18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Received: 25 November 2021 Revised: 16 January 2022 Accepted: 24 January 2022
DOI: 10.1002/jbio.202100365

RESEARCH ARTICLE

Collagen signature as a novel biomarker to predict

axillary lymph node metastasis in breast cancer using
multiphoton microscopy

Ye Fang1 | Deyong Kang2 | Wenhui Guo3 | Qingyuan Zhang4 | Shuoyu Xu5 |

Xingxin Huang1 | Gangqin Xi1 | Jiajia He1 | Shulian Wu1 | Lianhuang Li1 |
Xiahui Han1 | Jianhua Chen6 | Liqin Zheng1* | Chuan Wang3* |
Jianxin Chen1*
1
Key Laboratory of OptoElectronic Science and Technology for Medicine of Ministry of Education, Fujian Provincial Key Laboratory of Photonics
Technology, Fujian Normal University, Fuzhou, China
2
Department of Pathology, Fujian Medical University Union Hospital, Fuzhou, China
3
Department of General Surgery, Fujian Medical University Union Hospital, Fuzhou, China
4
Department of Medical Oncology, Harbin Medical University Cancer Hospital, Harbin, China
5
Department of General Surgery, Nanfang Hospital, Southern Medical University, China
6
College of Life Sciences, Fujian Normal University, Fuzhou, China

*Correspondence
Liqin Zheng, Key Laboratory of Abstract
OptoElectronic Science and Technology Accurate identification of axillary
for Medicine of Ministry of Education,
lymph node (ALN) status is crucial
Fujian Provincial Key Laboratory of
Photonics Technology, Fujian Normal for tumor staging procedure and
University, Fuzhou 350007, China. decision making. This retrospec-
Email: lqzheng@fjnu.edu.cn
tive study of 898 participants from
Chuan Wang, Breast Surgery Ward,
two institutions was conducted.
Department of General Surgery, Fujian
Medical University Union Hospital, The aim of this study is to evaluate
Fuzhou 350001, China. the diagnostic performance of clinical parameters combined with collagen sig-
Email: dr_chuanwang@fjmu.edu.cn
natures (tumor-associated collagen signatures [TACS] and the TACS
Jianxin Chen, Key Laboratory of
OptoElectronic Science and Technology
corresponding microscopic features [TCMF]) in predicting the probability of
for Medicine of Ministry of Education, ALN metastasis in patients with breast cancer. These findings suggest that
Fujian Provincial Key Laboratory of TACS and TCMF in the breast tumor microenvironment are both novel and
Photonics Technology, Fujian Normal
University, Fuzhou 350007, China.
independent biomarkers for the estimation of ALN metastasis. The nomogram
Email: chenjianxin@fjnu.edu.cn based on independent clinical parameters combined with TACS and TCMF
yields good diagnostic performance in predicting ALN status.
Funding information
National Natural Science Foundation of
KEYWORDS
China, Grant/Award Numbers: 81700576,
81901787, 82171991; Fujian Major axillary lymph node metastasis, breast cancer, multiphoton microscopy, TACS
Scientific and Technological Special corresponding microscopic features, tumor-associated collagen signatures
Project for “Social Development”, Grant/
Award Number: 2020YZ016002; Special
Funds of the Central Government Guiding
Local Science and Technology

J. Biophotonics. 2022;15:e202100365. www.biophotonics-journal.org © 2022 Wiley-VCH GmbH. 1 of 11

https://doi.org/10.1002/jbio.202100365

2 of 11
Development, Grant/Award Number:
2020L3008; Natural Science Foundation of
Fujian Province, Grant/Award Numbers:
2019J01269, 2020J011008, 2020J01154
1 | INTRODUCTION
Breast cancer is the most commonly diagnosed cancer
and the leading cancer-cause death [1]. Axillary lymph
node (ALN) status is one of the most important prog-
nostic factors in breast cancer. The presence of ALN
metastasis is a strong predictor of recurrence [2]. Accu-
rate identification of ALN status is crucial for tumor
staging procedure and decision making. ALN metasta-
sis is a multifactorial case and the pathogenesis is still
under investigation. Currently, sentinel lymph node
(SLN) biopsy (SLNB) is a standard of care for patients
to determine ALN status and avoid ALN dis-
section (ALND) [3]. SLNB has fewer complications than
ALND, but it is not risk-free surgery [4]. SLNB is an
invasive procedure and has a risk of long-time morbid-
ity. There are some complications happened after
undergoing SLNB such as axillary paresthesia, numb-
ness, lymphedema, axillary seromas and wound infec-
tions [5–7]. Thus, a novel biomarker is urgently needed
to predict metastatic extent of ALN.
Up to now, many institutions have developed pre-
diction tools to identify ALN status. Memorial Sloan-
Kettering Cancer Center (MSKCC) model has been the
most widely validated. In MSKCC model, clinical
parameters (tumor type, lymphovascular invasion
[LVI], tumor size, tumor location, age, multifocality,
estrogen receptor [ER] status and progesterone receptor
[PR] status) have been described as predictors of ALN
metastasis in breast cancer [8]. Clinically, various non-
invasive methods based on imaging examination, such
as ultrasonography, computed tomography (CT), mam-
mography and magnetic resonance imaging (MRI),
were used to detect the axillary nodal status. Ultraso-
nography has been routinely used to assess the ALN
metastatic disease preoperatively. If a suspicious lymph
node is found on imaging examination, patients may
undergo core needle biopsy or US-guided fine-needle
aspiration to obtain a cytologic or histologic diagnosis
[9]. In recent study, some institutions have indicated
the potential value of imaging examination combining
with radiomics to characterize breast lesions and deter-
mine ALN status. Santucci and their colleagues used
radiomics of 3 T MRI combined with histological data
to preoperatively predict ALN status [10]. Yang et al.
developed and validated a mammography-based radio-
mics nomogram for the preoperative prediction of ALN
status [11]. Zheng et al. demonstrated that clinical
18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
FANG ET AL.
parameter combined deep learning radiomic on breast
conventional ultrasonography and shear wave
elastography images provided a practical way for
predicting the extent of ALN involvement [12]. Guo
et al. constructed a model by combining ultrasonogra-
phy with deep learning radiomics, which yielded a good
diagnostic performance in predicting ALN status with
an AUC of 0.860 [13].
Multiphoton microscopy (MPM) is a nonlinear optical
imaging tool by detecting signals generated by the inter-
action of femtosecond laser with internal components of
biological tissue. It provides bioinformation about the tis-
sue structure and cell morphology at micron/submicron
resolution through a combination of second harmonic
generation (SHG) and two-photon excitation fluorescence
(TPEF) [14, 15]. MPM has the potential to be applied to
intraoperative real-time diagnosis and has gradually
transformed from laboratory to clinical practice [16]. Due
to its noncentrosymmetric structure, collagen fibers can
produce SHG signals [17]. SHG has been useful to moni-
tor tumor progression and carcinogenesis, providing
information about complex interactions between cancer
cells and extracellular matrix (ECM) [18]. Moreover,
SHG imaging could be converted into high-dimensional
and quantitative components of collagen via automatic
extraction of multiple features [19].
ECM provides structural and mechanical support to
tissues and cells. ECM involves in tumor progression and
regulates cancer behavior [20, 21]. Collagen is the main
structural component of ECM and has been proven to be
an indicator of tumor growth, metastasis, infiltration and
tumor cell dormancy [22–24]. Conklin et al. demon-
strated that the orientation of collagen fibers around duc-
tal carcinoma in situ could predict recurrence [25].
Kakkad et al. came to a conclusion that the increased of
collagen I fibers was associated with lymph node positive
breast cancer [26]. Keely et al. put forward three different
tumor-associated collagen signatures (TACS) in mouse
mammary gland models as biomarkers in the progression
of mammary carcinoma and proved TACS3 was associ-
ated with low survival rates in human breast cancer
patients [27]. Recently, our group recognized five new
tumor-associated collagen signatures (TACS4-8) in the
primary tumor. We also developed and validated a
TACS1-8 model that could predict individual disease-free
survival rate for breast cancer patients [28]. TACS1-8
strengthened the often-underappreciated role of the
tumor microenvironment in breast cancer prognosis.
 18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
FANG ET AL. 3 of 11

In this study, we extracted the TACS corresponding 2 | METHODS

microscopic collagen feature (TCMF) by converting MPM
images into multiple morphological and texture features
of collagen in microscale. Here we have initially per-
formed a study to test the relationship between ALN
metastasis and TACS and TCMF in breast cancer micro-
environment using MPM method. The purpose of this
study is to evaluate the diagnostic performance of clinical
parameters combined with collagen signatures (TACS
and TCMF) in predicting the extent of ALN involvement
in patients with breast cancer.
2.1 | Patients and ethic
This retrospective study used anonymous data which was
approved by the Institutional Review Board of Fujian Medical
University Union Hospital (Fuzhou, China) and Harbin Med-
ical University Cancer Hospital (Harbin, China). And we con-
firmed that all methods were carried out in accordance with
the relevant guidelines. Because of the retrospective nature of
this study, informed consent of patients was waived and

F I G U R E 1 (A) Patient recruitment pathway. (B) Flow diagram showing the development of the nomogram. The first step: Formalin-
fixed, paraffin-embedded (FFPE) breast tumor samples were collected. Two 5 μm thick continuous sections were cut from each paraffin
sample. The second step: ROIs were selected in the hematoxylin and eosin (H&E)-stained image and the corresponding multiphoton
imaging including TPEF and SHG was obtained. The third step: Calculate the frequencies of each TACS, and capture the most characteristic
areas in ROI for feature extraction, including morphological features and texture features. The fourth step: The TACS score and TCMF score
combined with clinical parameters to construct a nomogram

4 of 11
patients' information was protected. A total of 898 patients
were involved. 679 samples were obtained between
November 2003 and June 2017, from patients treated at the
Fujian Medical University Union Hospital. 219 patients regis-
tered between July 2009 and December 2014 at Harbin Medi-
cal University Cancer Hospital were also obtained.
The inclusion criteria included the followings:
(a) clinicopathological characteristics and follow-up
information were complete; (b) patients underwent
breast surgery and SLNB or ALND. The exclusion criteria
were as follows: (a) incomplete information or images;
(b) unqualified image. The patient recruitment pathway
was shown in Figure 1A.
2.2 | Sample preparation and data
collection
In this study, tissue samples of breast cancer patients,
fixed with formalin and embedded with paraffin, were
used. Two 5 μm thick consecutive sections were cut from
each paraffin sample using an ultra-thin semi-automatic
microtome. One piece of them was used for multiphoton
imaging after dewaxing by alcohol and xylene, and the
another one was stained with H&E staining. Pathologists
confirmed the region of the invasive margin and tumor
center. Several square nonoverlapping regions of inter-
ested (ROIs) were drawn and positioned in the H&E-
stained images (Figure 1B) by two independent panelists
who were blinded to the pathological results of ALN
metastasis. ROIs were across the invasive margin and
adjacent to the tumor area. ROIs with ductal carcinoma
in situ (DCIS) were preferably selected inside the tumor,
while ROIs were extensively sampled at the invasive
front. Because the size of each sample may be slightly dif-
ferent, the number of ROI per slice was range 7 to 20.
The size of each ROI is about 4500  4500 pixels.
Clinical and histopathologic data were obtained from
the medical records [8, 29]. Baseline clinical characteris-
tics were retrospectively collected, including age at surgi-
cal intervention, tumor size (T1, T2 and T3), LVI (present
and absent), histological grade (G1, G2 and G3), ER sta-
tus (positive, negative), PR status (positive, negative) and
HER2 status (positive, negative). The histological grade
(G1, G2 and G3) was evaluated according to the Notting-
ham histological grade [30]. The expression of ER, PR
and HER2 was detected by immunohistochemistry.
2.3 | Multiphoton image acquisition
The imaging system was built on a commercially laser
scanning microscope platform (LSM 880 Zeiss, Germany)
18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
FANG ET AL.
using a mode-locked femtosecond Ti: Sapphire laser (tun-
ing range from 690 to 1064 nm). In our experiments, the
810 nm excitation light was used for multiphoton imag-
ing. The backscattered signals were obtained via two
independent channels at the same time: one channel for
detecting SHG signal (green color) was set between
395 and 415 nm, whereas the other channel for detecting
TPEF signal (red color) was set between 428 and 695 nm.
A Plan-Apochromat 20 objective (NA = 0.8, Zeiss,
Germany) was employed for acquiring large field images
and for collecting the backward signals from tissue sam-
ples. The lateral resolution was ~0.8 μm while the imag-
ing field of view was 0.5  0.5 mm2, typically. Larger
scale imaging was enabled by a motorized stage under
computer control. In this study, multiphoton image
acquisition was performed on unstained section and com-
pared with H&E-stained image.
2.3.1 | Tumor-associated collagen signatures
(TACS) score quantification
According to TACS1-8 model [28], TACS reflected the
distribution of collagen fibers in tumor microenviron-
ment. TACS1-8 were extracted from tumor center and
tumor-stromal interface. The description of TACS1-8 was
shown in Table S3.
Three independent reviewers evaluated the MPM
images which were contained with all feature regions on
each tissue slice. TACS1-8 in each region was recorded
required at least two reviewers rated YES. TACS of the
same kind were recorded only once and different types of
TACS could be recorded multiple times in a single region.
TACS score was the frequency of TACS1-8 appearing in
the effective region. Ridge regression, a regularization
method, was used in this study [31]. All the regressor var-
iables stay in the model because of regression coefficients
do not become exactly zero [32]. We used ridge regres-
sion to calculate the weight value of each TACS. TACS
score was a multiple score which was calculated for each
patient via a combination of the score of each TACS that
was weighted by their, respectively, coefficients. The for-
mula of TACS score was shown in the supplementary
material.
2.3.2 | Tumor-associated collagen signatures
corresponding microscopic features (TCMF)
score quantification
The image tiles with captured with the TACS characteris-
tics on the MPM image were selected. The size of them
were 273  273 pixels. Then, SHG collagen signal was

FANG ET AL.
selected from MPM image. TCMF were computed over
the SHG images from each sample. A total of 142 TCMF
were extracted from SHG images using Matlab 2016b [33,
34]. The features were grouped into morphological (eight
features) and texture features (134 features). Among
142 TCMF, not all of them are associated with ALN
metastasis, and irrelevant features might potentially
lower the prediction quality of predictor. High-
dimensional feature selection could remove redundant
features and retain the most relevant features for building
a predictive model. LASSO (least absolute shrinkage and
selection operator) logistic regression was used to choose
the most predictive factors in the training cohort. LASSO
logistic regression is suitable for high-dimensional data
reduction and particularly well-suited for problems with
a small sample size [35]. At length, 10 TCMF were
selected as the best potential predictors by LASSO logistic
regression on the basis of 444 patients in training cohort,
including four morphological features and six texture fea-
tures. In Table S1 and S2, the categorical concepts of
TCMF and the mathematical definition of TCMF were
described. TCMF score was linear combination of chosen
signatures with their weighted value. The formula of fea-
ture calculation was shown in the supplementary
material.
2.4 | Development and evaluation of
prediction model
Seven clinical parameters (age, tumor size, histological
grade, ER status, PR status, HER2 status and LVI) and
collagen signatures (TACS and TCMF) were included in
the univariate analysis to explore the association with
ALN metastasis in the training cohort, and variables with
P < 0.05 were chosen for the multivariate analysis. Back-
ward stepwise regression was used to choose the inde-
pendent factors. The multicollinearity of the multivariate
model was assessed using the tolerance and variance
inflation factor. A nomogram was constructed according
to independent predictors. The flow diagram of the devel-
opment of the nomogram was shown in Figure 1B. The
predictive accuracy and discriminative ability of the
nomogram was determined by the area under the
receiver operating characteristic curve (AUC). A 95% CI
was calculated for each AUC. Calibration curves were
plotted to evaluate the nomogram model. To assess the
feasibility of the prediction model for ALN metastasis,
the diagnostic performance indices, including sensitivity
and specificity were evaluated. A decision curve analysis
was performed to determine the clinical value of the
nomogram by calculating the net benefits at different
threshold probabilities [36].
18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5 of 11
2.5 | Statistical analysis
A multivariate logistic regression was performed to esti-
mate the odds ratio (OR) with a 95% CI and to identify
the independent predictors for ALN metastasis and to
generate nomogram. Statistical analysis was conducted
with R software, version 3.6.3 (R Foundation for Statisti-
cal Computing) and IBM SPSS Statistics 24. Differences
between the two sets of groups were tested using Mann–
Whitney U test or Chi-square test. All statistics were two-
tailed, and P value <0.05 were considered statistically
significant.
3 | RESULTS
3.1 | Participants
In total 898 patients, 444 of these patients from Fujian
Medical University Union Hospital were selected to the
training cohort and 235 patients to the internal validation
cohort by computer-generated random numbers. The
external validation cohort included 219 patients from
Harbin Medical University Cancer Hospital. Table 1
showed the characteristics of patients in the training,
internal validation and external validation cohorts.
According to the SLNB and ALND results, of all
898 patients, 448 (49.9%) patients had no metastatic bur-
den of axillary disease and 450 (50.1%) patients had
positive ALNs.
3.2 | Association of collagen signature
with ALN metastasis
The correlation between each TACS and ALN metastasis
was assessed by computing the Spearman's rank correla-
tion coefficient. Figure 2 showed the correlation analysis
between individual TACS and ALN metastasis for the
three cohorts. In the training cohort, TACS1, TACS4 and
TACS7 showed negative correlation with ALN metasta-
sis, and TACS4 showed significantly negative correlation
with ALN metastasis, the correlation coefficient of them
was 0.232. Besides, TACS2, TACS3, TACS 5, TACS6
and TACS8 showed positive correlation with ALN metas-
tasis, and TACS5 and TACS6 had significantly positive
correlation with ALN metastasis, the correlation coeffi-
cient is 0.191 and 0.277, respectively. While TACS cor-
relation coefficients exhibited difference among
different cohorts, TACS4, 1 were consistently correlated
with negative ALN status and TACS6, 5 were consis-
tently correlated with positive ALN status. In the binary
logistic regression analysis, TACS score was an
 18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 of 11 FANG ET AL.

TABLE 1 Baseline patients and clinicopathological characteristics

Internal validation External validation

Training cohort cohort cohort

ALN Without ALN Without ALN Without

Variables metastasis metastasis P metastasis metastasis P metastasis metastasis P
Age 0.445 0.064 0.232
≤50 119 127 75 64 64 52
>50 103 95 40 56 48 54
Tumor size <0.0001 0.005 0.001
≤2 cm 64 114 36 62 50 71
2-5 cm 133 101 69 53 59 35
>5 cm 25 7 10 5 4 0
LVI <0.0001 0.025 0.020
Present 61 19 33 11 22 9
Absent 161 203 82 109 91 97
ER status 0.040 0.045 0.730
Positive 149 128 85 74 75 68
Negative 73 94 30 46 38 38
PR status 0.096 0.107 0.219
Positive 129 111 76 67 69 56
Negative 93 110 39 53 44 50
HER2 status 0.187 0.925 0.476
Positive 78 65 39 40 37 30
Negative 144 157 76 80 76 76
Histological grade 0.007 0.005 0.039
G1 24 48 13 26 2 7
G2 130 108 58 69 90 89
G3 68 66 44 25 21 10

Abbreviations: ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; LVI, lymphovascular invasion; PR, progesterone receptor.

FIGURE 2 Correlation analysis between individual TACSs and ALN metastasis for three cohorts

independent risk factor of ALN metastasis. TACS score
indicated a prediction of ALN metastasis with an AUC
of 0.680 (95% CI, 0.631–0.730) in the training cohort,
0.634 (95% CI, 0.563–0.706) in the internal validation
cohort and 0.682 (95% CI, 0.612–0.752) in the external
validation cohort.
The potential association of the TCMF score with
ALN metastasis was first assessed by the binary logistic
regression analysis and receiver operating characteristics
(ROC) analysis in the training cohort and then validated
in the internal and external validation cohorts. TCMF
score was an independent risk factor of ALN metastasis.
 18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
FANG ET AL. 7 of 11

The TCMF score indicated a prediction of ALN metasta-
sis with an AUC of 0.670 (95% CI, 0.620–0.719) in the
training cohort, 0.662 (95% CI, 0.592–0.731) in the inter-
nal validation cohort and 0.638 (95% CI, 0.565–0.711) in
the external validation cohort.
3.3 | Nomogram for predicting risk of
ALN metastasis
The binary logistic regression was used to as univariable and
multivariable analysis to evaluate risk factors in the training
cohort. In univariate analysis, variables that were associated
with ALN metastasis in breast cancer were tumor size, LVI,
ER status, histological grade, TACS score and TCMF score
(Table 2). Furthermore, in multivariable analysis, tumor size,
LVI, histological grade, TACS score and TCMF score were
identified as independent predictors for ALN metastasis
(Table 2). The variance inflation factor of each predictor was
<10, and the corresponding tolerance was more than 0.1.
Therefore, there is no multicollinearity between these predic-
tors (Table S4). The model containing these independent pre-
dictors were established and presented as the nomogram
(Figure 3A). Nomogram is a statistical model that provides
individualizing risk assessment. In the training cohort, the
AUC for nomogram is 0.783 (95% CI, 0.743–0.823). The inter-
nal validation of nomogram produced an AUC of 0.747 (95%
CI, 0.686–0.801). In the external validation cohort, nomogram
achieved an AUC of 0.746 (95% CI, 0.683–0.803). The calibra-
tion curve for the probability of ALN metastasis demonstrated
good agreement between prediction and observation
(Figure 3B). The Hosmer–Lemeshow test demonstrated a

TABLE 2 Univariate and multivariate logistic regression of ALN metastasis in the training cohort

Univariate analysis Multivariate analysis

Variable OR (95% CI) P OR (95%CI) P

Age
≤50 Reference
>50 1.037 0.705 1.527 0.852 NA NA NA NA
Tumor size
≤2 cm Reference
2-5 cm 2.346 1.550 3.522 <0.0001 2.348 1.494 3.692 <0.0001
>5 cm 6.362 2.606 15.527 <0.0001 6.459 2.421 17.230 <0.0001
LVI
Absent Reference
Present 4.048 2.324 7.051 <0.0001 3.702 2.026 6.764 <0.0001
ER status
Negative Reference 1.404 0.885 2.228 0.150
Positive 1.499 1.019 2.206 0.040
PR status
Negative Reference
Positive 1.412 0.971 2.054 0.071 NA NA NA NA
HER2 status
Negative Reference
Positive 1.308 0.878 1.951 0.187 NA NA NA NA
Histological grade
G1 Reference NA NA NA NA
G2 2.407 1.386 4.183 0.002 2.000 1.079 3.708 0.028
G3 2.061 1.136 3.738 0.017 1.947 0.980 3.868 0.057
TACS score 2.841 2.055 3.929 <0.0001 2.101 1.432 3.082 <0.0001
TCMF score 3.640 2.400 5.520 <0.0001 2.634 1.600 4.337 <0.0001

Abbreviations: ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; LVI, lymphovascular invasion; OR, odds ratio; PR, progesterone
receptor; TACS, tumor-associated collagen signatures; TCMF, TACS corresponding microscopic features.
 18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 of 11 FANG ET AL.

F I G U R E 3 (A) Establishing a nomogram that combines clinical variables (tumor size, LVI, histological grade), with TACS score and
TCMF score to predict the risk of having any positive ALNs. (B) Calibration curves of the radiomics nomogram in each cohort.
(A) Calibration curve of the nomogram in the training cohort. (B) Calibration curve of the nomogram in the internal validation cohort.
(C) Calibration curve of the model in the external validation cohort

nonsignificant statistic (training cohort, P = 0.815; internal
validation cohort, P = 0.610; external validation cohort,
P = 0.287), which suggested that there was no departure from
perfect fit. In the training cohort, the maximum Youden
index of 0.4279 was selected as the cutoff value, and the
cohort had a sensitivity of 84.23%, a specificity of 58.56%. The
internal validation cohort had a sensitivity of 74.78%, a speci-
ficity of 65.83%. The external validation cohort had a sensitiv-
ity of 63.72% and a specificity of 76.42% (Table S5).
efficiency of the nomogram combined with clinical
parameters, TACS and TCMF was significantly higher
than clinical model alone (0.783 vs. 0.700 [95% CI, 0.670–
0.758]) in the training cohort, 0.747 vs. 0.664 [95% CI,
0.600–0.724] in the internal validation cohort and 0.746
vs. 0.651 [95% CI, 0.583–0.714] in the external validation).
Figure 4 showed the comparison of ROC curves between
different models for predicting disease-free axilla and any
ALN metastasis in the training and validation cohort.

3.4 | Comparison with the clinical model 3.5 | Decision curve analysis

The clinical model was built on the basis of clinical The decision curve analysis for the clinical model and
parameters (tumor size, LVI and histological grade). nomogram was presented (Figure 5) to evaluate the
Compared with the clinical model, the predictive added clinical utility of nomogram model in predicting

FANG ET AL.
F I G U R E 4 Comparison of receiver operating characteristic (ROC) curves between different models for predicting disease-free axilla and
any ALN metastasis in the training cohort, internal validation cohort and external validation cohort
F I G U R E 5 Decision curve for the nomogram and clinical
parameters. The y-axis represents the net benefit. The green line
represents the nomogram. The red line represents the clinical
parameters. The gray line represents the hypothesis that all patients
had ALN metastases. The black line represents the hypothesis that
no patients had ALN metastases. The x-axis represents the
threshold probability
ALN metastasis. This analysis indicated that, when the
threshold probability was within the range 0 to 0.9, using
the nomogram to predict ALN metastasis added more net
benefit than the treat-all or treat-none strategies.
4 | DISCUSSION
Accurate assessment of the extent of ALN metastasis
involvement is vital for tumor staging and decision mak-
ing. Clinically, SLNB remains the golden standard in axil-
lary staging in node-negative breast cancer. Axillary
treatment is undergoing a paradigm shift and several
researchers are being verified on whether SLNB may be
18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
9 of 11
omitted in low-risk patients. Up to now, many studies
have developed noninvasive method based on imaging
examination to confirm the nodal status in preoperative,
such as ultrasonography, CT, mammography and MRI.
Those techniques combined with radiomics could be
used to preoperatively characterize breast lesions and
determine ALN status. In our study, we aimed to develop
a high diagnostic sensitivity model to predict the proba-
bility of ALN metastasis based on multiphoton imaging.
We developed and validated a nomogram based on clini-
cal parameters combined with TACS and TCMF. The
nomogram showed diagnostic performances in predicting
ALN status. Compared with the clinical model, signifi-
cant improvement in the AUC was observed in the
nomogram.
Previous studies developed noninvasive method based
on imaging examination to confirm the nodal status in
preoperative, and showed that multi-modal multi-sources
features fusion could be conducted thus enhancing the
diagnostic performance of predicting nodal status. Sev-
eral groups have reported the possible intraoperative
applications of optical imaging means in different organs,
especially photoacoustic imaging [37, 38]. MPM method
is also a promising method for realizing real-time in vivo
optical biopsy and can be applied to intraoperative diag-
nosis [16]. It provides bioinformation about the tissue
structure and cell morphology at micron/submicron reso-
lution [39]. The MPM distinguishability is high enough
in imaging collagen fiber morphology. In this study, we
detected collagen fibers in breast tumor samples using
MPM method. The percentages of TACS1-8 in each
patient were calculated to represent the large-scale fea-
tures of collagen. Furthermore, the microscopic features
of collagen including morphological and texture features
from multiphoton images were extracted after image
processing. Our results showed that both TACS and

10 of 11
TCMF are independent factor of ALN metastasis. Cur-
rently, we used postoperative specimens to capture MPM
images. If our results confirmed in a future, larger study,
we foresee real-time intraoperative applications of MPM
fused with other imaging examinations would have the
potential to determine appropriate axillary treatment
options for patients with breast cancer. Our study also
found that LVI could be detected by MPM imaging in
breast tissue samples (Figure S1). It showed the potential
that LVI could be detected by scanning the surgical mar-
gin of breast tumor in vivo using MPM imaging during
mastectomy. Scanning the surgical margin of breast tumor
in vivo by MPM has prospective biotechnological applica-
tion, especially in the field of intraoperative biopsy.
Chen et al. reported that collagen signature in tumor
microenvironment was closely associated with lymph
node metastasis in early gastric cancer. Straighter colla-
gen and high cross-link density were correlated with pos-
itive nodal status [19]. In addition, Kakkad et al. reported
that high collagen I density was associated with positive
nodal breast cancer [26]. Previous studies showed that
detection and quantification of collagen fibers had the
potential to explore the role of collagen in cancer meta-
static processes. In our study, TACS and TCMF were both
novel biomarkers to predict ALN status in breast cancer.
TACS4, 5, 6 were significantly associated with ALN sta-
tus, TACS4 showing negative correlation and TACS5,
6 showing positive correlation. The TACS score emerged
as a tumor microenvironment based structural prognosti-
cator. Because calculating TACS score relied on a qualita-
tive yes/no judgment performed by trained persons
without software involvement or quantification, we used
automated software that quantified microscopic collagen
signatures. Compared with TACS score, TCMF score was
reflected the collagen signature in micro-scale. In our
study, collagen fiber straightness (TCMF5) and orienta-
tion (TCMF8) were important signatures in ALN status
predicting. The straighter collagen fiber and chaotic ori-
ented collagen fiber signified higher risk in ALN metasta-
sis. Integrating multiple biomarkers into a single
signature, rather than performing individual biomarker
analysis, is a promising approach that would improve
clinical management [40]. In this study, collagen signa-
ture combined with clinical parameters that held higher
value than a single biomarker.
This study had its limitations. It was a retrospective
study subject to potential selection biases, although a
large number of patients from multi-centers were
included. Moreover, this nomogram was based on Chi-
nese patients alone. Hence, the nomogram should also be
validated by western population. Despite these limita-
tions, we believe that our study has demonstrated the
potential use of MPM method for clinical practice.
18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
FANG ET AL.
5 | CONCLUSION
In conclusion, we developed and validated a diagnostic
nomogram for the individualized prediction of ALN
metastasis in patients with breast cancer. The current
study indicated that collagen signatures (TACS and
TCMF) in the tumor microenvironment could be novel
biomarkers associated with ALN metastasis. With further
study, extracting collagen signature from MPM images
will stand the good chance to apply in a clinical trial.
ACKNOWLEDGMENTS
This study was supported by the Natural Science Founda-
tion of China (Nos. 82171991, 81901787, 81700576),
Fujian Major Scientific and Technological Special Project
for “Social Development” (No. 2020YZ016002), Special
Funds of the Central Government Guiding Local Science
and Technology Development (No. 2020L3008), Natural
Science Foundation of Fujian Province (Nos. 2019J01269,
2020J011008, 2020J01154).
CONFLICT OF INTEREST
The authors declare no conflict of interest.
AUTHOR CONTRIBUTIONS
Conception and design were provided by Jianxin Chen.
Development of methodology were contributed by Ye
Fang, Shuoyu Xu, Gangqin Xi, Xingxin Huang and Jiajia
He. Acquisition of data were supplied by Chuan Wang,
Qingyuan Zhang, Wenhui Guo and Deyong Kang. Analy-
sis and interpretation of data were done by Ye Fang.
Writing, review, and/or revision of the manuscript were
imparted by Ye Fang, Liqin Zheng, Lianhuang Li, Xiahui
Han, Jianhua Chen and Jianxin Chen. Administrative,
technical, or material supports were provided by Ye
Fang, Wenhui Guo and Deyong Kang. Study supervision
was supplied by Liqin Zheng and Jianxin Chen.
DA TA AVAI LA BI LI TY S T ATE ME NT
The data that support the findings of this study are avail-
able from the corresponding author upon reasonable
request.
ORCID
Jianxin Chen https://orcid.org/0000-0001-8519-4462
RE FER EN CES
[1] H. Sung, J. Ferlay, R. L. Siegel, M. Laversanne, I. Soerjomataram,
A. Jemal, F. Bray, CA A Cancer J Clin 2021, 71(3), 209.
[2] H. E. Campbell, A. M. Gray, A. L. Harris, A. H. Briggs, M. A.
Taylor, Br. J. Cancer 2010, 103(6), 776.
[3] G. H. Lyman, A. E. Giuliano, M. R. Somerfield, A. B. Benson,
D. C. Bodurka, H. J. Burstein, A. J. Cochran, H. S. Cody, S. B.
Edge, S. Galper, J. A. Hayman, T. Y. Kim, C. L. Perkins, D. A.

FANG ET AL.
Podoloff, V. H. Sivasubramaniam, R. R. Turner, R. Wahl, D. L.
Weaver, A. C. Wolff, E. P. Winer, J. Clin. Oncol. 2005, 23(30),
7703.
[4] H. Verry, S. J. Lord, A. Martin, G. Gill, C. K. Lee, K. Howard,
N. Wetzig, J. Simes, Br. J. Cancer 2012, 106(6), 1045.
[5] T. Ashikaga, D. N. Krag, S. R. Land, T. B. Julian, S. J.
Anderson, A. M. Brown, J. M. Skelly, S. P. Harlow, D. L.
Weaver, E. P. Mamounas, J. P. Costantino, N. Wolmark,
J. Surg. Oncol. 2010, 102(2), 111.
[6] E. D. Gournay, A. Guyomard, C. Coutant, S. Boulet, P.
Arveux, S. Causeret, S. Gouy, M.-M. Padeano, C. Loustalot,
J.-M. Sauzedde, M. Smail, J.-P. Combier, P. Chevillote, C.
Rosburger, F. Bonnetain, J. Fraisse, T. S. Dabakuyo-Yonli, Br.
J. Cancer 2013, 109(11), 2783.
[7] A. Lucci, L. M. Mccall, P. D. Beitsch, P. W. Whitworth, D. S.
Reintgen, P. W. Blumencranz, A. M. Leitch, S. Saha, K. K.
Hunt, A. E. Giuliano, J. Clin. Oncol. 2007, 25(24), 3657.
[8] J. L. B. Bevilacqua, M. W. Kattan, J. V. Fey, H. S. Cody, P. I.
Borgen, K. J. Van Zee, J. Clin. Oncol. 2007, 25(24), 3670.
[9] J. Cools-Lartigue, S. Meterissian, World J. Surg. 2012,
36(1), 46.
[10] D. Santucci, E. Faiella, E. Cordelli, R. Sicilia, C. de Felice,
B. B. Zobel, G. Lannelo, P. Soda, Cancer 2021, 13(9), 2228.
[11] J. Yang, T. Wang, L. Yang, Y. Wang, H. Li, X. Zhou, W. Zhao,
J. Ren, X. Li, J. Tian, L. Huang, Sci. Rep. 2019, 9(1), 4429.
[12] X. Zheng, Z. Yao, Y. Huang, Y. Yu, Y. Wang, Y. Liu, R. Mao,
F. Li, Y. Xiao, Y. Wang, Y. Hu, J. Yu, J. Zhou, Nat. Commun.
2020, 11(1), 1236.
[13] X. Guo, Z. Liu, C. Sun, L. Zhang, Y. Wang, Z. Li, J. Shi, T. Wu,
H. Cui, J. Zhang, J. Tian, J. W. Tian, EBioMedicine 2020, 60,
103018.
[14] W. R. Zipfel, R. M. Williams, R. Christie, A. Y. Nikitin, B. T.
Hyman, W. W. Webb, Proc Natl Acad Sci USA 2003, 100(12),
7075.
[15] J. Chen, S. Zhuo, X. Jiang, X. Zhu, L. Zheng, S. Xie, B. Lin, H.
Zeng, J. Biomed. Opt. 2011, 16(5), 051305.
[16] M. Jain, N. Narula, A. Aggarwal, B. Stiles, M. M. Shevchuk, J.
Sterling, B. Salamoon, V. Chandel, W. W. Webb, N. K. Altorki,
S. Mukherjee, Archs Pathol Lab Med 2014, 138(8), 1037.
[17] E. Brown, T. McKee, E. diTomaso, A. Pluen, B. Seed, Y.
Boucher, R. K. Jain, Nat. Med. 2003, 9(6), 796.
[18] K. Burke, P. Tang, E. Brown, J. Biomed. Opt. 2013, 18(3),
031106.
[19] D. Chen, G. Chen, W. Jiang, M. Fu, W. Liu, J. Sui, S. Xu, Z. Liu,
X. Zheng, L. Chi, D. Lin, K. Li, W. Chen, N. Zuo, J. Lu, J. Chen,
G. Li, S. Zhuo, J. Yan, JAMA Surg. 2019, 154(3), e185249.
[20] C. Buchheit, K. Weigel, Z. Schafer, Nat. Rev. Cancer 2014,
14(9), 632.
[21] A. Bergamaschi, E. Tagliabue, T. Sørlie, B. Naume, T. Triulzi,
R. Orlandi, H. G. Russnes, J. M. Nesland, R. Tammi, P.
Auvinen, V.-M. Kosma, S. Ménard, A.-L. Børresen-Dale,
J. Pathol. 2008, 214(3), 357.
[22] A. Lochter, M. J. Bissell, Semin Cancer Biol 1995, 6(3), 165.
[23] P. P. Provenzano, D. R. Inman, K. W. Eliceiri, J. G. Knittel, L.
Yan, C. T. Rueden, J. G. White, P. J. Keely, BMC Med. 2008, 6, 11.
18640648, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/jbio.202100365 by Thirion Paul - Dge, Wiley Online Library on [02/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
11 of 11
[24] J. S. Di Martino, A. R. Nobre, C. Mondal, I. Taha, E. Farias, E.
Fertig, A. Naba, J. Aguirre-Ghiso, J. J. Bravo-Cordero, Nat
Cancer 2022, 3, 90.
[25] M. W. Conklin, R. E. Gangnon, B. L. Sprague, L. Van Gemert, J. M.
Hampton, K. W. Eliceiri, J. S. Bredfeldt, Y. Liu, N. Surachaicharn,
P. A. Newcomb, A. Friedl, P. J. Keely, A. Trentham-Dietz, Cancer
Epidemiol Biomarkers Perv 2018, 27(2), 138.
[26] S. M. Kakkad, M. Solaiyappan, P. Argani, S. Sukumar, L. K.
Jacobs, D. Leibfritz, Z. Bhujwalla, K. Glunde, J. Biomed. Opt.
2012, 17(11), 116017.
[27] P. P. Provenzano, K. W. Eliceiri, J. M. Campbell, D. R. Inman,
J. G. White, P. J. Keely, BMC Med. 2006, 4, 38.
[28] G. Xi, W. Guo, D. Kang, J. Ma, F. Fu, L. Qiu, L. Zheng, J. He,
N. Fang, J. Chen, J. Li, S. Zhuo, X. Liao, H. Tu, L. Li, Q. Zhang,
C. Wang, S. A. Boppart, J. Chen, Theranostics 2021, 11(7), 3229.
[29] L. Dihge, P.-O. Bendahl, L. Rydén, Br. J. Surg. 2017, 104(11),
1494.
[30] M. H. Galea, R. W. Blamey, C. E. Elston, I. O. Ellis, Breast
Cancer Res Treat 1992, 22(3), 207.
[31] M. N. Çiftsüren, S. Akkol, Arch Anim Breed 2018, 61(3), 279.
[32] A. Acharjee, R. Finkers, R. G. Visser, C. Maliepaard, Met-
abolomics 2013, 3(3), 1000126.
[33] S. Xu, Y. Wang, D. C. Tai, S. Wang, C. L. Cheng, Q. Peng, J.
Yan, Y. Chen, J. Sun, X. Liang, Y. Zhu, J. C. Rajapakse, R. E.
Welsch, P. T. C. So, A. Wee, J. Hou, H. Yu, J. Hepatol. 2014,
61(2), 260.
[34] S. Xu, C. Kang, X. Gou, Q. Peng, J. Yan, S. Zhuo, C. Cheng, Y.
He, Y. Kang, W. Xia, P. T. C. So, R. Welsch, J. C. Rajapakse,
H. Yu, J. Biophotonics 2016, 9(4), 351.
[35] W. Sauerbrei, P. Royston, H. Binder, Stat. Med. 2007, 26(30),
5512.
[36] K. F. Kerr, M. D. Brown, K. Zhu, H. Janes, J. Clin. Oncol. 2016,
34(21), 2534.
[37] Q. Yu, S. Huang, Z. Wu, J. Zheng, X. Chen, L. Nie, J. Nucl.
Med. 2020, 61(7), 1079.
[38] F. Duan, H. Ma, J. Zhang, S. Li, H. Li, Z. Wu, F. Hong, L.
Zeng, L. Nie, Chin Opt Lett 2020, 18(12), 121701.
[39] A. Zoumi, A. Yeh, B. J. Tromberg, Proc. Natl. Acad. Sci.
U. S. A. 2002, 99(17), 11014.
[40] Y. Huang, C. Liang, L. He, J. Tian, C. Liang, X. Chen, Z. Ma,
Z. Liu, J. Clin. Oncol. 2016, 34(18), 2157.
SU PP O R TI N G I N F O RMA TI O N
Additional supporting information may be found in the
online version of the article at the publisher's website.
How to cite this article: Y. Fang, D. Kang,
W. Guo, Q. Zhang, S. Xu, X. Huang, G. Xi, J. He,
S. Wu, L. Li, X. Han, J. Chen, L. Zheng, C. Wang,
J. Chen, J. Biophotonics 2022, 15(6), e202100365.
https://doi.org/10.1002/jbio.202100365