ArticlePDF Available

Abstract

Nematodes are ubiquitous in soil representing different trophic levels and occupying a central position in the detritus food web. Nematodes have been widely used for biomonitoring of soil quality and health. However, their application in bio-indicator is limited due to the taxonomic identification is laborious, and largely relying on morphological characters which require substantial prior nematology experiences. In present study, we developed the web application NemaRec using the I-Nema dataset. NemaRec adopts the deep convolutional neural networks approach for nematode image identification, incorporate I-Nema dataset in a user friendly interface, and further extends it to the calculation of feeding types, c-p values, MI and PPI indices for environment evaluation. Within 19 studied genera, the model can properly identify up to 60% genera, 76% of c-p values and 76% feeding types in specimen-based dataset (images taken from microscopy), and 94%–97% in augmented dataset (image treated with random flip and Gaussian noise). The pipeline was further incorporated into a user-friendly web application NemaRec. NemaRec offers high-throughput online identification while simultaneously collect images uploaded by users for future modeling training. To our knowledge this is the first soil nematode image identification system using deep learning approach. NemaRec is available http://168.138.167.251:8080/).
1
NemaRec: a deep learning-based web application for nematode image
1
identification and ecological indices calculation
2
3
Xue Qing a #, Yihao Wang a, b #, Xuequan Lu b, Xuan Wang a, Hongmei Li a*, Xiaojun Xiec*
4
a Department of Plant Pathology, Nanjing Agricultural University, Nanjing, 210095, China
5
b Deakin University, Waurn Ponds, VIC 3216, Australia
6
c College of Artificial Intelligence, Nanjing Agricultural University, Nanjing, 210095, China
7
# These authors contribute equally to this work.
8
9
*Corresponding author, E-mail address: lihm@njau.edu.cn, xxj@njau.edu.cn
10
Dr. Hongmei Li, Dr. Xiaojun Xie
11
Telephone: +86 25 84396432
12
Fax: +86 25 84395240
13
Mailing address: Department of Plant Pathology, Nanjing Agricultural University, Tongwei
14
Road No. 6, Xuanwu District, Nanjing 210095, China
15
16
2
ABSTRACT
1
Nematodes are ubiquitous in soil representing different trophic levels and occupying a central
2
position in the detritus food web. Nematodes have been widely used for biomonitoring of soil
3
quality and health. However, their application in bio-indicator is limited due to the taxonomic
4
identification is laborious, and largely relying on morphological characters which require
5
substantial prior nematology experiences. In present study, we developed the web application
6
NemaRec using the I-Nema dataset. NemaRec adopts the deep convolutional neural networks
7
approach for nematode image identification, incorporate I-Nema dataset in a user friendly
8
interface, and further extends it to the calculation of feeding types, c-p values, MI and PPI
9
indices for environment evaluation. Within 19 studied genera, the model can properly identify
10
up to 60% genera, 76% of c-p values and 76% feeding types in specimen-based dataset
11
(images taken from microscopy), and 94%-97% in augmented dataset (image treated with
12
random flip and Gaussian noise). The pipeline was further incorporated into a user-friendly
13
web application NemaRec. NemaRec offers high-throughput online identification while
14
simultaneously collect images uploaded by users for future modeling training. To our
15
knowledge this is the first soil nematode image identification system using deep learning
16
approach. NemaRec is available http://8.140.109.225:8866/).
17
18
Keywords: Maturity index; Machine learning; Soil ecology; Biodiversity; Convolutional
19
neural networks
20
21
22
3
1. Introduction
1
Nematodes are an evolutionarily successful group of organisms occupying nearly all
2
habitats that provide available organic carbon sources. With different feeding behavior and
3
life strategies, nematodes are classified to at least five functional or trophic groups (Yeates et
4
al., 1993) and placed in a central position in the detritus food web (Moore and de Ruiter,
5
1991). Nematodes are considered as useful ecological indicators due to several unique
6
attributes (Freckman, 1988). Based on reproductive potential from explosive opportunists to
7
conservative survivalists, families of nematodes can be ordered on a colonizerpersister (c-p)
8
scale. The maturity index (MI) and plant-parasite index (PPI) are consequently calculated
9
based on the weighted mean c-p value of the individuals.
10
The MI and PPI indices have been widely used to evaluate soil health or condition
11
(Yeates et al., 1993; Neher, 2001; Ferris and Bongers, 2006; Wilson and Khakouli-Duarte,
12
2009), for example, a google scholar search of “nematode, environmental indicators” returned
13
41,600 items up to January of 2020. These indices are based on the taxonomic identification
14
which largely relies on morphological characters. This is practically laborious as ecological
15
studies are typically large in samples size and rich in nematode species. Moreover, the high
16
phenotypic plasticity (Coomans, 2002; Nadler, 2002), the vague diagnostic characters (Erwin,
17
1982; Derycke et al., 2008), and frequently encountered juveniles (Anderson, 2000) upgrade
18
the difficulties in identification and thus a well-established experience and expertise is hence
19
essential. Given the number of taxonomists are rapidly declining worldwide, the nematode
20
taxonomic identification often become the bottleneck for the application of these ecological
21
indicators. With the rapidly declining cost and improved availability of genetic sequencing
22
and related technologies, molecular barcoding holds great promise as a tool to simplify and
23
standardize species identification (Hebert et al., 2003). The molecular barcoding further
24
evolved as metabarcoding technique for batch evaluation of environmental samples, and this
25
4
method has been proofed as a powerful tool in resolving soil nematode community
1
(Porazinska et al., 2009; Porazinska et al.,2010). However, due to the PCR bias, lack of
2
effective primers, and the requirement in bioinformatics experiences, their application remain
3
an alternative choice for nematode community study.
4
Beside taxonomic identification, manual calculation of indices are usually performed in
5
Microsoft Excel or other spreadsheet software and may require manual adjustment of
6
formulae and data format during the process. Although a computational SAS code (Neher and
7
Campbell, 1996) and a more user-friendly automatic program were developed
8
(Sieriebriennikov, 2014), they all need manual taxonomic identification and preparation of
9
data sheet.
10
Image-based automatic identifications have shown promising results in various computer
11
vision problems including the nematode detection and counting (Silva et al., 2003; Holladay,
12
2016). Deep learning methods have been introduced to this task driven by the success of the
13
Convolutional Neural Networks (CNNs). Several efforts have been made using CNNs (Chou
14
et al., 2017; Javer et al., 2019; Liu et al., 2018; Chen et al., 2020; Wang et al., 2020). However,
15
these applications are predominantly designed for model organism Caenorhabditis elegans or
16
concentrated on image stacking, and limited in detection of behavior, age, development stages
17
and morphological adaptation to stress. Recently, CNNs model and nematode images dataset
18
I-Nema (Lu et al., 2021) was created. However, I-Nema focus on dataset, with all source as
19
command code, limit their application among biologists. In present study we aimed to
20
develop an automatic ecological indices calculation pipeline by using CNN-based deep
21
learning in image identification, and incorporated it in a user friendly interface. To our
22
knowledge this is the first soil nematode image identification system using deep learning
23
approach.
24
2. Material and methods
25
5
2.1. Sample collection and nematode identification
1
Soil samples were collected from various ecosystems in China. Nematodes were extracted
2
from soil samples using a Baermann tray, concentrated using a 500 mesh sieve (25 µm
3
opening). After removing water, nematodes were manually picked up, fixed with 4% formalin
4
and gradually transferred to anhydrous glycerin (Seinhorst, 1959). Slides were examined and
5
photographed using an Olympus BX51 DIC Microscope (Olympus Optical, Tokyo, Japan),
6
equipped with an Olympus C5060Wz camera. Specimens were identified to genus level based
7
on morphology using available keys (e.g. Andrassy, 2005; Andrassy, 2007). For each studied
8
genera, 2-3 representatives were further examined by 28S rRNA sequences to confirm the
9
identity. Briefly, nematode was kill by heat and mounted in a temperate slide for microscopy
10
inspection. DNA of single individual was freshly extracted by adding 10 µL solution of 0.05N
11
NaOH and 1 µL 4.5% Tween 20, incubating for 15 min at 95°C. Then 30 µL of
12
double-distilled water was added to each DNA sample, and subsequently amplified in PCR
13
using primer sets D2A (5’-ACA AGT ACC GTG AGG GAA AGT TG-3’) and D3B (5’-TCG
14
GAA GGA ACC AGC TAC TA-3’) (De Ley et al., 1999).
15
A total of 2769 individuals belonging to 19 genera (Acrobeles, Acrobeloides,
16
Amplimerlinius, Aphelenchoides, Aporcelaimus, Axonchium, Discolaimus, Ditylenchus,
17
Dorylaimus, Eudorylaimus, Helicotylenchus, Mesodorylaimus, Miconchus, Mylonchulus,
18
Panagrolaimus, Pratylenchus, Pristionchus, Rhabditis, and Mesocriconema) were
19
photographed, resulting 24 to 347 images per genus. These genera were selected as they are
20
cosmopolitan in soil, covering all nematode trophic guilds, and presenting in a large
21
population that allow us to acquire sufficient images for model training. Both juveniles and
22
adult stages were included in images, and head regions were given a higher priority, as they
23
involve more informative taxonomic characters. Tail, middle body regions and overall body
24
shape were also photographed, since these regions are also informative for species recognition
25
6
and easy to locate. Images were taken at three image planes, with magnification of 100X
1
objective for head and tail, and 10X for overall body.
2
2.2. Image pre-processing
3
The acquired images were pre-processed to facilitate the upcoming training process. The
4
images were gray-scale transformed to keep the consistency between each image, followed by
5
the edge detection. Canny edge detection algorithm was used to crop the region of interest
6
(ROI), which allowing nematodes be separated from the background while maximum
7
valuable information being preserved (Fig. 1). Briefly, the edge detector firstly computes the
8
horizontal and vertical coordinates of the ROI, then the nematode structures are automatically
9
cropped from the original image. After the cropping and Gray-scale transformation, the
10
processed images were rescaled to the size 224×224 pixels to meet the input requirement of
11
the chosen CNN.
12
2.3. Classification architecture and model training
13
The present study use the classification architecture and model developed by Lu et al.
14
(2021). Briefly, ResNet 101 pre-trained on ImageNet (He et el., 2016) was selected to
15
construct our model as it outperformed other mainstream networks in our preliminary tests.
16
The detail architecture for our model is presented in Fig. 2. Pytorch framework was used to
17
construct our neural network as it has been widely used in the research and computer
18
visualization industry. To avoid overfitting and to correlate with our dataset, dropout layers
19
were added and the number of SoftMax output was changed to 19 in ResNet 101 structure.
20
The acquired data were divided into trainset and testset at the ratio of 8:2, while augmentation
21
methods including random flip and Gaussian noise addition were also conducted to during
22
training. With more diverse training data, the classifier can learn features more thoroughly
23
while the immunity to certain level of noise can also be gained, which will improve the
24
7
general classification accuracy and prevent overfitting.
1
2.4. Web application and ecological indices calculation
2
To improve accessibility of above method for broad users, our trained model was
3
incorporated into a straightforward and user-friendly web application NemaRec (Fig. 3). The
4
NemaRec is composed of both front and back end. The front end was written in Javascript on
5
Visual Studio code, which includes general layout and user interaction scheme. The back end
6
was written with Flask in Python, which integrates the trained model as well as the
7
transformational pre-processing of the uploaded images.
8
During the identification, users can upload a serial of images taken from a specific
9
sample. As the classification results are returned to the front end, the number of each
10
identified species will be displayed and counted. The images will be turned into binary form
11
before being fed into the model, which subsequently outputs the predicted class and sends the
12
results to the front end. Based on CNN identification output, the NemaRec annotates
13
nematodes to five feeding types (bacterivore, fungivore, plant-parasitic, omnivore and
14
predator) and further calculates MI and PPI indices using the c-p values stated in Bongers
15
(1990). The web app is available at http://8.140.109.225:8866/, which allows the users to
16
upload their images via the front end and hence receive the results.
17
2.5. User data collection
18
Given the accuracy of a CNN model is related to the number of images used in model
19
training, the NemaRec can collect user uploaded images for future model adjustment. During
20
operation, users will upload their images (preferable the head and tail parts) via the front end.
21
The data will then be sent to the back end, where the trained model will analyse the output
22
and make the prediction. Finally, the results will be sent back to the front end to display, while
23
the ecological indices will also be computed (Fig. 3). After this procedure, user uploaded
24
8
images will be store in website server, available for download in a flexible time interval
1
depends on the amount of images. These images will be manually labelled to genus level
2
based on morphology, and subsequently split them into trainset and testset and used from
3
model training.
4
2.6. Test of effectiveness
5
The performance of NemaRec was evaluated using newly photographed nematode
6
samples and augmented images that independent from those used for model training.
7
First, a total of 190 images (10 images for each of genera) were acquired under the same
8
conditions (microscopy, nematode materials and photographer) of generating training set.
9
These images are identified in NemaRec, and the recognition rates are shown in Table 1.
10
To further verify the performance of NemaRec in multi-scenario, the newly
11
photographed nematodes were assigned to four groups of datasets. The datasets A, B and C
12
consist of nematodes from grassland, the wheat field, and the temperate deciduous broad-leaf
13
forest, respectively. These ecosystems have contrasting nematode communities, feeding types,
14
and ecological indices. For these datasets, nematode images were collected by students
15
without nematology experiences, containing blurred (Fig. 4 A, B), debris (Fig. 4C), broken
16
(Fig. 4D) and crushed nematode under cover slide (Fig. 4E), but representing the frequently
17
encountered scenarios. Since we only trained model with a limited number of soil nematode
18
taxa, we test the NemaRec in a more realistic scenario by adding 4-10 images which do not
19
belong to any of trained genera to these datasets A-C (e.g. Fig. 4F-J). The dataset D contained
20
nematodes from grassland, and were carefully prepared by first author, presenting images
21
with high quality (Fig. 4 K-O). All these datasets are available upon request from first author.
22
Aside from specimen-based dataset, we also generated the augmented dataset from the
23
images having been used in model training. This increases the data size and allows us to test
24
performance at optimal condition. We implemented two types of augmentation in which
25
9
testset images were randomly flip (vertical and horizontal) and Gaussian blurred. A total
1
number of 67 images were generated and used for modelling test.
2
3. Results
3
3.1. Model evaluation using augmented and specimen-based dataset
4
Currently, our developed application is capable to identify 19 genera (see the genera used
5
for photograph). To evaluate our models, we tested identification success for each of genera
6
and results are given in Table 1. In general, the success rate various greatly among genera
7
from 0% (Dorylaimus) to 90% (Acrobeles and Mesocriconema) with an average of 54.7%.
8
Apart from 6 genera (Dorylaimus, Discolaimus, Amplimerlinius, Helicotylenchus,
9
Mesodorylaimus and Aphelenchoides), all test genera have a success rate of 50% or higher.
10
Although our model failed to identify any of Dorylaimus, it mostly assigns query images as a
11
morphologically similar genus Eudorylaimus (70%). Indeed, these two genera often confused
12
by taxonomist. Similar misidentification is lead to most of low success rate, e.g.
13
Amplimerlinius was identified as other stylet-bearing genera (80%) and Helicotylenchus was
14
identified as Pratylenchus (70%). In compare, Discolaimus was surprisingly assigned to
15
several unrelated genera, regardless its distinct morphology.
16
Besides, we report the accuracy in a more realistic scenario, by using real ecological
17
datasets and different image qualities (Fig. 5). c-p values were calculated based on
18
identification results to examine how much it can be accidently correct, as different taxa may
19
share same value thus mask error in genus identification. As expected, the augmented dataset
20
outperforms the specimen-based dataset. Within specimen-based dataset, images with low
21
quality are generally hard to be identified, with only 29%-39% been correctly identified to
22
genus level, 53%-63% been correctly assigned with feeding type, and 51%-64% been
23
correctly assigned with c-p value. In contrast, high quality images can improve model
24
performance up to 76% in c-p value assignment.
25
10
3.3. Evaluation of maturity index calculation
1
The MI obtained from our model generally underestimated actual value, ranging from
2
17.1% in dataset B to 21.8% in dataset A, with the only exception of augmented dataset where
3
predicted value is 2.25% higher than the actual value. The augmented dataset predicts most
4
cases accurately while high quality dataset D surprisingly did not perform better than poor
5
images datasets in MI evaluation.
6
4. Discussion
7
The deep convolutional approaches have been a growing trend in the computer vision
8
field demonstrating impressive results in various tasks involving natural images. In present
9
study, we developed the web application NemaRec using the I-Nema dataset (Lu et al., 2021).
10
NemaRec adopts the deep CNN approach for nematode image identification, incorporate
11
I-Nema dataset in a user friendly interface, and further extends it to the calculation of MI and
12
PPI indices for environment evaluation. Compare with morphology based identification
13
which is time-consuming and expert-demanding, NemaRec is incorporated into a
14
user-friendly web application, and allows high-throughput image identification at a few
15
minutes. Such advantage is especially important for ecological studies as these researches
16
typically involve large dataset and majority ecologists are not expected to be nematode
17
taxonomists.
18
Although our model fails to achieve accuracy in a confident level (e.g. 95%), it is capable to
19
capture the general trends among samples. Besides, we observed that the manual
20
identification by non-specialists can inflict substantial mistakes that appears in similar
21
magnitude or even higher. Therefore, the advantage of our model can still serve as an valuable
22
tool for those without nematology experience.
23
24
The bottleneck that limits the accuracy in machine learning is the size of training
25
11
dataset. Specifically, sufficient number of nematode images is essential to achieve a
1
satisfactory identification result. However, this is barely possible by the efforts from one
2
single lab. Although the performance of current application is deemed far from satisfactory,
3
we demonstrated that the machine learning is promising in nematode identification, as well as
4
in MI and PPI indices calculation. More importantly, NemaRec opens a window of
5
international collaboration, and will be substantially benefit from collecting numerous user
6
uploaded images for the model training. Consequently, a much more functional and accurate
7
program is expected in future.
8
Funding
9
The work was fully supported by the National Natural Science Foundation of China
10
(grant number 32001876), and partly supported by the research grants PJ03906, PG00507 and
11
F002.251301 from Australia.
12
13
Credits authorship contribution statement
14
XQ conceived this work, collected image data, wrote this manuscript. YW and XL
15
developed software and analyzed data, XX, conducted model test and will maintain the web
16
application. XQ, YW, XL, XW and HL revised this manuscript.
17
Declaration of competing interest
18
We declare that there is no conflict of interest in the submission of this manuscript.
19
20
References
21
Anderson, R.C., 2000. Nematode parasites of vertebrates: their development and transmission,
22
12
2nd ed. CABI, Wallingford, UK.
1
Andrássy, I., 2005. Free-living nematodes of Hungary: Nematoda Errantia. Vol. 1. Hungarian
2
Natural History Museum, Budapest, Hungary.
3
Andrássy, I., 2007. Free-living nematodes of Hungary (Nematoda, Errantia). II. Hungarian
4
Natural History Museum. Budapest, Hungary.
5
Bongers, T., 1990. The maturity index: an ecological measure of environmental disturbance
6
based on nematode species composition. Oecologia 83, 1419.
7
https://doi.org/10.1007/BF00324627
8
Chen, L., Strauch, M., Daub, M., Jiang, X., Jansen, M., Luigs, H.-G., Schultz-Kuhlmann, S.,
9
Krüssel, S., Merhof, D., 2020. A CNN framework based on line annotations for detecting
10
nematodes in microscopic images. 2020 IEEE 17th International Symposium on
11
Biomedical Imaging (ISBI), Iowa City, IA, USA, pp. 508512. https://doi.org/
12
10.1109/ISBI45749.2020.9098465.
13
Chou, Y., Lee, D. J., Zhang, D., 2017. Edge detection using convolutional neural networks for
14
nematode development and adaptation analysis. In: International Conference on
15
Computer Vision Systems, pp. 228238. Springer.
16
https://doi.org/10.1007/978-3-319-68345-4_21
17
Coomans, A., 2002. Present status and future of nematode systematics. Nematology 4,
18
573582. https://doi.org/10.1163/15685410260438836.
19
Neher, D., Campbell, C., 1996. Sampling for regional monitoring of nematode communities
20
in agricultural soils. J. Nematol. 28, 196208.
21
De Ley, P., Felix, M.A., Frisse, L.M., Nadler, S.A., Sternberg, P.W., Thomas, W.K., 1999.
22
Molecular and morphological characterisation of two reproductively isolated species
23
with mirror-image anatomy (Nematoda: Cephalobidae). Nematology 1, 591612.
24
https://doi.org/10.1163/156854199508559
25
13
Derycke, S., Fonseca, G., Vierstraete, A., Vanfleteren, J., Vincx, M., Moens, T., 2008.
1
Disentangling taxonomy within the Rhabditis (Pellioditis) marina (Nematoda,
2
Rhabditidae) species complex using molecular and morhological tools. Zool. J. Linn. Soc.
3
152 (1), 115.
4
Erwin, T.L., 1982. Tropical forests: their richness in Coleoptera and other arthropod species.
5
Coleopts. Bull. 36, 7475.
6
Ferris, H., Bongers, T., 2006. Nematode indicators of organic enrichment. J. Nematol. 38(1),
7
312.
8
Freckman, D.W., 1988. Bacterivorous nematodes and organicmatter decomposition. Agr.
9
Ecosyst. Environ. 24, 195217.
10
He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition. In:
11
Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
12
770778. 2016. https://doi.org/10.1109/CVPR.2016.90
13
Hebert, P.D.N., Cywinska, A., Ball, S.L., deWaard J.R., 2003. Biological identifications
14
through DNA barcodes. Philos. Trans. R. Soc. Lond. B Biol. Sci. 270, 313321.
15
https://doi.org/10.1098/rspb.2002.2218
16
Holladay B., Willett, D., Stelinski, L., 2016. High throughput nematode counting with
17
automated image processing. BioControl 61(2), 177183.
18
https://doi.org/10.1007/s10526-015-9703-2
19
Javer, A., Brown, A., Kokkinos, I., Rittscher, J., 2019. Identification of C. elegans strains
20
using a fully convolutional neural network on behavioural dynamics. In: Leal-Taixé, L.,
21
Roth, S., eds. Computer Vision ECCV 2018 Workshops Springer International
22
Publishing. pp. 455464. https://doi.org/10.1007/978-3-030-11024-6_35.
23
Liu, M., Wang, X., Zhang, H., 2018. Taxonomy of multi-focal nematode image stacks by a
24
CNN based image fusion approach. Comput. Meth. Prog. Bio. 156, 209215.
25
14
https://doi.org/10.1016/j.cmpb.2018.01.016
1
Lu, X., Wang, Y., Fung, S., Qing, X., 2021. I-Nema: a biological image dataset for nematode
2
recognition. arXiv preprint. https://arxiv.org/abs/2103.08335.
3
Moore, J.C., de Ruiter, P.C., 1991. Temporal and spatial heterogeneity of trophic interactions
4
within below-ground food webs. Agr. Ecosyst. Environ. 34, 371397.
5
https://doi.org/10.1016/0167-8809(91)90122-E
6
Nadler, S., 2002. Species delimitation and nematode biodiversity: phylogenies rule.
7
Nematology 4, 615625. https://doi.org/10.1163/15685410260438908
8
Neher, D.A., 2001. Nematode communities as ecological indicators of agroecosystem health
9
In: Gliessman, S.R. ed. Agroecosystem sustainability: developing practical strategies.
10
Boca Raton, FL, CRC/Lewis Press, pp. 105120.
11
Porazinska, D.L., Giblin‐Davis, R.M., Esquivel, A., Powers, T.O., Sung, W.A.Y., Thomas,
12
W.K., 2010. Ecometagenetics confirm high tropical rainforest nematode diversity.
13
Molecular Ecology 19: 55215530. https://doi.org/10.1111/j.1365-294X.2010.04891.x
14
Porazinska, D.L., Giblin-Davis, R.M., Faller, L., Farmerie, W., Kanzaki, N., Morris, K.,
15
Powers, T.O., Tucker, A.E., Sung, W., Thomas, W.K., 2009. Evaluating high-throughput
16
sequencing as a method for metagenomics analysis of nematode diversity. Mol. Ecol.
17
Resour. 9, 14391450. https://doi.org/10.1111/j.1755-0998.2009.02611.x
18
Seinhorst, J.W., 1959. A rapid method for the transfer of nematodes from fixative to
19
anhydrous glycerin. Nematologica 4, 6769. https://doi.org/10.1163/187529259X00381
20
Sieriebriennikov, B., Ferris, H., de Goede, R.G.M., 2014. NINJA: an automated calculation
21
system for nematode-based biological monitoring. Eur. J. Soil Biol. 61, 9093.
22
https://doi.org/10.1016/j.ejsobi.2014.02.004
23
Silva, C.A., Magalhaes, K.M.C., Doria Neto, A.D., 2003. An intelligent system for detection
24
of nematodes in digital images. In: Proceedings of the International Joint Conference on
25
15
Neural Networks, vol. 1, pp. 2024. https://doi.org/10.1109/IJCNN.2003.1223431.
1
Wang, L., Kong, S., Pincus, Z., Fowlkes, C., 2020. Celeganser: automated analysis of
2
nematode morphology and age. In: Proceedings of the IEEE/CVF Conference on
3
Computer Vision and Pattern Recognition Workshops, pp. 968969.
4
https://doi.org/10.1109/CVPRW50498.2020.00492.
5
Wilson, M.J., Khakouli-Duarte, T., 2009. Nematodes as environmental indicators. Wallingford,
6
UK, CABI. https://doi.org/10.1079/9781845933852.0000.
7
Yeates, G.W., Bongers, T., de Goede, R.G.M., Freckman, D.W., Georgieva, S.S., 1993.
8
Feeding habits in soil nematode families and generaan outline for soil ecologists. J.
9
Nematol. 25, 315331.
10
11
16
Legends of figures.
1
2
Fig. 1. The example showing the effects of Canny edge detection and cropping algorithm
3
4
Fig. 2. The overall architecture of ResNet101 used in this study.
5
17
1
Fig. 3. The NemaRec user interface and the model re-training system by collecting user
2
uploaded images in program back end.
3
18
1
Fig. 4. Examples of dataset used for modelling test. Example of low quality images for (A)
2
and (B) Blurred. (C) Contains debris. (D) Broken specimen. (E) Crushed under cover slide.
3
(F-I) Example of species that were not used for modeling training but included in modelling
4
test. Example of images considered as high quality for (J) Acrobeles sp. (K) Pristionchus sp.
5
(L) Mylonchulus sp. (M) Ditylenchus sp.; (N) Mesodorylaimus sp.; (O) Aphelenchoides sp.
6
19
1
Fig. 5. Evaluation of NemaRec using different datasets. (A-C) Datasets with low quality
2
images and contain species not been included in training datasets. (D) Dataset with high
3
quality images. (E) Augmented dataset with images been randomly flipped and blurred. (F)
4
Comparison of true maturity index calculated based on manually identification (T) and
5
predicted value from NemaRec (P).
6
7
... I-NEMA Dataset: This dataset is a nematode public dataset containing multiple characteristic parts of different nematode genera (Xue et al., 2022). ...
... The method gave satisfactory results and made the identification process faster. Nemarec, a deep learning-based web application for the nematode identification process, was proposed by [20], [21]. They use their self-collected dataset for the research. ...
Article
Full-text available
Nematodes represent very abundant and the largest species diversity in the world. Nematodes, which live in a soil environment, possess several functions in agricultural systems. There are two huge groups of soil nematodes, a non-parasitic nematode, which contributes positively to ecological processes, and a plant-parasitic nematode, which cause various disease and reduces yield losses in the agricultural system. Early detection and classification in the agricultural area infected with plant-parasitic nematode and interpreting the soil level condition in this area required a fast and reliable detection system. However, nematode identification is challenging and time-consuming due to their similar morphology. This study applied a pre-trained faster region-based convolutional neural network (RCNN) for plant-parasitic and non-parasitic nematodes detection. These deep learning-based object detection models gave satisfactory results as the accuracy reached 87.5%.
Article
Full-text available
土壤动物学面临以全新知识体系为科学研究框架的变革时期,其核心内容是以数据驱动为主要特征的人工智能技术方法。目前广泛应用的基于数据库的数据处理分析方法,面临着数据多源异构、快速增长和处理能力不足之间的矛盾。基于快速发展的大数据科学和人工智能技术的数据挖掘方法在解决前述矛盾中有突出优势。但需要依赖一个强大的领域知识库。然而土壤动物领域知识图谱的研究十分匮乏。土壤动物知识图谱是一个具有有向图结构的知识库,其中图的节点代表与土壤动物相关的实体或概念ꎬ图的边代表实体或概念之间的各种语义关系,提出了土壤动物知识图谱的定义、内涵、理论模型和构建方法,以浙江天目山土壤螨类多样性为例,分析了构建山地土壤动物知识图谱的技术方法,以土壤动物多样性研究关注的物种分布、物种共存、环境条件对物种的影响作用为例,探讨了基于山地土壤动物知识图谱可以解决的相关科学问题。 研究表明,土壤动物知识图谱在解决生物多样性重要科学问题方面具有独特的潜力和优势,有力推动了土壤动物学、信息科学和数据科学交叉的土壤动物信息学的发展。
Article
Full-text available
Although much biological research depends upon species diagnoses, taxonomic expertise is collapsing. We are convinced that the sole prospect for a sustainable identification capability lies in the construction of systems that employ DNA sequences as taxon 'barcodes'. We establish that the mitochondrial gene cytochrome c oxidase I (COI) can serve as the core of a global bioidentification system for animals. First, we demonstrate that COI profiles, derived from the low-density sampling of higher taxonomic categories, ordinarily assign newly analysed taxa to the appropriate phylum or order. Second, we demonstrate that species-level assignments can be obtained by creating comprehensive COI profiles. A model COI profile, based upon the analysis of a single individual from each of 200 closely allied species of lepidopterans, was 100% successful in correctly identifying subsequent specimens. When fully developed, a COI identification system will provide a reliable, cost-effective and accessible solution to the current problem of species identification. Its assembly will also generate important new insights into the diversification of life and the rules of molecular evolution.
Conference Paper
The nematode C. elegans is a promising model organism to understand the genetic basis of behaviour due to its anatomical simplicity. In this work, we present a deep learning model capable of discerning genetically diverse strains based only on their recorded spontaneous activity, and explore how its performance changes as different embeddings are used as input. The model outperforms hand-crafted features on strain classification when trained directly on time series of worm postures.
Chapter
The nematode C. elegans is a promising model organism to understand the genetic basis of behaviour due to its anatomical simplicity. In this work, we present a deep learning model capable of discerning genetically diverse strains based only on their recorded spontaneous activity, and explore how its performance changes as different embeddings are used as input. The model outperforms hand-crafted features on strain classification when trained directly on time series of worm postures.KeywordsBehavioural phenotypingClassificationDeep learning
Article
Background and objective In the biomedical field, digital multi-focal images are very important for documentation and communication of specimen data, because the morphological information for a transparent specimen can be captured in form of a stack of high-quality images. Given biomedical image stacks containing multi-focal images, how to efficiently extract effective features from all layers to classify the image stacks is still an open question. Methods We present to use a deep convolutional neural network (CNN) image fusion based multilinear approach for the taxonomy of multi-focal image stacks. A deep CNN based image fusion technique is used to combine relevant information of multi-focal images within a given image stack into a single image, which is more informative and complete than any single image in the given stack. Besides, multi-focal images within a stack are fused along 3 orthogonal directions, and multiple features extracted from the fused images along different directions are combined by canonical correlation analysis (CCA). Because multi-focal image stacks represent the effect of different factors - texture, shape, different instances within the same class and different classes of objects, we embed the deep CNN based image fusion method within a multilinear framework to propose an image fusion based multilinear classifier. Results The experimental results on nematode multi-focal image stacks demonstrated that the deep CNN image fusion based multilinear classifier can reach a higher classification rate (95.7%) than that by the previous multilinear based approach (88.7%), even we only use the texture feature instead of the combination of texture and shape features as in the previous work. Conclusions The proposed deep CNN image fusion based multilinear approach shows great potential in building an automated nematode taxonomy system for nematologists. It is effective to classify multi-focal image stacks.
Conference Paper
The Antarctic nematode Plectus Murrayi is an excellent model organism for the study of stress and molecular mechanisms. Biologists analyze its development and adaptation by measuring the body length and volume. This work proposes an edge detection algorithm to automate this labor-intensive task. Traditional edge detection techniques use predefined filters to calculate the edge strength and apply a threshold to it to identify edge pixels. These classic edge detection techniques work independently of the image data and their results are sometimes inconsistent when edge contrast varies. Convolutional Neural Networks (CNNs) are regarded as powerful visual models that yield hierarchies of features learned from image data, and perform well for edge detection. Most CNNs based edge detection methods rely on classification networks to determine if an edge point exists at the center of a small image patch. This patch-by-patch classification approach is slow and inconsistent. In this paper, we propose an efficient CNN-based regression network that is able to produce accurate edge detection result. This network learns a direct end-to-end mapping between the original image and the desired edge image. This image-to-edge mapping is represented as a CNN that takes the original image as the input and outputs its edge map. The feature-based mapping rules of the network are learned directly from the training images and their accompanying ground truth. Experimental results show that this architecture achieves accurate edge detection and is faster than other CNN-based methods.
Article
Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers---8x deeper than VGG nets but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.
Article
Nematode counting forms the basis for almost every assay in nematology: population surveys and culture density estimates all rely on accurate, rapid nematode counting. Accurate, rapid nematode counting is especially important for bioassays of entomopathogenic nematodes used for biological control. While manual microscope-based counting has traditionally been the standard, automated image processing holds promise for high-throughput nematode counting. Here we develop image capture and processing techniques to facilitate standard curve development and automated counting of two species of entomopathogenic nematodes. The techniques not only produce accurate nematode counts but also are rapid: timesavings over traditional manual counting are large and increase with increasing sample size. These techniques will likely be generally useful for quantification of all nematode species and potentially other small animals requiring quantification using microscopy. © 2015 International Organization for Biological Control (IOBC)