent (adipose) tissue and the subsequent host’s inflammatory
reaction [16 ].
Tumor angiogenesis is well known and is associated with malignant breast lesions.
Color doppler ultrasound serves as a useful complementary tool to differentiate between
breast lesions [17 ]. Several studies showed that malignant lesions present more often with detectable
intralesional vascularization [18 ] as well with typical tumoral vessels with an irregular course, sinusoids and arteriovenous
shunts [19 ]
[20 ]. Lee et. al demonstrated that malignant lesions often showed both peripheral and
central vascularity, penetrating vessels, and the presence of branching vessels [21 ]. Our results were consistent with the literature showing that a prominent peripheral
vessel (“feeding vessel”) and a chaotic vascular architecture are predictors for malignancy.
Raza and Baum found that malignant lesions often showed prominent peripheral vessels
and an irregular branching pattern, while benign lesions commonly present avascular
or with only small central vessels or vessels that are located around the periphery
[22 ]. Benign lesions like fibroadenomas can also present with vessels within the lesions
but more commonly with peripheral capsular and central segmental vessels and less
commonly with prominent feeding vessels as shown by Strano et al. [23 ].
Ultrasound elastography is another tool for the characterization of masses in the
breast by measuring tissue stiffness [24 ], and several studies showed that sonoelastography is useful for differentiating
benign from malignant masses [25 ]
[26 ]. In our study we demonstrated that evaluating lesions with elastography provides
additional information and that a larger area of increased stiffness beyond the tumor
surface margin is a strong surrogate parameter for malignancy as previously demonstrated
[27 ]. Malignant breast lesions tend to appear larger on elastograms than on conventional
B-mode images. This might be due to tumor cells infiltrating adjacent tissue [24 ]. For illustrative cases, please refer to [Fig. 3 ].
Fig. 3 Comparison of a fibroadenoma (upper row) in a 21-year-old patient exhibiting a modestly
uniform hypoechoic endotexture, oval shape, parallel orientation to the chest wall,
sharp borders and mild posterior acoustic enhancement (a ), no vascularization (b ), and nonspecific elastographic finding without peritumoral stiffening (c ) and an invasive breast cancer of no specific type (grade 3), (lower row) showing
a marked hypoechoic appearance, oval shape, parallel orientation to the chest wall,
partially irregular borders, posterior acoustic enhancement (d ) as well as feeding vessels (e ) and peritumoral stiffening (f ).
Yet, also of note are those predictors without any significant contribution to the
correct classification of tumors with intermediate or malignant differentiation (histologically
B3 or B5) in a multiparametric evaluation, commonly associated with a higher risk:
age and growth rate. Fibroadenomas are thought to rarely present after the age of
40 years. Therefore, a heightened suspicion of malignancy in newly diagnosed breast
lesions mimicking fibroadenomas is considered necessary [4 ]. Furthermore, breast cancer often displays rapid growth depending on the subtype,
grade, and stage of the tumor, among other things [28 ]. However, we showed that neither age nor doubling time had any impact on correct
classification of fibroadenomas and MTMFs when also weighing other factors. Our results
highlight that there are only a few somewhat robust imaging features in this subset
of breast lesions, and BI-RADS classification may significantly overestimate the likelihood
for intermediate or malignant differentiation in a given lesion. While 71.2 % of lesions
were radiologically classified as BI-RADS IV or V, indicating suspicion for malignancy,
histopathology revealed that 79.2% of these lesions were benign. We found that, when
employing a J48-based machine learning approach to identify malignant tumors (B5),
correct classification and odds ratios were much higher than when trying to identify
B3 und B5 tumors, thus illustrating that B3 lesions (i.e., phyllodes tumors) show
a significant demographic and imaging factor overlap with benign fibroadenomas. Development
of novel predictors – possibly derived from computer-aided approaches – may help solve
this current impasse.
Limitations
We acknowledge that this study presents several limitations. Firstly, our analysis
was performed only at a single center and followed a retrospective approach. Therefore,
some details regarding patient history and clinical symptoms may not have been recorded.
Retrospective evaluation of ultrasonography imaging studies can be challenging as
its accuracy depends on correct initial lesion depiction. Secondly, breast ultrasound
is an imaging method that is very examiner-dependent. In the clinical routine, though,
even ultrasound BI-RADS classification is usually done by 2 consultants before biopsy.
For possible future investigations, a prospective trial with multimodal imaging and
histopathological correlation would be advisable.
Conclusion
Ultrasound-based differentiation between fibroadenomas and malignant tumors of the
breast mimicking fibroadenomas still remains challenging. Our study underscores the
difficulties in accurately classifying these breast lesions, particularly when comparing
BI-RADS classifications with final histopathological results. The most robust ultrasound
features identifying fibroadenoma-mimicking malignant lesions are those also found
in common malignant tumors of the breast such as a hyperechoic rim, irregular border,
perilesional stiffening, pectoral contact, irregular shape, and irregular vascularity.