Received: 23 May 2023 Revised: 25 September 2023 Accepted: 10 October 2023

DOI: 10.1002/cjce.25130

NOTE

Automated nanofibre sizing by multi-image processing

and deep learning with revised UNet model

Hesheng Yu 1,2,3 | Erqian Gao 2 | Zhongchao Tan 1,2

1
Eastern Institute for Advanced Study,
Eastern Institute of Technology, Abstract
Zhejiang, China Nanofibres have been widely used in many chemical engineering applications
2
Department of Mechanical & and their performance greatly depends on the size distribution of the nanofibres.
Mechatronics Engineering, University of
Researchers have developed automated tools to determine nanofibre diameters,
Waterloo, Waterloo, Ontario, Canada
3
School of Chemical Engineering, China
primarily using commercial MATLAB software package. However, no
University of Mining & Technology, researchers have reported automatic processing of multiple images, which is
Xuzhou, China essential to the consistency and accuracy of results. Nor has anyone reported
Correspondence nanofibre sizing using deep learning. Therefore, this paper reports an automated
Zhongchao Tan, Eastern Institute for tool to measure the size distribution of electrospun nanofibres by simultaneous
Advanced Study, Eastern Institute
multi-image processing. This tool determines the diameters of nanofibres using
of Technology.
Email: zhongchaotan@eias.ac.cn deep learning based on UNet model. Results show that the UNet-based deep
learning approach is more accurate than those obtained using existing methods,
Funding information
compared to experimental data.
GCI Ventures, Toronto; University of
Waterloo
KEYWORDS
deep learning, electrospinning, multi-image processing, nanofibre, size distribution

1 | INTRODUCTION therefore expedites the dissolution of water-soluble

Nanofibres have been extensively studied in the past two
decades aiming at applications to energy, environment,
health care, and so on. For example, nanofibres can be
used as electrode or separator materials in solar cells,[1]
fuel cells,[2] electrochemical hydrogen production,[3] and
batteries.[4,5] The high porosity and flexibility of nanofi-
bre structures enable them to outperform other materials
in energy applications.[6,7] Nanofibres are also used for
environmental protection. The sizeable surface-area-to-
volume ratio and large porosity of nanofibres enable
them to absorb more contaminants.[8,9] Regarding health,
nanofibres can also be used as drug delivery media and
wound dressings in the biomedical field. The high surface
area to volume ratio of nanofibres also helps increase the
contact area of drugs with the vivo environment, and
drugs.[10,11] Therapeutic agents can be either loaded in
the post-treatment of nanofibres[12] or added in a polymer
solution before electrospinning.[13] Electrospun nanofibres
overlapping scaffolds are similar to extracellular matrix in
terms of structure, giving them an advantage as wound
dressings. The structure of nanofibre naturally mimics the
extracellular matrix of the human body. Compared to tradi-
tional wound dressings, nanofibres are more favourable to
body tissue and have better cell attachment and prolifera-
tion, thus lowering the possibility of scar tissue formation.[14]
Nanofibres can also be used for circulating tumour cell
analysis for the early diagnosis of cancer, evaluating cancer
progression and assessing treatment efficacy.[15]
The size (a.k.a., diameter) distribution of nanofibres
plays a vital role in the physical and chemical properties of
nanofibre mats.[16,17] The nanofibre morphology depends

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided
the original work is properly cited.
© 2023 The Authors. The Canadian Journal of Chemical Engineering published by Wiley Periodicals LLC on behalf of Canadian Society for Chemical Engineering.

Can J Chem Eng. 2023;1–13. wileyonlinelibrary.com/journal/cjce 1


2 YU ET AL.
on many factors, including polymer molecular weight,[18–21]
solution concentration,[22] solvent volatility,[23] electrospin-
ning voltage,[24] solvent flow rate,[25] and needle-to-collector
distance.[25] In this context, accurate and effective charac-
terization is particularly important. The characterization
of morphology can offer not only information about fibre’s
properties, but also guidance for fabrication protocol
optimization.[26]
However, the current nanofibre characterization tech-
niques are still in their infancy. Currently, the nanofibre
diameters are usually determined by single-image proces-
sing. Manual measurement is time-consuming, restrict-
ing the number of samples processed. More importantly,
subjective manual measurement makes it difficult to
compare different results, based on which important
conclusions are drawn and distributed to the society at
large. Alternatively, Tomba et al.[27] developed a multivari-
ate image analysis (MIA) system for the determination of
nanofibre diameters. The MIA can extract essential infor-
mation from a scanning electron microscope (SEM) image
and reduce the data size and data dimensionality. Multiway
principal component analysis (MPCA) is also performed on
the SEM image. The MPCA linearly combines the original
variables with principal components to help identify the
greatest variability in data. However, it is noted that the
MPCA cannot effectively separate fibres from each other in
the overlapped areas, resulting in deviation. MPCA also
assumes that all fibres are straight, which can be a signifi-
cant source of error. Finally, the thin fibres in low-resolution
images are likely to be overlooked because of repeated
denoising operations on single images.
To tackle the challenge of overlapping fibres,
Ho Shin et al.[28] reported another method for automati-
cally measuring nanofibre diameters using the Canny edge
detection to enhance the accuracy and effectiveness in
separating overlapped fibres in SEM images. However,
Canny edge detection cannot be used on highly oriented
nanofibres. When dealing with high-oriented fibres, this
method can easily outline the fibre boundaries, but has dif-
ficulty distinguishing the fibre sides and the background.
Thus, Murphy et al.[29] proposed an alternative method for
automatically measuring nanofibre diameter. This method
consists of the following three main steps: (1) morphological
reconstructing, (2) image skeletonization, and (3) fibre
diameter calculation. First, the morphologically reconstruct-
ing binarizes the image using the Otsu method.[30] Then,
the centreline of the fibre is determined and reduced into
1 pixel. Finally, the diameter is calculated by doubling the
distance between the fibre boundary and fibre centreline.
Alternatively, Dehghan et al.[31] used the fuzzy clustering
method (FCM) to determine the nanofibre diameters. The
SEM image is segmented into a binary image by FCM
method. The optimal cluster value is determined by trying
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
out different clustering numbers ranging from one to six.
The optimal clustering value is the one that results in the
most similarity between the binary image and the original
image. Finally, the image similarity is calculated using the
structural similarity index method.
Image binarization is a major point of difference for all
the earlier studies. The algorithm for the image binariza-
tion can be divided into two categories: global threshold
and local threshold. The global threshold chooses a single
threshold value, and all values smaller than the threshold
value are regarded as 0, corresponding to pure black.
All values greater than the threshold value are regarded
as 255, corresponding to pure white. Conversely, local
threshold is the method in which multi-threshold values
are used at different spots on the image. It is a necessary
means for images with uneven luminance. The global and
local thresholds are not totally opposite. Many global
threshold-based methods can also be applied locally. To
locally apply a global threshold algorithm, the image needs
to be first divided into many sub-images. Then the global
threshold algorithm is applied to each sub-image, and the
result is obtained by stitching together all sub-images. That
is why some algorithms can be applied either locally or
globally. With the recent advances in algorithms, the third
category appears. It includes the clustering algorithm, such
as K-means and FCM,[31] using the FCM to measure the
diameter of fibres.
Existing image processing platforms for nanofibre
characterization can be divided into commercial software
and open-source platforms. MATLAB is the most widely
used commercial software, which has been used by many
researchers such as Ho Shin’s group,[28] Tan’s group,[32–37]
MIA,[27] and SIMpoly.[29] The most used open-source
software is ImageJ[38] and its plug-in DiameterJ.[39] Both
commercial software and open-source software have
their own advantages and disadvantages.
The determination of nanofibre diameters consists of
the following three steps: (1) image segmentation, (2)
skeletonization, and (3) distance measurement. The
image segmentation can be done by deep learning. There
are three main categories for image segmentation in the
deep learning field: (1) semantic segmentation, (2) instance
segmentation, and (3) panoptic segmentation. Semantic
segmentation is the most basic type of segmentation that
identifies different objects in an image. The objects can
be a person, book, flower, car, and so forth. Typical
algorithms for the semantic segmentation include
UNet,[40] DeepLab,[41,42] and FCN.[43] An instance segmenta-
tion combines target detection and semantic segmentation.
The instance segmentation first identifies the bounding
box of the instance by target detection. Then it performs
the semantic segmentation in the detection box, and
each segmentation result is output as a different instance.

YU ET AL.
The instances of each semantic category are often regarded
as an extension of the semantic segmentation. For
example, if the semantic segmentation method identifies a
flock of birds in an image, then the instance segmentation
further eases the object detection by identifying individual
birds in the flock. Frequently used algorithms for instance
segmentation include Mask R-CNN,[44] YOLCAT,[45] and
Gated-SCNN.[46] The panoptic segmentation is an emerging
method compared to the other two methods,[47] first
proposed by Kirillov et al. This algorithm combines the
semantic and instance segmentations and generates a
coherent scene segmentation.
In this work, semantic segmentation is sufficient in
the context of nanofibre morphology analysis because
there is only one object category (i.e., fibre) throughout
this work. Locating the classified items is unneeded. In
addition, all fibres are equally analyzed for characteriza-
tion and distinguishing the fibres is unnecessary. For
these two reasons, the semantic segmentation is enough
for this project. The algorithms of semantic segmentation
reported earlier include UNet,[40] Deeplab,[42] and FCN.[43]
Among all these algorithms, UNet is one of the earliest
segmentation algorithms that has been repeatedly studied
and implemented. Thus, the principles and implementa-
tion paths of UNet are relatively clear, and useful informa-
tion is easily available in the literature. In addition, UNet
is well known for its low requirement of training data and
short training time, which can significantly reduce the
time for preparing the database and reduce the require-
ments for computer hardware. For example, Ronneberger
et al.[40] obtained a high accuracy with R2 of 0.9203 using
the PhC-U373 dataset, which contains only 35 partially
annotated images for training. Furthermore, in the ISBI
2015 Challenge, the UNet model outperformed all other
models with an R2 of 0.7756 on DIC-HeLa dataset, while
the second-best algorithm can only achieve an R2 of 0.46.
To the best of our knowledge, however, no one has
attempted to determine the nanofibre diameters (i.e., sizes)
by deep learning, particularly, the UNet model. This paper
therefore introduces an automated tool for the characteriza-
tion of nanofibres by simultaneously processing multiple
SEM images. The basic procedure is inspired by earlier
works.[28,34–37] This work contributes to the field by two
improvements: one is the ability to simultaneously process
multiple images, and the other is the optimized denoising
procedure. This code in this work, solved using MATLAB,
has the following two improvements.
The first improvement is to increase the number of
images that can be processed at the same time. This
feature allows end users to simultaneously process
multiple SEM images with different magnifications and
to combine the results for accurate presentation. Our
research group has realized that the accuracy of calculated
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
3
nanofibre diameters depends on the magnification of
the SEM image used for processing. Normally, the SEM
image with a higher magnification gives a smaller mean
diameter with a smaller standard deviation, and vice
versa. Thus, more statistically reliable data can be
obtained by combining the results from the images with
different magnifications. After all, combining the results
of different images still offers a larger sample area and
more sampling points, leading to a more accurate result,
even if the SEM images of the sample are obtained
under the same magnification.
The second improvement is to optimize the denoising
step. The earlier works only use linear filtering, that is, the
median filter on the original SEM image for denoising.
Compared to linear filtering, adaptive filtering is more
selective and can preserve edges and high-frequency parts
better.[48] A Wiener filter is applied adaptively to the image
to achieve a better result.
2 | METHODS
Ten steps for image pre-processing before deep learning
practice is detailed in the Supplemental Information.
Briefly, the image is cropped into an image with dimen-
sion of 1024  680 pixels to remove the title bar from the
original image. The Wiener filter is used to minimize
the mean square error among the pixel intensities. By per-
forming histogram equalization, the intensity peak is spread
out, and objects in the image become more prominent.
Opening and closing are performed to further reduce
noises. Image binarization converts a grayscale image into a
binary image, greatly reducing the information contained
in the image from 256 shades of grey to two shades, black
and white. In this study, the Sauvola local threshold
method is used for consistency because our research group
has been using this approach. The Sauvola local threshold
allows retrieval of information from a grid of 4  4 pixels,
which is used to determine a localized value for the
binarization. With the binary image, all graphics on the
image with an area less than 3  3 pixels are eliminated
after the ‘majority’ and ‘clean’ operations. The Canny
edge detection is used to detect the fibre boundaries.
The dilation operation is used to obtain output, in which
the output pixel is the maximum value of all pixels in
the neighbourhood. Then, the thinning operation is
performed to reduce the thickness of the boundary. This
step helps fix the edge profile and avoids inaccurate
thickening. After the skeletonization with the identified
centrelines, the pruning operation is then performed to
delete sporadic branches at the ends of the skeletons.
Finally, the Euclidean distance transformation is chosen
for this research project.

4 YU ET AL.
The UNet model is used for the segmentation of the
SEM images, followed by central line determination and
distance transformation for diameter determination. The
network architecture of UNet is illustrated in Figure 1.[40]
The most notable feature of this network is its U-shape
structure. The left side of the U shape is a contracting
path consisting of a series of convolutional layers. The
right side is an expansive path consisting of a series of
up-convolutional layers. There are also four concatenations
between the contacting path and the expansive path.
The contracting path contains four convolutional
steps. Each step contains two convolution layers and a
max pooling layer. In each step, a 3  3 convolutions
kernel is first applied twice, each followed by a rectified
linear activation function (ReLU). This convolutional
operation is unpadded and results in the loss of boundary
pixels. After two convolutional layers, a 2  2 max pooling
operation with stride two is applied to the feature map.
The number of feature channels is doubled after each
down-sampling step.
As for the expansive path, each up-sampling step of
the feature map is followed by a 2  2 up-convolution
layer that halves the number of feature channels. ReLU
is used as activation function. A SoftMax layer is placed
after all neural network layers.
In the original UNet model,[40] border pixels may be
lost during each convolution. Therefore, it is necessary
to crop the image into the desired size before the concat-
enation operation. There are 23 layers in this network.
FIGURE 1 The structure of UNet.
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
The loss function used is the cross-entropy function.
Apart from the network structure, overlap-tile strategy[49]
is also used in this model to improve performance. When
a pixel is predicted in the image, the surrounding pixels
are also considered to provide environmental information
for enhanced accuracy. For this reason, padding is con-
ducted by preprocessing to expand the original image size.
In this paper, the implementation of the UNet model is
not exactly the same as the one originally reported.[40]
Unlike medical image segmentation, the images segmented
in this project are post-processed to determine the diameter
of the nanofibre. The result of diameter is highly related to
the number of pixels in the image. A smaller overall size
of the image results in a smaller diameter too, deviating
from the fact.
To minimize the deviation mentioned in the preced-
ing paragraph, padding is added to each convolutional
step in this work. Therefore, the pixels lost during the
convolution will be complemented by zeroes, and the size
of the image remains unchanged during the convolution
operation. The size of the output image then equals
that of the input image. Batch normalization[50] is
added between the convolution operation and the ReLU
activation function.[51]
We implement the UNet code on the SEM images of
nanofibres collected in our research group to segment the
nanofibres from the background. We change the padding
method of the original model without major change to
the original neural network structure.

YU ET AL.
2.1 | Procedure for implementation of
UNet model
The implementation of deep learning comprises of three
steps: (1) database build-up, (2) implementation of UNet
model, and (3) optimization of hyperparameters. They
are introduced one-by-one as follows.
2.1.1 | Dataset build-up
The nanofibre samples were prepared in our research lab
by other graduate students studying electrospun nanofi-
bres for energy and clean air applications. Data presented
in this thesis project include those from the same research
group. All electrospun nanofibre samples are polymer-
based, including polyacrylonitrile (PAN) and polyvinyl
pyrrolidone (PVP). A total of 100 SEM images are col-
lected. They are randomly split into training data (80%),
validation data (10%), and test data (10%) for model train-
ing, validation, and testing, respectively.
F I G U R E 2 Procedure for mask
correction using Photoshop.
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
The SEM images are prepared as follows. After
removing the title bar (i.e., the word description under
the main image), the size of SEM images becomes around
600  1000. The original SEM images come in different
sizes because they were taken at different times using
different equipment, but all images are cropped into
512  512 pixels. During cropping, the best effort has
been made to ensure as many non-repeating positions as
the image size allows. We wrote a Python script to
quickly batch process images.
Segmentation masks are generated from MATLAB
using the same binarization and edge detection algorithm
in the diameter determination part, specifically the
Sauvola local threshold and the Canny edge detection
method. In addition, the resultant segmentation mask is
manually corrected using Photoshop to further enhance
accuracy. The procedure is depicted in Figure 2.
Figure 2A is an example of SEM image after cropping,
and Figure 2B is the segmentation mask generated using
a MATLAB script. Figure 2C is a composite image made
in Photoshop. The opacity of the MATLAB segmentation

6 YU ET AL.
mask is adjusted to 50% and overlapped with the original
SEM image to show the deviation between them and the
imperfection of this segmentation result. Figure 2D is
the manually corrected segmentation mask with the changes
highlighted.
During the segmentation mask generation, the
masks generated by MATLAB may have a minor dis-
placement compared to the original image. Again, the
mask was generated by combining the Sauvola local
threshold method and the Canny edge detection. Dur-
ing the Sauvola local thresholding, the image informa-
tion was retrieved from the upper left corner of the
image grids by 4  4 pixels. The pixels at the very upper
left edge do not have enough neighbours for the algo-
rithm to calculate the threshold value, and they are
ignored during the thresholding process, resulting in a
shift of 4 pixels in both left and upward directions
within the segmentation mask. There is an additional
step in the image processing to line up the image and
the segmentation mask for further training.
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
As illustrated in Figure 3, the segmentation mask was
4 pixels downward and 4 pixels right compared to the
original image. Thus, a Python script is used to quickly
batch process the images by cropping the displacement
boundary quickly.
2.1.2 | Implementation of UNet model
The images with thick fibre boundaries are shown in
Figure 4. Figure 4A–C are images of fibre that are 1-pixel
wide, 3-pixel wide, and 5-pixel wide, respectively. However,
we still could not get a non-trivial result, even if the edge of
the fibre increased to five pixels thick.
Another attempt was made by inverting the image
colour (see Figure 5), changing the image from pure
black to pure white. However, the output was still
pure black images.
After excluding all the preceding possible sources of
problem, we believe that the problem lies in the format
F I G U R E 3 Displacement of
mask generated by MATLAB.

YU ET AL.
FIGURE 4 The segmentation mask image with different fibre thickness.
FIGURE 5 The segmentation mask image with inverted colour.
and properties of the images used as training dataset.
The original images obtained from the experimental
equipment are in tag image file format (.tif format). Thus,
they were converted into PNG format with 8-bits bit
depth using Photoshop. A bit depth of eight bits allows
256 different intensities (shades of grey) to be recorded.
After this attempt, we finally got non-trivial results from
the model.
2.1.3 | Optimization of the hyperparameter
The learning rate, batch size, and epoch were tuned and
optimized by trial and error. The loss function used for
training is binary cross-entropy. The range of learning
rate is from 1E 3 to 1E 5 as a best practice in the field.
The range of batch size varies from 2 to 10. The epoch
varies from 5 to 50 epochs. Figure 6 shows the effects
of epoch on the accuracy and loss of the testing set.
The greater the epoch number, the less the loss of the
model, which indicates a smaller variance between the
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
7
calculated and generated mask. However, the accuracy
did not increase further, but rather decreased when the
epoch number reached 25. This result shows that
25 epochs are already sufficient for the model training.
Otherwise, problems such as overfitting may occur when
the epoch number increases further.
2.1.4 | Diameter determination
Figure 7 shows the generated boundary result of the
nanofibre SEM images. The Python libraries used for
the image processing are skimage and quanfima. The
overlapping picture (Figure 7C) shows a promising result
for the fibre segmentation. When used for diameter
determination, precision and integrity are the two most
essential criteria for image segmentation. Precision
means the generated image boundary should be located
precisely at the place where the fibre boundary appears.
The dislocation of fibre boundary will lead to deviation in
diameter. Integrity means that both sides of the fibre

8 YU ET AL.
FIGURE 6 The accuracy and loss under different epochs.
FIGURE 7 (A) Original SEM image; (B) The generated boundary of nanofibre; (C) Overlap of (A) and (B) for comparison.
need to appear in the image segmentation mask. The
fibres in the background, which do not appear on
the segmentation mask, do not introduce deviation to
the diameter result. Therefore, the segmentation result
does not need to outline every single fibre on the image.
However, the algorithm for diameter calculation may
misrecognize fibres with only one side boundary and
bring deviation into the result. Therefore, the segmenta-
tion is regarded as satisfactory only when both criteria
are satisfied.
One misjudgement problem is noticed during the
diameter determination as illustrated in Figure 8. As
seen from the skeleton of the image in Figure 8A, the
centrelines are drawn between not only the fibres, but
also some gaps, and the misjudged fibres are highlighted
in Figure 8B. This means that some gaps between fibres
are also recognized as the nanofibre boundary, and the
distance between gaps is calculated as the fibre diameter.
The algorithm itself causes this error and it is considered
challenging for the program to distinguish between the
gap and the fibre. However, this is not a big concern
because the gap usually leads to an extremely large
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
number, which is hundreds of times larger than the
average fibre diameter. This obvious outlier can be
easily identified and excluded.
3 | RESULTS A ND DISCUSSION
3.1 | Results of multiple image
processing using MATLAB
Figure 9 compares the fibre diameter distributions
obtained using the two-image processed by MATLAB. It
clearly indicates the dependence of diameter distribution
on the SEM magnification. Considering the dependence
of diameter distribution on the SEM magnification,
results from images at different magnifications should be
combined for more statistically reliable data. For the
combined result, the average diameter and standard
deviation are closer to the low-magnification image
result. This is likely because the low-magnification
images contain a larger area of sample, and therefore are
more representative. The mean diameter obtained from
 1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
YU ET AL. 9

F I G U R E 8 Skeletonization
result of the image.

F I G U R E 9 (A) Scanning
electron microscope (SEM) image of
nanofibre at high ( 20K)
magnification; (B) SEM image of
nanofibre at low ( 5K)
magnification; (C) Diameter
distributions from two images at
different magnifications.

the image with a higher magnification is smaller. It is
likely because the high magnification means a zoom-in
on the sample, and thus characterizes the finer fibres that
are ignored in the low-magnification scenario. Therefore,
the combination of the results from different magnifica-
tions is deemed more comprehensive and accurate in
diameter determination.
3.2 | Results obtained by deep
learning (UNet)
Figure 10 compares the diameter distributions obtained by
image-processing using MATLAB, manual measurement,
and deep learning (UNet). The manual measurement
is regarded as the benchmark for accuracy evaluation.

10
FIGURE 10 (A) The subject scanning electron microscope (SEM) image; (B) The results measured using MATLAB, UNet, and manual
counting.
F I G U R E 1 1 (A) The scanning electron microscope (SEM) image with the manually measured location marked; (B) Image with the
UNet method measuring area highlighted in yellow.
The manual measurement is from the doctoral thesis of
Givehchi,[33] who also uses the manual measurement as
a benchmark. The same benchmark is used herein to
compare the UNet method with the MATLAB method.
Figure 10B shows that that the UNet method is closer to
the manual measurement than MATLAB does. However,
their difference is not distinctive. Further in-depth
analyses are as follows.
The difference between manual measurement and
UNet result may be induced by the difference in the area
of examination. For example, Figure 11 shows the origi-
nal SEM image (Figure 11A) for manual measurement
and the image for the UNet method highlighted in yellow
(Figure 11B). Figure 11A shows the image with manually
measured points marked—the short white lines crossing
the fibres in this figure mark the random locations for
manual measurement. Note that all fibres in the original
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
YU ET AL.
SEM image are counted when using the MATLAB
method. The size of the complete SEM image is 1024  681
pixels, which can be measured twice in a lateral direction
using UNet because when using UNet for image segmenta-
tion, the input image size is fixed at 512  512 pixels. This
means that the UNet method cannot measure the entire
area of the SEM image.
To exclude the interference of different examination
areas, we compared the results using the same area,
which are the squares marked by yellow lines. Figure 12
compares the results obtained using the same size
of SEM image, and the corresponding mean diameters
and standard deviations from all methods are shown
in Table 1. According to the table, compared to the
MATLAB method, mean diameter from UNet method has
a smaller deviation with the manual result. This means
the UNet method has a better accuracy compared to the
 1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
YU ET AL. 11

FIGURE 12 The comparison of diameter distribution curve from different methods.

TABLE 1 The mean and standard deviation of diameter results.

Manual By MATLAB By deep learning (UNet)

Sample
Mean’s relative Mean’s relative
deviation with deviation with
Mean (nm) Stdeva (nm) Mean (nm) Stdev (nm) manual (%) Mean (nm) Stdev (nm) manual (%)
#1 178.61 43.98 166.75 80.19 6.64 177.80 71.96 0.50
#2 177.07 54.50 156.02 72.72 11.89 156.77 58.02 11.46

’Stdev’ stands for standard deviation.

a

MATLAB method. Overall, the UNet model is deemed
reliable. The distributions of UNet data points are very
close to the manual measured data points and the overall
trend is also the same.
4 | C ON C L U S I ON S
This paper presents two approaches to fibre diameter
measurement: image processing by MATLAB and UNet-
based deep learning. The influence of image magnification
in the MATLAB approach was verified and considered
during the diameter determination procedure. Two
images at different magnifications can be processed at the
same time and their measurements are combined as the
final result, which is deemed more accurate. Deep
learning by UNet is used for diameter determination, and
the results obtained by deep learning were more accurate
than those obtained by image-processing using MATLAB.
Deep learning (UNet) was also implemented using Python,
which can be run on a free platform instead of commercial
software like MATLAB. This change reduces the cost of
the operation.
AUTHOR CONTRIBUTIONS
Hesheng Yu: Conceptualization; formal analysis; inves-
tigation; supervision; writing – review and editing.
Erqian Gao: Formal analysis; investigation; methodology;

12
visualization; writing – original draft. Zhongchao Tan:
Funding acquisition; project administration; resources;
supervision; validation; writing – original draft; writing –
review and editing.
FUNDING INFORMATION
Funding was received from the University of Waterloo
and GCI Ventures, Toronto, ON.
P EE R R EV IE W
The peer review history for this article is available at
https://www.webofscience.com/api/gateway/wos/peer-
review/10.1002/cjce.25130.
DATA AVAILABILITY STATEMENT
The data that support the findings of this study are
available on request from the corresponding author. The
data are not publicly available due to privacy or ethical
restrictions.
R EF E RE N C E S
[1] F. Pierini, M. Lanzi, P. Nakielski, S. Pawlowska, O. Urbanek,
K. Zembrzycki, T. A. Kowalewski, Macromolecules 2017, 50,
4972.
[2] K. M. Skupov, I. I. Ponomarev, D. Y. Razorenov,
V. G. Zhigalina, O. M. Zhigalina, I. I. Ponomarev,
Y. A. Volkova, Y. M. Volfkovich, V. E. Sosenkin, Macro-
mol. Symp. 2017, 375, 1600188.
[3] D. Sebastian, J. C. Calder on, J. A. Gonzalez-Exp osito,
E. Pastor, M. V. Martínez-Huerta, I. Suelves, R. Moliner,
M. J. Lazaro, Int. J. Hydrogen Energy 2010, 35, 9934.
[4] Y. Chen, X. Li, K. Park, J. Song, J. Hong, L. Zhou, Y. W. Mai,
H. Huang, J. B. Goodenough, J. Am. Chem. Soc. 2013, 135,
16280.
[5] Y. E. Miao, G. N. Zhu, H. Hou, Y. Y. Xia, T. Liu, J. Power
Sources 2013, 226, 82.
[6] T. Yu, B. Lin, Q. Li, X. Wang, W. Qu, S. Zhang, C. Deng, Phys.
Chem. Chem. Phys. 2016, 18, 26933.
[7] N. Iqbal, X. Wang, A. A. Babar, G. Zainab, J. Yu, B. Ding, Sci.
Rep. 2017, 7, 15153.
[8] K. Yoon, B. S. Hsiao, B. Chu, J. Mater. Chem. 2008, 18, 5326.
[9] R. Givehchi, Q. Li, Z. Tan, Fuel 2016, 181, 1273.
[10] Z. K. Nagy, A. Balogh, B. Vajna, A. Farkas, G. Patyi,
Á. Kramarics, G. Marosi, J. Pharm. Sci. 2012, 101, 322.
[11] I. Sebe, P. Szabo, B. Kallai-Szabo, R. Zelk
o, Int. J. Pharm 2015,
494, 516.
[12] K. Kataria, A. Sharma, T. Garg, A. K. Goyal, G. Rath, Drug
Delivery Lett. 2014, 4, 79.
[13] W. Xu, W. Yang, Y. Yang, Biotechnol. Prog. 2009, 25, 1788.
[14] J. S. Choi, K. W. Leong, H. S. Yoo, Biomaterials 2008, 29, 587.
[15] Y. Zhang, Y. Li, Z. Tan, Analyst 2021, 146, 7048.
[16] R. J. Moon, A. Martini, J. Nairn, J. Simonsen, J. Youngblood,
Chem. Soc. Rev. 2011, 40, 3941.
[17] A. Barhoum, M. Bechelany, A. Salam, H. Makhlouf, Handbook
of Nanofibres, Springer International Publishing, Cham,
Switzerland 2019.
[18] G. Eda, S. Shivkumar, J. Appl. Polym. Sci. 2007, 106, 475.
[19] A. Koski, K. Yim, S. Shivkumar, Mater. Lett. 2004, 58, 493.
1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
YU ET AL.
[20] M. G. Mckee, J. M. Layman, M. P. Cashion, T. E. Long, Science
2006, 311, 353.
[21] Y. Y. Zhao, Q. B. Yang, X. F. Lu, C. Wang, Y. Wei, J. Polym.
Sci., Part B: Polym. Phys. 2005, 43, 2190.
[22] N. Amiraliyan, M. Nouri, M. H. Kish, J. Appl. Polym. Sci. 2009,
113, 226.
[23] B. Veleirinho, M. F. Rei, J. A. Lopes-Da-Silva, J. Polym. Sci.,
Part B: Polym. Phys. 2008, 46, 460.
[24] M. Dhanalakshmi, A. K. Lele, J. P. Jog, Mater. Today
Commun. 2015, 3, 141.
[25] X. Y. Yuan, Y. Y. Zhang, C. Dong, J. Sheng, Polym. Int. 2004,
53, 1704.
[26] R. Khajavi, M. Abbasipour, Electrospun Nanofibres, Elsevier,
Duxford, UK 2017, p. 109.
[27] E. Tomba, P. Facco, M. Roso, M. Modesti, F. Bezzo, M. Barolo,
Ind. Eng. Chem. Res. 2010, 49, 2957.
[28] E. Ho Shin, K. Soo Cho, M. Hwo Seo, H. Kim, Macromol. Res.
2008, 16, 314.
[29] R. Murphy, A. Turcott, L. Banuelos, E. Dowey, B. Goodwin,
K. O. H. Cardinal, Tissue Eng., Part C 2020, 26, 628.
[30] N. Otsu, IEEE Transactions on Systems, Man, and Cybernetics
1979, 9, 62.
[31] N. Dehghan, P. Payvandy, M. A. Tavanaie, Journal of Textiles
and Polymers 2016, 4, 83.
[32] J. Ahne, Q. Li, E. Croiset, Z. Tan, Text. Res. J. 2018, 89,
3137.
[33] R. Givehchi, PhD Thesis, University of Waterloo (Waterloo, ON)
2016.
[34] R. Givehchi, Q. Li, Z. Tan, Materials 2018, 11, 1277.
[35] R. Givehchi, B. Du, M. Razavi, Z. Tan, J. Siegel, Aerosol Sci.
Technol. 2021, 55, 805.
[36] Y. Li, Q. Li, Z. Tan, J. Power Sources 2019, 443, 227262.
[37] Y. Li, Z. Tan, Energy Fuels 2020, 34, 14915.
[38] T. Ferreira, W. S. Rasband, ImageJ User Guide, ImageJ/
Fiji1.46. 2012, https://imagej.nih.gov/ij/docs/guide/user-guide.
pdf (accessed: March 2022).
[39] N. A. Hotaling, K. Bharti, H. Kriel, C. G. Simon, Biomaterials
2015, 61, 327.
[40] O. Ronneberger, P. Fischer, T. Brox, ArXiv preprint 2015,
1505.04597, http://arxiv.org/abs/1505.04597 (accessed: May
2015).
[41] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy,
A. L. Yuille, ArXiv preprint 2014, 1412.7062, http://arxiv.org/
abs/1412.7062 (accessed: July 2022).
[42] L. C. Chen, G. Papandreou, I. Kokkinos, K. Murphy,
A. L. Yuille, IEEE Transactions on Pattern Analysis and
Machine Intelligence 2018, 40, 834.
[43] J. Long, E. Shelhamer, T. Darrell, ArXiv preprint 2014, 1411.4038,
http://arxiv.org/abs/1411.4038 (accessed: July 2022).
[44] K. He, G. Gkioxari, P. Dollar, R. Girshick, Proceedings of the
IEEE International Conference on Computer Vision (ICCV),
IEEE, Venice, Italy 2017, p. 2961.
[45] D. Bolya, C. Zhou, F. Xiao, Y. J. Lee, ArXiv preprint 2019,
1904.02689, http://arxiv.org/abs/1904.02689 (accessed: July
2022).
[46] T. Takikawa, D. Acuna, V. Jampani, S. Fidler, ArXiv preprint
2019, 1907.05740, http://arxiv.org/abs/1907.05740 (accessed:
July 2022).
[47] A. Kirillov, K. He, R. Girshick, C. Rother, P. Dollar, ArXiv
preprint 2018, 1801.00868, http://arxiv.org/abs/1801.00868
(accessed: July 2022).
 1939019x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cjce.25130 by CAPES, Wiley Online Library on [01/12/2023]. 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
YU ET AL. 13

[48] J. Chen, J. Benesty, Y. Huang, S. Doclo, IEEE Transac- SU PP O R TI N G I N F O RMA TI O N

tions on Audio, Speech, and Language Processing 2006, 14, Additional supporting information can be found online
1218. in the Supporting Information section at the end of this
[49] S.-H. Tsang, Review: U-Net (Biomedical Image Segmentation).
article.
Towards Data Science, https://towardsdatascience.com/review-
u-net-biomedical-image-segmentation-d02bf06ca760 (accessed:
March 2022). How to cite this article: H. Yu, E. Gao, Z. Tan,
[50] S. Ioffe, C. Szegedy, Proceedings of Machine Learning Research
Can. J. Chem. Eng. 2023, 1. https://doi.org/10.1002/
2015, 37, 448.
cjce.25130
[51] A. F. Agarap, ArXiv preprint 2019, 1412.7062, https://arxiv.
org/abs/1803.08375 (accessed: July 2022).