Automatic Detection and Identification of Trichomonas Vaginalis

from Fluorescence Microscopy Images

Yongjian Yu

and Jue Wang

Axon Connected, LLC, Earlysville, VA 22936, U.S.A.

Department of Mathematics, Union College, Schenectady, NY 12308, U.S.A.

Keywords: Detection, Segmentation, Multiscale, Trichomoniasis, Trichomonas Vaginalis, Fluorescence Microscopy.

Abstract: Trichomonas vaginalis (TV) causes sexually transmitted infections that, if unresolved timely, can lead to

adverse health conditions. We construct a software platform integrating a novel, robust multiscale image

analysis pipeline for automatic detection and characterization of TV from dual-resolution, multi-band digital

fluorescence microscopy scans. We develop two spectral indices to highlight the TV in the spectrally

contaminated image. The system employs a search algorithm that incorporates the spectral indices to locate

the microorganisms from the low-resolution scans across the sample slide, and then identifies the TV using

a multiscale edge-sensitive automatic thresholding segmentation and index-driven ranking in the high-

resolution view. Method capability is demonstrated through the discriminability in the feature classification

and in the TV test pipeline, both showing a high sensitivity. This technique can be used to enable automatic,

fast diagnosis of trichomoniasis at the point-of-care clinics.

1 INTRODUCTION

Trichomoniasis (or trich for short) is the most

prevalent non-viral sexually transmitted infection

(STI) in the world (Bahadory et. al. 2021; WHO

2021a). It is caused by infection with a protozoan

parasite called Trichomonas vaginalis (T. vaginalis,

or TV for short). In 2020, the World Health

Organization (WHO) estimated 156 million new

infections of trich (WHO 2021b). According to the

latest WHO report, the estimates for trich were 6.3%

(95% UI: 4.0–7.2) in women and 0.6% (95% UI:

0.4–0.9) in men (Rowley et. al. 2019). Because no

recommendations are available for general screening

for TV, the epidemiology of trich has largely come

from population-based and clinic-based surveillance

studies (CDC 2021).

TV infection can be overlooked by clinicians, as

the process generally follows a benign course and is

frequently asymptomatic. The majority of people

who have trich (70–85%) either have minimal or no

genital symptoms, and untreated infections might

last from months to years (CDC 2021). Symptoms of

TV may be non-specific, making it difficult to

differentiate TV from other STIs clinically, which

require different treatment approaches. TV infection

in women is associated with vaginitis, urethritis,

cervicitis, and pelvic inflammatory disease. TV

infection can cause adverse pregnancy complica-

tions, such as tubal infertility, preterm delivery, low

birth weight, and premature rupture of membranes

(Webb et. al. 2021). TV transmission from mother to

child has been associated with neonatal morbidities,

including vaginitis, urinary tract infection, and

respiratory disease. TV infection has also been

linked to cervical human papillomavirus (HPV)

infection and cervical cancer (Amorim et. al. 2017;

Bahadory et. al. 2021). Among men, TV infection

has been associated with benign prostatic

hyperplasia and invasive prostate cancer (Bahadory

et. al. 2021; Webb et. al. 2021). Moreover, TV

infection is associated with an increased risk of

human immunodeficiency virus (HIV) transmission

and acquisition (Masha et. al. 2019). Up to 53% of

women with HIV have TV infections (CDC 2021).

Without inclusive testing protocols, TV infections

largely go undiagnosed and untreated. Therefore,

improving the detection of TV is of high

significance and impact.

The standard test for trich at the near point-of-

care (nPOC) clinics uses microscopic examinations

of the fresh saline wet mount to identify moving

organisms as TV. Wet mount microscopy is

considered insensitive compared to the culture

method (Nathan et. al. 2015). The sensitivity is

reported to be 52% in a recent study of 136

participants (Hsieh et. al. 2020). Current screening

for trich is performed using tests with limited

190

Yu, Y. and Wang, J.

Automatic Detection and Identiﬁcation of Trichomonas Vaginalis from Fluorescence Microscopy Images.

DOI: 10.5220/0010993400003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 2: BIOIMAGING, pages 190-197

ISBN: 978-989-758-552-4; ISSN: 2184-4305

sensitivity, including wet mount microscopy

(sensitivity 38–68%), culture of specimens (44–

88%), nucleic acid hybridization (roughly 60%),

rapid antigen detection (83–86%), and nucleic acid

amplification tests (NAATs) (76–100%) (Andrea

and Chapin 2011; CDC 2021; Gaydos et. al. 2017;

Hsieh et. al. 2020; Hobbs and Seña 2013; Nathan et.

al. 2015; Patil et. al. 2012). The culture method has

a relatively longer turn-around time and is limited by

non-viable organisms in the specimen. The FDA-

cleared NAATs can improve the quality of trich

diagnostics, but they use large robotic platforms and

are time-consuming, thus not feasible for nPOC.

Moreover, the result is typically yes or no, rather

than being quantitative. The histochemical methods

use non-specific stains, which require interpretations

by the cytologists, pathologists, or other experts.

The emerging immunofluorescence (IF) assay

that binds the trich specific molecules enables an

easy interpretation and thus lowers the critical

errors. The IF assay makes it possible to locate,

highlight, and count all living, immobile, dead, and

even partially destroyed TV. In addition, the IF

assay is more sensitive than the wet mount

brightfield microscopic examinations. An automatic

IF assay equipped with motorized scanner and

automatic data analyzer is highly desired. It is a

challenging task to automatically detect the rare TV

from the IF micrographs, especially in the early

stage of infection. The TV is embedded among other

microorganisms such as white blood cells, epithelial

cells, yeast, hyphae, and swab debris, etc. The

multiband image suffers from degradation due to

channel crosstalk, local defocusing, biological

sample variability, and preparation variability.

Example IF micrographs of clinical samples are

shown in Figure 1 at a 4X objective and Figure 2 at

a 40X objective, respectively. The color is assigned

as follows: red for channel 3 (epithelial cells), green

for channel 2 (yeast and hyphae of candida

albicans), grayscale for channel 1 (nuclei of various

microorganisms including the bacteria, epithelial

cells, TV, white blood cells, and yeast), and blue for

channel 4 (TV). The pseudo-color for display is

different from the LED illumination or fluorophore

emission waveband, which is further elaborated in

section 2.1. The non-specific staining can be seen in

the composite image. The TV is oval shaped and has

a faint eccentric, elongated nucleus. The organism

varies in size but is typically around 10 μm in length

and 7 μm in width, slightly larger than the white

blood cell (WBC). It can assume an amoeboid form

when attached to the vaginal epithelial cells.

Towards detection efficiency and robustness, the

algorithm needs to first find the TV candidates using

a 4X objective, and then identify the TV within a

40X objective high-resolution view.

We propose two normalized TV spectral indices

along with a pipeline of data processing algorithms

for a rapid search and identification of TV from IF

micrographs. We analyze the discriminability of the

indices and demonstrate their potential after being

integrated into an automated digital IF microscopic

system for a computer-aided diagnosis of trich.

Figure 1: Composite images of pseudo-colored four-band

IF micrographs at a 4X objective. The locations of

possible TV (blue) are marked by the red circles.

Figure 2: Composite image of pseudo-colored four-band

IF micrograph at a 40X objective. The TV approximately

centered in the image.

2 METHODS

2.1 Problem Statement

A research digital IF microscopic platform under

development and optimization is utilized for this

Automatic Detection and Identiﬁcation of Trichomonas Vaginalis from Fluorescence Microscopy Images

191

study. The platform runs with dual resolutions, four

spectral channels, and is configurable to scan and

analyze vaginal microbiome swab samples for

aiding diagnosis of bacterial vaginosis, candida

albicans infection, and trichomoniasis (trich). The

four-channel image can be expressed by

𝑰



𝐼



𝐼



𝐼



𝐼





. (1)

The 𝐼



, 𝐼



, 𝐼



and 𝐼



represent the fluorescence

images in the four emission spectral channels (lime,

green, blue, and far red). The LED illumination

spectrum for each channel image is at a shorter

wavelength than the respective stain emission

wavelength. These images contain numerous

different microorganisms, in the form of isolated

cells or colony of cells, such as the epithelium,

fungus, DNA materials/bacteria, and TV. For the

diagnosis of trich, a TV-specific antibody staining

enables TV expression mostly in 𝐼



. Likewise, 𝐼



, 𝐼



and 𝐼



express selectively the stained epithelium,

fungus, and DNA materials/bacteria, respectively.

Because of the channel crosstalk problem, these

images are not mutually independent in terms of the

contents. In the TV expression channel 𝐼



, there

exist distractive interfering images of irrelevant

objects. Likewise, the other channels have their own

latent objects of interest, but also contain the ghost

images of the other unwanted objects. The presence

of channel crosstalk reduces the sensitivity and

accuracy of finding, segmenting, and identifying

TV. Detection of TV with a high sensitivity becomes

a challenging task in the case of only a single TV

presence. The difficulty is further compounded by

the background fluorescence and debris that increase

the false detection rate.

2.2 Segmentation and Identification

The dual-resolution system uses a 4X and a 40X

objective to find and identify the TV, respectively.

2.2.1 TV Search

A 4X scan of the sample slide yields 25 sub-images,

each with 2K-by-2K pixels. The pixel pitch of the

image is approximately 1.6 μm. The 4X TV search

algorithm consists of finding the boundary of the

organism and its nucleus, blob quantification, and

support vector machine (SVM) classification of the

blobs into two classes: TV alike and non-TV.

The external boundary segmentation of TV is

performed via adaptive thresholding of the TV-

sensitive channel 4X 𝐼



. Considering the tiny size of

the trich nucleus (1 to 3 pixels), we apply the top-hat

Figure 3: 4X TV search and segmentation. The contour

delineates the TV external boundary.

filtering of DNA-sensitive channel 4X 𝐼



for bright

objects to extract the trich nuclei. Top-hat filtering

computes the morphological opening of the image

and then subtracts the result from the original image.

The TV candidates are masked as nucleated blobs in

the 4X 𝐼



segmentation map. Any clump of nuclei is

identified by a size thresholding and then split into

individual TV nucleus using a marker-controlled

watershed method. The nuclear markers are located

as the local extended maxima. Similarly, a touching

cluster of TV is identified by the number of nuclei

in the cluster and spilt via the watershed controlled

by the individually segmented nucleus. Finally, the

BIOIMAGING 2022 - 9th International Conference on Bioimaging

192

strong cytokeratin and yeast areas are masked out

using a set union:

𝐻𝐼



𝐼



𝜀∪𝐻𝐼



𝐼



𝜀, (2)

where 𝐻∙ is the Heaviside function acting as a

thresholding operator; the ∪ denotes the union of

sets; 𝜀 is a smaller number in the range of 0-10. An

example of the 4X TV segmentation is shown in

Figure 3.

The initial TV candidates are numerically

labelled. Each blob will be further quantified by the

features outlined in the later sections and classified

into TV alike or non-TV. Each 4X TV candidate is

then imaged with a 40X objective and examined to

identify if it is a true TV.

2.2.2 TV Segmentation with Multiscale

Edge-sensitive Automatic

Thresholding

In order to accommodate for the shadow depth of

focus with the 40X objective imaging, which may

cause the local out-of-focus issue, we develop a

multiscale edge-sensitive automatic thresholding

technique for 40X TV segmentation. The edge

information facilitates locating the objects in dim

regions where they are likely to be missed by

grayscale thresholding, and the automatic

thresholding recovers the cells with undefined edges

due to local defocusing. The method optimizes an

objective functional, i.e., maximizing the Dice shape

similarity between the multiscale edge synthesized

segmentation and automatic thresholding

segmentation of an input image 𝐼 , expressed as

follows,

𝑇

∗

max















∩





|











|





, (3)

where 𝐼



is an L-level linear quantized image of 𝐼

(40X 𝐼



or 𝐼



); the quantization compresses the range

of the threshold value 𝑇 to an integer set {1, 2, …,

L}. 𝐻∙ is the Heaviside function acting as a

thresholding operator. 𝐶 denotes the edge

synthesized contours by morphological linking of

the multiscale edges 𝐸 of the input image 𝐼, 𝐶



𝐸⊕𝑆

∥



⊕𝑆



, where ⊕ stands for the dilation

operator; 𝑆

∥

and 𝑆



are the horizontal and vertical

line structure elements of length 3, respectively; 𝐹



∙



is the filling operator on 𝐶 . The final 40X TV

segmentation is given by the fusion 𝐻



𝐼



𝑇

∗



∪

𝐹



𝐶



The morphological linking synthesized edge 𝐶

may contain false contours resulted from the

clustered edges of noise. Moreover, the algorithm

itself is unable to determine if the segmentation is

good or bad without the ground truth. To overcome

these issues, we build and incorporate some ground

truth data into the proposed segmentation paradigm.

In recognition of the approximate circular shape

prior of the objects of interest, we perform a circular

Hough transform to 𝐼 to find all circles with radii in

the radius range of TV (or TV nuclei if segmenting

the TV nuclei). This results a network of circles 𝐶



Since 𝐶



is much resilient to image noise and fuzzy

edges, 𝐹



𝐶





is well deserved to be an approximate

ground truth semantic segmentation. We regularize

𝐹



𝐶



using 𝐹



𝐶





by maintaining only the

segmentations in 𝐹



𝐶



that are also in 𝐹



𝐶





. When

the foreground is well segmented, the shapes of

𝐹



𝐶



and 𝐹



𝐶





are close to each other, leading to a

high shape similarity. When 𝐶 deviates from 𝐶



significantly, 𝐶 becomes unreliable. We set the Dice

metric threshold of 0.5 for an acceptable goodness

measure of 𝐶 with reference to 𝐶



. If the Dice metric

is below 0.5, we use 𝐶



rather than 𝐶 as the final

object segmentation.

The edge-sensitive automatic thresholding

segmentation with shape regularization and

approximate ground truth support solves the

problem (to a certain extent) of conventional

segmentation algorithms (Ray and Saha 2007),

where the algorithms are unable to tell if the results

are satisfactory or not unless a human user inspects

the results.

2.3 Normalized Tv Spectral Indices

Denote 𝐽



the latent fluorescence signal in the raw

image 𝐼



(k =1 to 4). The raw signal in the k

channel can be modelled as a linear combination of

the latent data across four channels,

𝐼





∑

𝛼



𝐽







𝑛



, (4)

where 𝛼



is a 4 by 4 real-coefficient matrix, with

unity diagonal elements, i.e., 𝛼



1; 𝑛



is the

background and noise. The off-diagonal non-zero

coefficients reflect the residual of channel crosstalk,

non-specific staining, and sample variability.

The research microscope system in use does not

provide channel crosstalk correction, and thus the

latent image is inaccessible. To solve the channel

crosstalk problem with the raw data, we proposed

two spectral indices for a robust search of TV alike

with a low-resolution view, followed by quality

verification with a high-resolution view. A pre-

processing is first performed to convert the raw

image into an image appropriate for human

Automatic Detection and Identiﬁcation of Trichomonas Vaginalis from Fluorescence Microscopy Images

193

visualization of the targets of interest, including

noise reduction, contrast enhancement, and

background subtraction. The TV spectral indices are

defined as

TVSI



𝐼





𝐼





/𝐼





𝐼





, (5)

TVSI



𝐼





𝐼





/𝐼





𝐼





, (6)

where 𝐼





𝑝



𝐼





,𝑘1 to 4; 𝑝𝐼 denotes a series

of operations on 𝐼, including the median filtering,

histogram stretching, and background subtraction.

The denominators in (5) and (6) normalize the

spectral indices to highlight the dim TV by

suppressing the distractive interference of ghost

images of unwanted objects. Both indices are less

than 1, but they can assume small negative values

because the illumination in each channel is

unbalanced. In section 3 results, we will mainly

demonstrate the discriminability of these two

indices, coupled with other features, for TV search

and identification.

2.4 TV Ranking for Identification

The TV is of similar size with many district clutter

organisms, such as the (candida albicans) yeast cells,

monocytes, and some debris co-existing in the trich

test sample. These clutter organisms express

fluorescence in the same waveband as the TV does.

Consequently, accurate TV enumeration is hindered

by the false positive rate resulted from

misidentifying those yeast cells, monocytes and

debris as TV. To increase the accuracy of trich

diagnosis, we develop a TV ranking model to

differentiate quantitatively the detected TV from

yeast or debris. This model incorporates the spectral

indices, sizes of TV and its nucleus, and a condition

that the TV must have a nucleus.

The DNA marker channel 𝐼



is first enhanced

around the 4X candidate region using a multiscale

blobness filter bank, followed by a binarization of

the blobness image using the clustering method

(Otsu 1979), biovolume elasticity method (Luo et.

al. 2018), or locally adaptive thresholding method

(Singh et. al. 2011). The binarized map contains the

TV nucleus, as well as other DNA materials (e.g.,

nuclei of WBC), which may cluster with the TV

nucleus under examination. The nuclear clumps are

split using the watershed transform. Denote 𝐵



and

𝐵



the segmentation masks of TV and its nucleus,

respectively. 𝐵



is further split until each area has

one nucleus in it. The true TV nucleus is determined

via a merit scoring and selection process. The

ranking model is defined by

𝑅 𝑆

〈

TVSI



〉

𝜇



𝜎



𝑆

〈

TVSI



〉

𝜇



𝜎





𝑓



𝐵



,𝐵





𝑔𝐴



,𝐴



; 𝑀



,𝑀



,𝐷



,𝐷



, (7)

where 𝑆∙ is the Sigmoid function parametrized by

the offset μ and spread σ;

〈

∙

〉

denotes the averaging

of cell segmentation; 𝑓



𝐵



,𝐵







𝐵



∩𝐵



𝐵



;

the ∩ and

∙

denote set intersection and cardinality,

respectively; 𝑔𝐴



,𝐴



; 𝑀



,𝑀



,𝐷



,𝐷



 is the size

driven probability given by

𝑔



𝐴



,𝐴



; 𝑀



,𝑀



,𝐷



,𝐷







exp



























exp



























, (8)

where 𝐴



and 𝐴



are the trich area and nucleus area,

respectively; 𝑀



,𝑀



,𝐷



,𝐷



are the means and

standard deviations of the sizes of TV and nucleus,

respectively.

3 RESULTS

We demonstrate the capability and power of the

proposed platform for TV testing. Firstly, the

spectral indices are employed as the discriminative

features in the classification framework for TV and

non-TV organisms. Secondly, the spectral indices,

segmentation and ranking algorithms are integrated

into a complete system for TV quantification.

In order to study how the normalized TV

spectral indices perform as features for TV

classification at a low resolution (4X objective), we

construct four discriminative feature sets as follows:

𝑆





〈

TVSI



〉

〈

TVSI



〉

〈

𝐼





〉

,

𝑆





〈

TVSI



〉

〈

TVSI



〉

〈

𝐼





〉

〈

𝐼





〉

,

𝑆





〈

TVSI



〉

〈

TVSI



〉

〈

𝐼





〉

〈

𝐼





〉

,𝐴



,𝐴



,

𝑆



 Union of 𝑆



and {nucleus eccentricity, cell-

minor-to-major-axis-length ratio}.

The nucleus eccentricity measures the distance

between the centroid of the cell and nucleus. We

apply the t-Distributed Stochastic Neighbor

Embedding (t-SNE) to visualize the clusters of

feature points in a two-dimensional plane based on

the feature relative similarities in a high-dimensional

feature space that corresponds to the user labels. The

optimal feature set 𝑆



is identified with t-SNE; the

result is shown in Figure 4. Using 𝑆



, we train a

linear and a nonlinear SVM classifier over 64 4X

cell images that contain 31 TV positive and 33 TV

negative. The samples are split randomly to 85%

(54) training and 15% (10) testing. An accuracy of

90% is achieved for the linear SVM and the

BIOIMAGING 2022 - 9th International Conference on Bioimaging

194

accuracy increases to 100% with the nonlinear SVM

using a radial basis function (RBF) kernel.

The TV detection and segmentation result at a

high resolution is shown in Figure 5 (b), with a

probability of 0.89 according to the ranking model.

The normalized spectral index maps are illustrated in

Figure 5 (c) and (d) that highlight the TV. A region

of interest (ROI) is selected for processing. The

procedure of nucleus detection and segmentation is

illustrated in Figure 6.

To quantify the performance of the integrated IF

system, 33 clinical TV samples are scanned and

analyzed. Slide preparation takes places in the slide

well. The sample is fixed with methanol for 5

minutes. A dilated solution of α-T. vaginalis, p65

adhesive antigen antibody is applied. The sample is

incubated for 15 minutes at room temperature. Then

DAPImount is applied and covered with a coverslip.

The overall time for slide preparation is less than 30

minutes including wash and dry time. The sample

slide is automatically scanned and trichomonads are

enumerated and reported in about 15 minutes.

The test results are compared to the experts’ data

reading. The sensitivity of the IF TV test is 100%,

specificity is 94%, and accuracy is 97%. We point

out that the relative lower classification accuracy in

the 4X classification is compensated by the 40X

ranking mechanism thus a high overall system level

performance is expected. The method comparison is

summarized in Table 1. Compared to the standard

wet mount microscopy and other methods used for

trich diagnosis, the IF test achieves a superior

performance. Coupling the 4X and 40X algorithms

in the sample evaluation, the developed technique is

able to deliver a test with a high sensitivity and

accuracy for quick TV detection and identification.

Table 1: Comparison of sensitivity.

Wet mount microscopy 38–68%

Culture 44–88%

Nucleic acid hybridization roughly 60%

Rapid antigen detection 83–86%

Nucleic acid amplification tests 76–100%

Proposed integrated IF 100%

(a) (b)

(

)

(

)

Figure 4: Classification of TV and other cells using the discriminative features. (a) Three-feature set 𝑆



. (b) Four-feature set

𝑆



. (c) The optimal six-feature set 𝑆



. (d) All eight features 𝑆



Automatic Detection and Identiﬁcation of Trichomonas Vaginalis from Fluorescence Microscopy Images

195

(a) (b)

(e)

Figure 5. (a) Composite image of the pseudo-colored four-band IF micrograph at a 40X objective. (b) Segmentation

contours of the TV and its nucleus, with the ROI box overlaid on the composite image. (c) Spectral index TVSI

map. (d)

Spectral index TVSI

map. (e) Segmentation masks of the TV (gray) and its nucleus (white).

(

)

(

)

(

)

(

)

Figure 6: (a) Preprocessed nucleus channel. (b) ROI blobness enhancement. (c) Blobness map. (d) TV nucleus.

BIOIMAGING 2022 - 9th International Conference on Bioimaging

196

4 CONCLUSIONS

We have developed a set of novel spectral indices

along with cell segmentation and ranking algorithms

for a quick, robust search and quantitative

identification of TV from fluorescence micrographs

of specimen samples for assisting trich diagnosis in

a point-of-care setting. We demonstrate that the

proposed spectral indices are strong discriminative

TV features against artifacts. A low-resolution

search algorithm and a high-resolution identification

algorithm are integrated into the testing framework

and data processing pipeline. We devise an edge-

sensitive automatic thresholding method that

incorporates a cost minimization with an implicit

shape regularization and self-validation. It is capable

of extracting the TV with a wide range of signal

levels and edge strengths. Moreover, this method is

generic for cell detection. Our technique has shown

promising results and achieved a high sensitivity and

accuracy. We implement a software system that

eliminates user variability in slide reading and offers

the ability to archive images. The system enables a

real-time, more accurate assessment of trich

infections.

REFERENCES

A. T. Amorim, L. M. Marques, G. B. Campos, et al., “Co-

infection of sexually transmitted pathogens and

Human Papillomavirus in cervical samples of women

of Brazil”, BMC Infect. Dis., 17(1):769, 2017.

S. B. Andrea and K. C. Chapin, “Comparison of Aptima

Trichomonas vaginalis transcription-mediated

amplification assay and BD Affirm VPIII for detection

of T. vaginalis in symptomatic women: performance

parameters and epidemiological implications”, J. Clin.

Microbiol., 49(3):866-869, 2011.

S. Bahadory, S. Aminizadeh, A. Taghipour, F. Bokharaei-

Salim, K. Khanaliha, M. H. Razizadeh, A. Soleimani,

L. beikzadeh and A. Khatami, “A systematic review

and meta-analysis on the global status of Trichomonas

vaginalis virus in Trichomonas vaginalis”, Microbial

Pathogenesis, Vol. 158, 105058, 2021.

Centers for Disease Control and Prevention (CDC),

Trichomoniasis, Sexually Transmitted Infections

Treatment Guidelines, 2021.

C. A. Gaydos, J. D. Klausner, N. P. Pai, H. Kelly, C.

Coltart and R. W. Peeling, “Rapid and point-of-care

tests for the diagnosis of Trichomonas vaginalis in

women and men”, Sex Transm. Infect., 93(S4):S31-

S35, 2017.

M. M. Hobbs and A. C. Seña, “Modern diagnosis of

Trichomonas vaginalis infection”, Sex Transm. Infect.,

89(6):434-438, 2013.

Y. H. Hsieh, M. K. Lewis, V. G. Viertel, D. Myer, R. E.

Rothman and C. A. Gaydos, “Performance evaluation

and acceptability of point-of-care Trichomonas

vaginalis testing in adult female emergency

department patients”, Int. J. STD AIDS, 31(14):1364-

1372, 2020.

T. L. Luo, M. C. Eisenberg, M. A. L. Hayashi, C.

Gonzales-Cabezas, B. Foxman, C. F. Marrs and A. H.

Rickard, “A sensitive thresholding method for

confocal laser scanning microscope image stacks of

microbial biofilms”, Sci. Rep., Vol. 8, 13013, 2018.

S. C. Masha, P. Cools, E. J. Sanders, M. Vaneechoutte and

T. Crucitti, “Trichomonas vaginalis and HIV infection

acquisition: a systematic review and meta-analysis”,

Sexually Transmitted Infections, 95(1):36-42, 2019.

B. Nathan B, J. Appiah J, P. Saunders, D. Heron, T.

Nichols, R. Brum, S. Alexander, P. Baraitser and C.

Ison, “Microscopy outperformed in a comparison of

five methods for detecting Trichomonas vaginalis in

symptomatic women”, Int. J. STD AIDS, 26(4):251-

256, 2015.

N. Otsu, “A Threshold Selection Method from Gray-Level

Histograms”, IEEE Trans. Syst. Man Cybern. Syst.,

Vol. 9, No. 1, 62-66, 1979.

M. J. Patil, J. M. Nagamoti and S. C. Metgud, “Diagnosis

of trichomonas vaginalis from vaginal specimens by

wet mount microscopy, in pouch tv culture system,

and PCR”, J. Glob. Infect. Dis., 4(1):22-5, 2012.

N. Ray and B. Saha, “Edge sensitive variational image

thresholding”, IEEE ICIP, 16-19, 2007.

J. Rowley, S. Vander Hoorn, E. Korenromp, N. Low, M.

Unemo, L. J. Abu-Raddad, R. M. Chico, A. Smolak,

L. Newman, S. Gottlieb, S. S. Thwin, N. Broutet and

M. M. Taylor, “Chlamydia, gonorrhoea,

trichomoniasis and syphilis: global prevalence and

incidence estimates, 2016”, Bulletin of the World

Health Organization, 97(8), 548-562P, 2019.

T. R. Singh, S. Roy, O. I. Siggh, T. Sinam and K. M.

Singh, “A new local apative thresholding technique in

binarization”, IJCSI International Journal of

Computer Science Issues, Vol. 8, Issue 6, No. 2, 271-

277, 2011.

B. Webb, A. Crampton, M. J. Francis, J. Hamblin, T. M.

Korman and M. Graham, “Increased diagnostic yield

of routine multiplex PCR compared to clinician

requested testing for detection of Trichomonas

vaginalis”, Pathology, Vol. 53, Issue 2, 257-263,

2021.

World Health Organization (WHO), “Global progress

report on HIV, viral hepatitis and sexually transmitted

infections”, July 2021.

World Health Organization (WHO), “Sexually transmitted

infections (STIs)”, Nov. 2021.

Automatic Detection and Identiﬁcation of Trichomonas Vaginalis from Fluorescence Microscopy Images

197