Access to this full-text is provided by Springer Nature.
Content available from Scientific Data
This content is subject to copyright. Terms and conditions apply.
1
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
A large-scale optical microscopy
image dataset of potato tuber
for deep learning based plant cell
assessment
Sumona Biswas & Shovan Barma
We present a new large-scale three-fold annotated microscopy image dataset, aiming to advance the
Background & Summary
Microscopy image analysis by employing machine learning (ML) techniques advances the critical understanding
of several characteristics of biological cells, ranging from the visualization of biological structures to quantica-
tion of phenotypes. In recent years, deep learning (DL) has revolutionized the area of ML, especially, in computer
vision technology by evidencing vast technological breakthroughs in several domains of image recognition tasks
including object detection, medical and bio-image analysis, and so on. In general, the ML implicates complex sta-
tistical techniques on a set of images and its recognition eciency heavily relies on the handcraed data features;
whereas, the DL processes the raw image data directly and crams the data representation automatically. Indeed,
its performance highly depends on the large number of diverse images with accurate and applicable labelling.
Following the trend, the DL is emerging as a powerful tool for microscopy image analysis, such as cell segmen-
tation, classication, and detection by exploring the dynamic variety of cells. Moreover, the DL pipeline allows
discovering the hidden cell structures, such as single-cell size, number of cells in a given area, cell wall thickness,
intercellular space distribution, subcellular components, and its density, etc. from microscopy images by extract-
ing the complex data representation in hierarchical way. Meanwhile, it expressively diminishes the burden of
feature engineering in traditional ML.
Certainly, several works have been attempted in cell biology and digital pathology domain to provide quanti-
tative support in automatic diagnosis and prognosis by detecting mitosis, nucleus, cells, and the number of cells
from breast cancer1,2, brain tumour3, and retinal pigment epithelial cell4 images in DL framework. Consequently,
the DL network successfully applied in plant biology for stomata classication, detection5–7, and counting8, plant
protein subcellular localization9, xylem vessels segmentations10, and plant cell segmentation for automated cells
tracking11,12. erefore, it is necessary to assemble a large number of annotated microscopy images and its ground
truth for the successful application of DL based microscopy image analysis13. Certainly, there are numerous pub-
licly available microscopy image datasets, mostly medical images for DL based diagnosis and prognosis, such as
Human Protein Atlas14, H&E-stained tissue slides from the Cancer Genome Atlas15, DeepCell Dataset16,17, Mitosis
Department of Electronics and Communication Engineering, Indian Institute of Information Technology Guwahati,
Guwahati, Assam, India. e-mail: sumona@iiitg.ac.in; shovan@iiitg.ac.in
Content courtesy of Springer Nature, terms of use apply. Rights reserved
2
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
detection in breast cancer18, lung cancer cell19. In contrast, there are very few number of publicly accessible bio-
logical microscopy image datasets of plant tissue cells, which are suitable for the DL framework. Furthermore, the
existing datasets have limited number of diverse images with proper annotation. In such context, we have gener-
ated an optical microscopy image dataset of potato tuber with a larger number of diverse images, and appropriate
annotation. is publicly available dataset will be benecial in analysing the plant tissue cells with great details by
employing DL based techniques.
Microscopy image analysis has become more reliable in understanding the structure, texture, geometrical
properties of plant cells and tissues which pay a profound impact on botanical research. Such studies have sig-
nicant aspects in interpreting the variety of dierent plant cells, tissues, and organs by discriminating cell size,
shape, orientation, cell wall thickness, distribution, and size of intracellular spaces20, tissue types, and mechani-
cal21,22 properties like shear, compressive stiness etc. For instance, the shape and size of cell guides to determine
the size and texture23 of a plant organ; while, the tissue digestibility and plant productivity24,25 are controlled by
the cell wall thickness; similarly, the mechanical properties of the cell wall plays a crucial role in plant stability
and resistance against pathogens26; whereas, the intercellular spaces inuence the physical properties of tissues,
like rmness, crispness, and mealiness27. Certainly, it has been practiced in various domains of plant cell research,
such as fruits and vegetables23,28. In this connection, there are various ways to generate microscopy images, such
as brighteld microscopy, uorescence microscopy, and electron microscopy. All these methods have their own
advantages and disadvantages as well. Besides, sample preparation is one of the crucial steps in microscopy image
generation which includes xation, paran embedding, and dierent staining techniques for better visualization
of cell segments. e most widely used stains are safranin-o29,30 and toluidine blue-o31 for visualizing cell walls
and lugol’s iodine32 for starch detection.
In this view, we present a large brighteld optical microscopy image dataset of plant tissues of potato tuber, as
it is one of the principal and high productive tuber crops and a valuable component of our regular diet. Usually,
potato tubers are of oval or round shape with white esh and pale brown skin with bud and stem end. ree major
parts of the tuber are cortex, perimedullary zone (outer core), and pith (inner core) with medullary rays, which
are made up of parenchyma cells. e cell structures are distinct for dierent tuber variety33, even within the same
tuber, especially inner core and outer core34. e same structural dierences can also be observed between the
stem and bud ends. In addition, the cell division and enlargement in various regions play an important role35 on
potato tuber growth. Following such variations in cell structure, we have generated a large dataset consisting of
15,938 fully annotated unstained and stained images with three-fold labelling. e labelling has been prepared
based on the tuber size (large, medium, and small), collections area (bud, middle and stem part), and tissue zones
(inner and outer core) and the images have been graded as physical, morphological and tissue grading respec-
tively. In addition, 60 ground truth segmentation labels of the images from the inner core have been prepared for
the dierent tuber weight. To check the quality of the images, technical validation has been conducted by the DL
based classication and segmentation tasks, which displayed signicant recognition accuracy. us, this dataset
is very much suitable for studying plant cell microstructures including cell size and shape, cell wall thickness,
intercellular space, starch, and cell density distribution in potato tubers using DL based pipeline. Indeed, such
properties can be explored explicitly as the dataset includes the images from the entire region of the tuber cov-
ering two tissue zones from stem to bud end for dierent tuber weights. In addition, large number of images in
this dataset will provide new opportunities for evaluating and developing DL based plant biology classication
and segmentation algorithms. Furthermore, the unstained along with stained images will be suitable to develop
virtual-staining algorithms in the DL framework. erefore, the dataset could enrich the DL based microscopy
cell assessment in plant biology substantially.
Methods
e raw potato tubers (Solanum
tuberosum L.) of an Indian variety, Kufri Pukhraj have been chosen in this work. e Kur Pukhraj, an excellent
source of vitamin C, potassium, and bre is one of the popularly grown commercial cultivars in India. e tubers
have been collected immediately aer harvesting in mid of December 2019 from Kamrup, a district of Assam
state, India. All the samples without any outer damages have been collected and stored in the temperature of
19.2 °C–29.2 °C with 70% relative air humidity. Based on the weight of the tuber, samples have been graded into
large, medium, and small of weight 80–100 gm, 40–50 gm, and 15–25 gm respectively. From each of these groups,
5 samples (total of 15 samples) have been selected for image generation at the laboratory maintaining stable
room temperature and humidity. e whole experiment including collection of the tuber samples and image
generation has been accomplished in 20 days. Dierent graded tuber samples are chosen alternate days during
the experiment.
e major parts of potato tuber, periderm (skin) with the bud and stem ends, cortex, perimedullary zone
(outer core), and pith (inner core) with medullary rays have been displayed in Fig.1b,c. e periderm, the outer-
most layer, protects a tuber from dehydration, infection, and wounding during harvest and storage. e cortex,
outer core, and inner core tissues appear successively aer the skin where starch granules are stored in paren-
chyma cells. e thickness of the cortex is about 146–189 µm36 and the largest cells are found here. e outer
core spreads about 75% of the total tuber volume and contains the maximum amount of starch37. e innermost
region i.e., inner core expands from stem to bud end38 along longitude direction; whereas, the medullary rays
spread toward the cortex. e samples have been collected from the inner and outer core which covers most of
the areas of a tuber. Besides, the cell structures34 and the amount of starch are distinct in these two tissue zones.
Similar samples have been collected from three areas, named Z1, Z2, and Z3 as indicated in Fig.1b. e samples
have been extracted with a cork borer of a diameter of 4 mm and rinsed in distilled water. Aer that, 5 thin sec-
tions from the inner core as well as the outer core of each of the three areas have been collected. erefore, from
a tuber sample, 30 thin sections (5 sections ×3 section areas ×2 tissue zones) have been analysed. Furthermore,
Content courtesy of Springer Nature, terms of use apply. Rights reserved
3
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
to capture images, fresh thin potato sections (i.e., unstained samples) have been placed under the microscope.
In addition, for the better visualization of cell boundaries and subcellular components, especially starch, the
samples have been stained. Safranin-o (1% solution) and toluidine blue-o (0.05% solution) has been used to vis-
ualize all cell walls; whereas, lugol’s iodine solution helped to distinguish starch granules. An optical microscope
(Labomed Lx300, Labomed America) accompanied by a smartphone camera (Redmi Note 7 Pro) was used to
generate and capture microscopy images as shown in Fig.1f. e brighteld microscopy images have been gen-
erated using a 10x lens (eld number = 18, numerical aperture = 1.25) which provides a eld of view (FOV) of
diameter 0.18 mm. e camera of the smartphone has been xed on the microscope eyepiece by using an adaptor.
Certainly, the exposure and white balance state has been secured by the adequate brightness level of the micro-
scope’s built-in light-emitting diode (LED) and a clear FOV. e exposure time of the smartphone camera has
been kept in the range of 1/200 s– 1/100 s which provides satisfactory brightness level; whereas, the focus setting
of the camera has been locked that maintains xed magnication among all the images. e images have been
captured in the highest quality JPG format with maximum of 10% compression only to retain the image quality
reasonably high. e mobile camera has been xed to 3x zoom which oers a FOV of 890 × 740 µm2 with an
approximate resolution of 0.26 µm/pixel. Following this setting, three images have been taken for each eld of
view by changing the focus distance of 3 µm. Similarly, around 15 images have been acquired from a section by
continuous precision shiing of the microscope stage along the x-y plane before the samples get dried. us, in
total 9,811 unstained and 6,127 stained images have been captured and saved in JPG format in 24-bit RGB color
and of resolution 3650 × 3000.
Previous studies identied that the potato tuber weight is directly associated with the num-
ber of cells and cell volume in dierent tissue zones. Nevertheless, the cell numbers are considered as a signicant
factor compared to mean cell volume for a tuber weight variation39. Hence, potato tuber weight has been recog-
nized as one of the important physical parameters to achieve versatility in the image database. erefore, in this
work, based on the weight, potato tubers are categorized into three groups as large, medium, and small. Certainly,
the captured microscopic images are composed of discrete cells with thin nonlignied cell walls surrounded by
starch granules40. In a tuber, the cell size diers considerably in the two tissue zones— inner and outer core34. In
general, the outer core occupies the maximum volume of the tuber and stores the largest number of starch gran-
ules as reserve material. On the contrary, the inner core cells are smaller34 with lower starch content which makes
this tissue zone wet and translucent as displayed in Fig.2. Such variation of cell sizes and starch distribution can
be observed in the stem, bud, and middle section of tubers as well. erefore, the images have been graded into
three categories namely (1) physical grading, (2) morphological grading and, (3) tissue grading based on tuber
weight, section areas, and tissue zones respectively.
Physical grading. Tubers of three dierent weight ranges have been selected for the image dataset, as it has a
correlation with the cell features. ree dierent weight groups of tubers, such as large (L), medium (M), and
small (S) with weight 80–100 gm, 40–50 gm and 15–25 gm respectively, have been considered for this microscopy
image dataset. e generated images have been labelled with L, M, and S followed by sample number 1–5 to dis-
tinguish tuber weight along with sample number; for instance, L1 refers to the rst sample of a large tuber. e
labels associate with weights and related parameters along with sample numbers for physical grading have been
listed in Table1.
Fig. 1 Demonstration of potato tuber anatomy, sample preparations, and image acquisition set up for
microstructure visualization: (a) A potato tuber sample. (b) Longitudinal cross-section of a tuber. e samples
have been divided into three parts, named Z1, Z2, and Z3 nearer to bud, middle and stem respectively as
indicated by dotted lines for microscopic observations. (c) Transverse cross-sections of the tuber where sample
collection areas, inner and outer core are highlighted by red circle. (d) Tissue samples have been collected
by using a cork borer of diameter 4 mm from specied zones. (e) in free-hand unstained sections have
been obtained. e stained samples have been prepared by using safranin-o (1%), toluidine blue-o (0.05%),
and lugol’s iodine. (f) Image capturing set up in which, the camera of the smartphone has been xed on the
microscope eyepiece by using an adaptor. Two types of microscopy images, unstained and stained images have
been captured independently without drying the sections.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
4
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
Morphological grading. e bud and stem ends of potato tubers are connected with the apical and basal end of
the shoot respectively. ese areas displayed compositional variations41,42 with distinct cell features. e images
of the tuber middle part (separates the bud, and stem end) have been incorporated in this dataset to visualize
structural variations along the longitudinal direction. erefore, for morphological grading, the tubers have been
divided into three parts namely Z1, Z2, and Z3 which specify the bud, middle, and stem areas respectively as
shown in Fig.1b. Certainly, the images have been captured from these areas for each physically graded sample
and labelled accordingly.
Tissue grading. A signicant variation in cell sizes within the same potato tuber can be observed in inner core
and outer core tissue zones. e cell size of the outer core is larger than that of the inner core and contains most of
the starch material. erefore, in tissue grading, these two zones have been identied. Certainly, the images have
been captured from these zones for each morphologically graded sample and labelled as IC and OC which indi-
cates the inner and outer core of the potato tuber respectively. Example of unstained and stained images of large,
medium, and small potato tuber from dierent section areas and tissue zones have been displayed in Figs.3–5
respectively.
Database Summary
ere are total 15,938 (9,811 unstained and 6,127 stained) numbers of images in this dataset. e images are
categorized based on dierent grading and labelling basis, and listed in Table2. e rst two columns refer to
grading and labelling basis followed by the number of images for unstained and stained cases. Furthermore, the
stained images with three stains (safranin-o, toluidine blue-o, and lugol’s iodine) are also specically mentioned.
Segmentation is performed
to split an image into several parts to identify meaningful features or objects. In microscopy image analysis,
a common problem is to identify distinct parts which correspond to a single cell or cell components to quan-
tify the spatial and temporal coordination. Furthermore, as a precursor to geometric analysis, such as cell size
and its distributions, image segmentation is essential. Such a task can be performed manually, which is very
much time-consuming, irreproducible, and tedious for larger image sets. Nonetheless, it can be automated by
the ML techniques which require proper ground truth labels. erefore, we have generated ground truth labels
of cell boundaries for the automated segmentation task. e images have been captured from dierent parts of
the tubers as mentioned earlier, and labelled accordingly. Certainly, to generate the ground truth labels for cell
boundary segmentation, the unstained images of inner core from the Z2 area have been selected, as cell bounda-
ries are comparatively prominent in this zone due to presence of fewer amounts of starch granules.
Segmentation of potato cell images can be very much challenging because of its complex cell boundaries and
non-uniformity in image background which leads to poor contrast between cell boundary and background.
erefore, to generate the ground truth cell boundaries, a few steps have been involved: (1) pre-processing (2)
thresholding, and (3) morphological operations. e pre-processing steps have been mainly implicated in back-
ground correction and image ltering. Generally, the uneven thickness of the tuber section results non-uniform
microscopy image background. us, to minimize such non-uniformity a well-known rolling ball algorithm43,44
has been employed. It eliminates the unnecessary background information by converting a 2D image I(x,y) into
a 3D surface; where, the pixel values are considered as the height. en, a ball of a certain radius (R) is rolled over
the backside of the surface which creates a new surface S(x,y). Furthermore, a new image with a uniform
Fig. 2 A schematic diagram displaying inner and outer core cell characteristics of a potato tuber.
SI. No. Weight
Categories Label Weight
(gm) Length
(mm) Width
(mm) ickness
(mm) Sa mp le’s No.
1Large L 80–100 65–80 50–55 40–45
1–52 Medium M 40–50 50–65 40–45 30–40
3Small S 15–25 38–50 25–35 25–30
Tab le 1. Summary of dierent parameters for physical grading including weight and associated measures of
potato tubers.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
5
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
background is created by,
=+−−IxyIxy Sx yR(, )((, )1)((, ))
new
44. To achieve an optimal image with the best
uniform background, the values of R must be selected carefully. In our work, empirically, the values of R have
been kept as 30 < R < 60. Next, for image ltering, bandpass lter has been used to enhance the cell edges by
eliminating shading eects. In this purpose, Gaussian ltering in Fourier space has been considered. A bandpass
lter having two cut-o frequencies, lower (fcl) and higher (fch) are kept within a range for intensity variation in
the captured image. Empirically, it has been kept as 10 < fcl < 30 and 60 < fch < 120. Furthermore, the adaptive
thresholding method45 has been implemented to binarize the images for discriminating the cell boundaries.
Moreover, morphological operators, such as opening, closing, and hole lling has been chosen to rene cell
boundaries. Several values of fcl, fch, and R have been chosen to get the best binary images. Although, very few
starch granules and some disconnected cell boundaries can be observed in the resultant binary images, which
could lead to a weak cell boundary segmentation. Certainly, such discrepancies have been further rened by very
well-known manual process46 which involves removal of the starch granules and contouring cell boundaries. e
whole process of cell boundary segmentation ground truth label generation has been shown in Fig.6.
is dataset is publicly available on gshare47 (https://doi.org/10.6084/m9.gshare.c.4955669) which can be
downloaded as a zip le. e zip le contains three folders named as “stain”, “unstain”, and “segmentation”. All
the images are in JPG format. e raw microscopy images of potato tubers can be found in “stain” and “unstain”
folder; whereas, the segmentation folder provides raw images with ground truth segmentation labels. e “stain”
Fig. 3 Example of unstained and stained images of large (80–100 gm) potato tubers. Rows and columns indicate
respective tissue zones (inner and outer core) and dierent staining agents. e rst column species the
unstained images, whereas, the subsequent columns are for stained images of safranin-o, toluidine blue-o, and
lugol’s iodine. e images are from (a) Bud Region (Z1), (b) Middle Region (Z2), and (c) Stem Region (Z3).
Note: All the images are with the scale at top-le corner on unstained image.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
6
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
folder contains dierent stained images in the respective folders named as “safranin”, “toluidine”, and “lugols”. e
image labels information can be extracted from the image lenames itself. e le naming format for unstained
image is as < physical grading with sample no. > _ < morphological grading > _ < tissue grading > _ < section
no. > _ < image no. > ; for example, M1_Z1_IC_Sec. 1_02.jpg refers to an image (rst section out of ve) taken
from the inner core of Z1 of medium weight potato tuber (sample no. 1). Similarly, the stained image le naming
format is as < physical grading with sample no. > _ < morphological grading > _ < tissue grading > _ < section
no. > _ < stain type > _ < image no.>; for example, S1_Z2_OC_Sec. 3_lugol_02.jpg refers to an lugol’s iodine
stained image (third section out of ve) taken from the outer core of Z2 of small weight potato tuber (sample
no. 1). e whole le naming format can be understood by following Table3. e segmentation folder con-
tains two subfolders, named as “images” and “groundtruth”. e image les naming format is as < physical grad-
ing > _ < image no. >; for example, “L_2.jpg” represents the second image of a large potato tuber sample. Besides,
the ground truth label images are kept in binary image format having the same dimensions of the raw images.
e technical validation has been conducted by employing the DL based classication and segmentation tasks
on the acquired image dataset as illustrated in Fig.7. Multi-label cell classication has been conducted to verify
the quality of the assigned labels. It has been examined by considering two specic image labels— physical (L,
M, and S) and tissue grading (IC and OC). Besides, to verify the ground truth segmentation labels, semantic
Fig. 4 Example of unstained and stained images of medium (40–50 gm) potato tubers. Rows and columns
indicate respective tissue zones (inner and outer core) and dierent staining agents. e rst column species
the unstained images, whereas, the subsequent columns are for stained images of safranin-o, toluidine blue-o,
and lugol’s iodine. e images are from (a) Bud Region (Z1), (b) Middle Region (Z2), and (c) Stem Region (Z3).
Note: All the images are with the scale at top-le corner on unstained image.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
7
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
Fig. 5 Example of unstained and stained images of small (15–25 gm) potato tubers. Rows and columns indicate
respective tissue zones (inner and outer core) and dierent staining agents respectively. e rst column
species the unstained images, whereas, the subsequent columns are for stained images of safranin-o, toluidine
blue-o, and lugol’s iodine. e images are from (a) Bud Region (Z1), (b) Middle Region (Z2), and (c) Stem
Region (Z3). Note: All the images are with the scale at top-le corner on unstained image.
Grading Basis Labelling Basis
Number of Images
Unstained
Stained
Safranin-O Toluidine Blue-O Lugol’s Iodine
Physical
Large (L) 3,311 674 683 680
Medium (M) 3,200 671 663 653
Small (S) 3,300 702 705 696
Morphological
Bud (Z1) 3,271 708 690 655
Middle (Z2) 3,327 660 698 674
Stem (Z3) 3,213 679 663 700
Tissue Inner Core (IC) 5,220 1,079 1,042 993
Outer Core (OC) 4,591 968 1,009 1,036
Tab le 2. e number of images under each grading for unstained and stained image dataset.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
8
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
segmentation has been performed using the DL pipeline. e rst test can yield information about the possible
separation of labels and the later can access individual cells in dierent tuber weights.
Multi-label cell classification. The CNN classification network, VGG1648 has been employed for
multi-label cell classication using input images of 256 × 256 pixels with two labels, physical (L, M, and S)
and tissue grading (IC and OC). e rst 13 layers of the neural network have been pre-trained on ImageNet
[ILSVRC2012] dataset. On top of it, task-specic fully-connected layers have been attached and activated by the
sigmoid function. e complete network has been ne-tuned on our datasets. e network performance has been
evaluated based on the train-test scheme. erefore, the entire image dataset (unstained and stained) has been
partitioned randomly into two subsets, with 80% for training and 20% for test. e network has been trained
using SGD49 optimizer with a learning rate of 10−2, momentum 0.9, and the binary cross-entropy as loss function
for both the image dataset. With the iterative learning technique, performance metrics, such as accuracy and
F2-score (assessing the correctness of the image labels), have been obtained for test images. e results have been
listed for the test set in Table4. It shows that for the same number of epochs (30), the unstained image dataset
gives a better result than the stained image dataset.
In this task, Unet50, a very well recognized image segmentation neural network has been
employed. It has shown remarkable performance in biomedical image segmentation. e input images have been
generated by subdividing each ground truth labels and raw images into 20 sub-images, which further resized to
512 × 512 pixels before training. e network has been trained using Adam51 optimizer with learning rate of 10−1.
Two types of inputs, namely raw and normalized images have been given separately into the network. e entire
Physical
Grading Sample
No. Morphological
Grading Tissue
Grading Section
No. Stain Type Image
No. Remark
Large (L)
1 to 5
Bud (Z1) Inner Core
(IC)
Sec <1 to
5>
Safranin-o
(safo)
Toluidine
blue-o (tolu)
Lugol’s iodine
(lugol)
n
L1_Z1_OC_Sec. 1_1
L1_Z1_IC_Sec. 4_safo_1
Medium (M) Middle (Z2) M1_Z1_OC_Sec. 1_2
Outer Core
(OC)
M2_Z1_IC_Sec. 1_tolu_1
Small (S) Stem (Z3) S2_Z2_OC_Sec. 3_3
S1_Z1_IC_Sec. 3_lugol_3
Tab le 3. Raw image le name format.
Fig. 6 Steps involved in generating the ground truth segmentation labels for the inner core tissues. e
original images pre-processed by employing rolling ball algorithm and bandpass ltering. Next, the adaptive
thresholding has been employed to obtain binary images. Furthermore, morphological operations have been
performed to rene the cell boundaries and remove the starch granules. By changing fcl, fch and R at pre-
processing steps, possible binary images have been generated. en, the best image has been selected for manual
correction.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
9
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
image dataset has been partitioned randomly into two subsets, with 80% for training and 20% for test. en, per-
formance evaluation has been conducted by employing normal adaptive learning rate-based training. During the
training period, early stopping has been used to choose the model with the highest validation performance. e
mean intersection of union (IOU) has been chosen as a performance metric that measures how much predicted
boundary overlaps with the ground truth (real cell boundary) and the results have been displayed in Table5. For
the same deep neural network, normalize input images give better result than the raw images. A representative
Fig. 7 Overall technical verication of image and ground truth segmentation label. Two types of microscopy
images have been chosen independently. e image labels have been veried by VGG1648 deep neural network
employing transfer learning. e Unet50 architecture has been used to employ the semantic segmentation using
the generated ground truth segmentation labels and hence verify the same.
Model Accuracy F2-score
Unstained images as train and test 0.9427 0.9207
Stained images as train and test 0.9205 0.8918
Tab le 4. Performance assessment based on accuracy and F2-score of unstained and stained image dataset.
Model Input typ e Mean IOU
Cell Boundary
as train and test
Raw Images 0.6964
Normalized Images 0.7020
Tab le 5. Performance assessment based on mean IOU of raw and normalize image dataset.
Fig. 8 Example of dierent image sets used during cell segmentation by Unet50 for ground truth labels
validation. e rst and second rows indicate the train and test images respectively. (a) Unstained raw RGB
images, (b) Ground Truth images, (c) Segmentation result for raw RGB input images, and (d) Segmentation
result for normalized input images.
Content courtesy of Springer Nature, terms of use apply. Rights reserved
10
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
result of cell segmentation for raw and normalized input images has been displayed in Fig.8 in which (a), (b),
(c), and (d) refer to raw RGB image, ground truth, the result for raw and normalized input images respectively.
Received: 29 April 2020; Accepted: 16 September 2020;
Published: xx xx xxxx
1. Cireşan, D. C., Giusti, A. & Gambardella, L. M. & Schmidhuber. Mitosis detection in breast cancer histology images with deep
neural networs. In Proc. 16th Int. Conf. Med. Image Comput. Comput. -Assist. Intervent. 8150, 411–418 (2013).
2. Veta, M., Van Diest, P. J. & Pluim. Cutting out the middleman: Measuring nuclear area in histopathology slides without
segmentation. In Proc. 19th Int. Conf. Med. Image Comput. Comput. -Assist. Intervent. 632–639 (2016).
3. X ing, F., Xie, Y. & Yang, L. An automat ic learning-based framewor for robust nucleus segmentation. IEEE Trans Med. Imaging. 35,
550–566 (2015).
4. Xie, W., Noble, J. A. & Zisserman, A. Microscopy cell counting and detection with fully convolutional regression networs. In Proc.
1st Worshop Deep Learn. Med. Image Anal. (MICCAI). 1–8 (2015).
5. Bhugra, S. et al. Deep Convolutional Neural Networs based Framewor for Estimation of Stomata Density and Structure from
Microscopic Images. In Proc. Eur. Conf. Comput. Vis. (ECCV). (2018).
6. Aono, A. H. et al. A stomata classication and detection system in microscope images of maize cultivars. Preprint at https://www.
biorxiv.org/content/10.1101/538165v1 (2019).
7. Saponaro, P. et al. Deepxscope: Segmenting microscopy images with a deep neural networ. In Proc. IEEE Conf. Comput. Vis. Pattern
ecognit. Worshops. 91–98 (2017).
8. Fetter, . C., Eberhardt, S., Barclay, . S., Wing, S. & eller, S. . Stomata Counter: a neural networ for automatic stomata
identication and counting. New Phytol. 223, 1671–1681 (2019).
9. Shao, Y.-T., Liu, X.-X., Lu, Z. & Chou, .-C. pLoc_Deep-mPlant: Predict Subcellular Localization of Plant Proteins by Deep
Learning. Nat. Sci. 12, 237–247 (2020).
10. Garcia-Pedrero, A. et al. Xylem vessels segmentation through a deep learning approach: a rst loo. IEEE Int. Wor Conf. Bioinspir.
Intell. (IWOBI). 1–9 (2018).
11. Jiang, W., Wu, L., Liu, S. & Liu, M. CNN-based two-stage cell segmentation improves plant cell tracing. Pattern ecognit. Lett. 128,
311–317 (2019).
12. Liu, M., Wu, L., Qian, W. & Liu, Y. Cell tracing across noisy image sequences via faster -CNN and dynamic local graph matching.
IEEE Int. Conf. Bioinform. Biomed. (BIBM). 455–460 (2018).
13. Moen, E. et al. Deep learning for cellular image analysis. Nat. Methods. 16, 1–14 (2019).
14. ul, P. J. et al. A subcellular map of the human proteome. Sci. 356, e aal3321 (2017).
15. umar, N. et al. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Trans. Med.
Imaging. 36, 1550–1560 (2017).
16. Bannon, D. et al. DeepCell 2.0: Automated cloud deployment of deep learning models for large-scale cellular image analysis.
Preprint available at https://www.biorxiv.org/content/early/2018/12/22/505032 (2018).
17. Fal, T. et al. U-Net: deep learning for cell counting, detect ion and morphometry. Nat. Methods. 16, 67–70 (2019).
18. oux, L. et al. Mitosis detection in breast cancer histological images. An ICP 2012 Contest. J. Pathol. Informatic. 4 (2013).
19. Cano, A., Masegosa, A. & Moral, S. ELVIA biomedical data set repository. (2005).
20. Pierzynowsa-ornia, G., Zadernowsi, ., Fornal, J. & Nesterowicz, J. e microstructure of selected apple varieties. Electron. J.
Pol. Agric. Univ. 5 (2002).
21. Sadowsa, J., Fornal, J. & Zgórsa, . e distribution of mechanical resistance in potato tuber tissues. Postharvest Biol. Tech. 48,
70–76 (2008).
22. Haman, J. & onstaniewicz, . Destruction processes in the cellular medium of a plant-theoretical approach. Int. Agrophys. 14,
37–42 (2000).
23. McAtee, P. A., Hallett, I. C., Johnston, J. W. & Schaer, . J. A rapid method of fruit cell isolation for cell size and shape measurements.
Plant Methods. 5 (2009).
24. Moghaddam, P. . & Wilman, D. Cell wall thicness and cell dimensions in plant parts of eight forage species. J. Agric. Sci. 131,
59–67 (1998).
25. Ahmad, N., Amjed, M., ehman, A. & ehman, A. Cell walls digestion of ryegrass and Lucerne by cattle. Sarhad J. Agric. 23, 475
(2007).
26. Vogler, H., Feleis, D., Nelson, B. J. & Grossnilaus, U. Measuring the mechanical properties of plant cell walls. Plants. 4, 167–182
(2015).
27. Volz, ., Harer, F., Hallet, I. & Lang, A. Development of Texture in Apple Fruit– a Biophysical Perspective. XXVI Int. Hortic.
Congress: Deciduous Fruit and Nut Trees. 636, 473–479 (2004).
28. onstaniewicz, ., Pawla, . & Zdune, A. Inuence of structural parameters of p otato tuber cells on their mechanical properties.
Int. Agrophys. 15, 243–246 (2001).
29. van de Velde, F., Van iel, J. & Tromp, . H. Visualisation of starch granule morphologies using confocal scanning laser microscopy
(CSLM). J. Sci. Food Agric. 82, 1528–1536 (2002).
30. Dürrenberger, M. B., Handschin, S., Conde-Petit, B. & Escher, F. Visualization of food structure by confocal laser scanning
microscopy (CLSM). LWT-Food Sci. Tech. 34, 11–17 (2001).
31. Souup, A. Selected simple methods of plant cell wall histochemistry and staining for light microscopy. Methods Mol. Biol. 1080,
25–40 (2014).
32. Smith, A. M. & Zeeman, S. C. X. Quantication of starch in plant tissues. Nat. Protoc. 1, 1342–1345 (2006).
33. Bordoloi, A., aur, L. & Singh, J. Parenchyma cell microstructure and textural characteristics of raw and cooed potatoes. Food
Chem. 133, 1092–1100 (2012).
34. onstaniewicz, . et al. Cell structural parameters of potato tuber tissue. Int. Agrophys. 16, 119–128 (2002).
35. Xu, X. & Vreugdenhil, D. & Lammeren, A. A. v. Cell division and cell enlargement during potato tuber formation. J. Exp. Bot. 49,
573–582 (1998).
36. Troncoso, E., Zúñiga, ., amírez, C., Parada, J. & Germain, J. C. Microstructure of potato products: Eect on physico-chemical
properties and nutrient bioavailability. Glob. Sci. Boos. 3, 41–54 (2009).
37. Salunhe, D. . & adam, S. Handboo of vegetable science and technology: production, compostion, storage, and processing.
(CC press, 1998).
38. Böl, M., Seydewitz, ., Leichsenring, . & Sewerin, F. A phenomenological model for the inelastic stress-strain response of a potato
tuber. J. Mech. Phys. Solids. 103870 (2020).
39. Liu, J. & Xie, C. Correlation of cell division and cell expansion to potato microtuber growth in vitro. Plant Cell Tiss. Organ Cult. 67,
159–164 (2001).
40. amaswamy, U. ., abel, M. A., Schols, H. A. & Gruppen, H. Structural features and water holding capacities of pressed potato bre
polysaccharides. Carbohydr. Polym. 93, 589–596 (2013).
Content courtesy of Springer Nature, terms of use apply. Rights reserved
11
SCIENTIFIC DATA | (2020) 7:371 | https://doi.org/10.1038/s41597-020-00706-9
www.nature.com/scientificdata
www.nature.com/scientificdata/
41. Liu, B. et al. Dierences between the bud end and stem end of potatoes in dry matter content, starch granule size, and carbohydrate
metabolic gene expression at the growing and sprouting stages. J. Agric. Food Chem. 64, 1176–1184 (2016).
42. Sharma, V., aushi, S., Singh, B. & aigond, P. Variation in biochemical parameters in dierent parts of potato tubers for processing
purposes. J. Food Sci. Tech. 53, 2040–2046 (2016).
43. Sternberg, S. . Biomedical image processing. Comput. 22–34 (1983).
44. Shvedcheno, D. & Suvorova, E. New method of automated statistical analysis of polymer-stabilized metal nanoparticles in electron
microscopy images. Crystallogr. ep. 62, 802–808 (2017).
45. Chan, F. H., L am, F. . & Zhu, H. Adaptive thresholding by variational method. IEEE Trans. Image Process. 7, 468–473 (1998).
46. White, A. E., Diow, . B., Baugh, M., Jenins, A. & Frandsen, P. B. Generating segmentation mass of herbarium specimens and a
data set for training segmentation models using deep learning. Apl. Plant Sci. 8, e11352 (2020).
47. Biswas, S. & Barma, S. A large-scale optical microscopy image dataset of potato tuber for deep learning based plant cell assessment.
gshare https://doi.org/10.6084/m9.gshare.c.4955669 (2020).
48. Simonyan, . & Zisserman, A. Very deep convolutional networs for large-scale image recognition. Preprint at https://arxiv.org/
abs/1409.1556 (2014).
49. uder, S. An overview of gradient descent optimization algorithms. Preprint at https://arxiv.org/abs/1609.04747 (2016).
50. onneberger, O., Fischer, P. & Brox, T. U-Net: convolutional networs for biomedical image segment at ion. Preprint at https://arxiv.
org/abs/1505.04597 (2015).
51. ingma, D. P. & Ba, J. Adam: A method for stochastic optimizat ion. Preprint at https://arxiv.org/abs/1412.6980 (2014).
Acknowledgements
e work has been supported by the Department of Science and Technology, Govt. of India under IMPRINT-II
with the le number IMP/2018/000538.
Author contributions
Both the authors have equal contribution in writing the manuscripts, carried outing main research, and analysis
tasks of this work.
e authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to S.B. or S.B.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional aliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre-
ative Commons license, and indicate if changes were made. e images or other third party material in this
article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons license and your intended use is not per-
mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
e Creative Commons Public Domain Dedication waiver http://creativecommons.org/publicdomain/zero/1.0/
applies to the metadata les associated with this article.
© e Author(s) 2020
Content courtesy of Springer Nature, terms of use apply. Rights reserved
1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com
