Pushing the limits of antibiotic susceptibility testing

UPTEC X 15 010

Examensarbete 30 hp

September 2015

Pushing the limits of antibiotic

susceptibility testing

 

Degree Project in Molecular Biotechnology

Masters Programme in Molecular Biotechnology Engineering, Uppsala University School of Engineering

UPTEC X 15 010

Date of issue 2015-09

Author

Pikkei Yuen

Title

Pushing the limits of microfluidic antibiotic susceptibility testing

– an image analysis approach

Abstract

Sepsis is a life-threatening condition caused by bacteria entering the blood flow. Delays in

effective antibiotic treatment have been shown to be associated with increased mortality

rate, longer hospital stays and higher hospital costs. A microfluidic assay has previously

been evaluated for determination of the minimal inhibitory concentration (MIC) using

clinical isolates of Staphylococcus aureus in combination with vancomycin. This study

investigates the optimal inoculum level in the microfluidic assay for S. aureus, Escherichia

coli, Klebsiella pneumoniae and Pseudomonas aeruginosa. The detection level varies with

analysis method and detection system; cluster-based analysis of the image data from the

microfluidic assay could lower the detection level 10-100 fold. Also, capture of image data

in the oCelloScope rather than darkfield microscope could decrease the time to detection.

Keywords

sepsis, rapid AST, antibiotic resistance, minimal inhibitory concentration, microfluidics,

im-age analysis, intensity, blob detection

Supervisors

Pernilla Lagerbäck, Christer Malmberg

Department of Medical Sciences, Uppsala University

Scientific reviewer

Karin Hjort

Department of Medical Biochemistry and Microbiology, Uppsala University

Project name

Sponsors

Gradientech AB

Language

English

Security

Secret until 2016-09

ISSN 1401-2138

Classification

Supplementary bibliographical information

Pages

57

Biology Education Centre

Biomedical Center

Husargatan 3, Uppsala

Pushing the limits of microfluidic antibiotic

susceptibility testing

- an image analysis approach

Pikkei Yuen

Populärvetenskaplig sammanfattning

Blodförgiftning (sepsis) är ett allvarligt tillstånd som hastigt kan försämras och bli livshotande.

Tillståndet uppstår när det finns bakterier i blodet som kroppen försvarar sig mot genom att

starta en inflammation. Symptom är bland annat hög feber, frossa, hjärtklappning och andnöd.

Patienter med sepsis behandlas normalt med antibiotika, men den globalt ökande förekomsten

av antibiotikaresistenta bakterier ökar sannolikheten att den antibiotika som man valt att

behan-dla patienten med är ineffektiv. I dagsläget tar det flera dagar innan man vet vilken bakterie som

orsakar tillståndet och ifall den har resistens mot någon antibiotika. Det finns därför ett behov

av metoder som kan detektera antibiotikaresistens mycket snabbare. Detta skulle då innebära

en lägre sjukvårdskostnad och högre överlevnadschans för patienten.

Tidigare studier har visat att det mikrofluidiska systemet CellDirector® 3D (Gradientech AB)

kan användas för att bestämma den bakteriehämmande antibiotikakoncentrationen inom några

timmar. I det här projektet undersöktes den lägsta detektionsgränsen i systemet för två av de

vanligaste bakteriearter som förekommer vid sepsis. För att kvantifiera bakterietillväxt i

Cell-Director 3D har intensitetsförändring använts som standard, men denna studie föreslår att

de-tektion av enskilda cellkluster kan sänka dede-tektionsgränsen och dede-tektionstiden för systemet.

Inom en snar framtid kommer en klinisk studie påbörjas där man utvärderar de fem vanligaste

bakterierna som är associerade med sepsis i ett liknande system. För att systemet ska vara

en-kelt att använda utvecklades även ett program med grafiskt användargränssnitt i detta projekt.

Programmet analyserar automatiskt data som genereras från CellDirector 3D och hittar den

bakteriehämmande antibiotikakoncentrationen.

Examensarbete 30 hp

T

ABLE OF

C

ONTENT

Abbrevations

6

Background

7

Impact of antimicrobial drugs

7

Fitness cost of antibiotic resistance

7

Standard antibiotic susceptibility testing methods

7

Sepsis and treatment

8

Clinical laboratory routines

8

Novel antibiotic susceptibility testing methods

10

A microfluidic assay for antibiotic susceptibility testing

10

Detection principles in the microfluidic assay

11

Image analysis

12

Aim of this study

13

Material and methods

13

Bacterial strains

13

Etest assay

13

CellDirector 3D assay preparation

14

oCelloScope preparation

14

Intensity analysis

15

Blob analysis based on Otsu’s thresholding

15

Blob analysis based on entropy thresholding

15

Results

15

Choice of imaging software

15

Software development for intensity analysis

17

Determination of the antibiotic concentration range

17

Detection limits using darkfield imaging and intensity analysis

17

Detection limit of blob analysis using cluster-based thresholding

20

Detection limit of blob analysis using entropy-based thresholding

23

Single cell detection for shorter detection times

25

Discussion

25

The automatic software

25

Issues of intensity analysis

25

Improvements of the cluster-based analysis algorithm

26

Improvements of the entropy-based analysis algorithm

27

Lowest inoculum size detection limit

29

Evaluation of the shortest time to detection

30

Conclusions

30

Acknowledgements

31

References

31

A

BBREVATIONS

ATCC

American Type Culture Collection

AST

Antibiotic Susceptibility Testing

CFU

Colony-Forming

Units

CLI

Command Line Interface

GUI

Graphical User Interface

EUCAST

European Committee on Antimicrobial Susceptibility Testing

ICU

Intensive Care Unit

MALDI-TOF MS

Matrix-Assisted Laser Desorption/Ionisation Time-Of-Flight Mass

Spectrometry

MIC

Minimum Inhibitory Concentration

ROI

Regions Of Interest

SIR

Susceptible, Intermediate, Resistant

B

ACKGROUND

Impact of antimicrobial drugs

The discovery and use of antimicrobial drugs has saved countless lives and together with vaccines

they are one of the most important factors behind the increase in life expectancy during the last

hundred years. Before antibiotics were discovered in the early 1900, the crude infectious disease

mortality rate in the United States was 797 deaths per 100 000 persons, and in 1995 the mortality

rate had dropped to 63 deaths per 100 000 persons (1). Without antibiotics, advancements in many

medical fields, including surgery, transplantation, cancer treatment, and neonatal care would not

have been possible (2).

Fitness cost of antibiotic resistance

The increasing prevalence of antibiotic resistant bacteria makes it more difficult for clinicians to

initially prescribe appropriate antibiotics. In many cases broad-spectrum antibiotics are prescribed

to minimise the risk of treatment failure, but this practice puts a strong selective pressure on bacteria

to evolve resistance to these antibiotics. In an environment free from antibiotics, resistant bacteria

tend to have a lower fitness compared to susceptible bacteria (3), so the susceptible bacteria will

eventually outcompete the resistant bacteria. However, when exposed to antibiotics, the growth of

susceptible bacteria is inhibited, leading to a rapid enrichment of resistant bacteria. All antibiotic

use provides a selection pressure that benefits resistant bacteria.

Standard antibiotic susceptibility testing methods

The most widely used antibiotic susceptibility testing (AST) methods are phenotypic, detecting

bacterial growth in the presence of antibiotics. These methods were the first to be introduced, and

they have not changed significantly since their introduction with the exception for automation. The

broth microdilution test investigates bacterial growth in liquid media containing predefined serially

diluted antibiotic concentrations (fig. 1A). The antibiotic concentration that inhibits visible bacterial

growth after approximately 24 hours is determined as the minimum inhibitory concentration

(MIC). This method is considered the gold standard of susceptibility testing and the breakpoints

formulated by the European Committee on Antimicrobial Susceptibility Testing (EUCAST) to

determine if a strain is susceptible, intermediate or resistant to a certain antibiotic, are based on

broth microdilution. Etest® (bioMérieux) is a different AST method that also is calibrated against

the MIC obtained from broth dilution. Etest is a strip with a predefined exponential antibiotic

Figure 1. Standard antibiotic susceptibility testing methods. (A) In the broth dilution assay the MIC value is determined by investigating the lowest antibiotic concentration that inhibits visible bacterial growth in liquid media. (B) A MIC value cannot be obtained in the disk diffusion test, instead the S, I, R value is given by measuring the zone of inhibition around the antibiotic patch. (C) Etest is a strip with predefined exponential antibiotic concentration gradient. The MIC value is determined as the lowest antibiotic concentration that inhibits visible bacterial growth.

B. Disk diffusion (S, I, R)

16-20 hours

C. Etest (MIC)

16-20 hours

A. Broth dilution (MIC)

24 hours

concentration gradient, and a printed scale for easy readout (fig. 1C). The lowest antibiotic

concentration that inhibits visible bacterial growth (or less than 80% for some antibiotics) can be

read off the strip after 16-20 hours of incubation. The MIC value can be compared to EUCAST

breakpoints to determine antibiotic susceptibility.

The disk diffusion test is the most widely used AST method in clinical settings. An

antibiotic-impregnated patch is placed on an agar plate that has been swabbed with bacteria. The antibiotic

diffuses into the agar plate and creates an antibiotic gradient around the disk. After 16-20 hours of

incubation, a visible zone of growth inhibition can be detected and the size of the zone diameter

is measured and compared with EUCAST breakpoints to determine is the strain is susceptible,

intermediate, or resistant (SIR) to that particular antibiotic (fig. 1B). It is important to note that

disk diffusion only determines susceptibility; it does not return a MIC value. A rapid disk diffusion

test can be carried out in a similar manner, except that the test is evaluated after 4-6 hours to give

a preliminary indication of resistance profile.

Broth microdilution, Etest, and disk diffusion are highly standardised and sensitive AST methods

for detecting antibiotic resistance and, in addition, they are recommended in international guidelines

published by the Clinical and Laboratory Standards Institute (CLSI) in the US and EUCAST in the

EU (4). The limitations of these methods include the high number of viable bacteria required, the

need of overnight incubation and that pure cultures normally are required.

Sepsis and treatment

Sepsis, also known as blood poisoning, is an acute condition caused by the innate immune response

to bacteria and bacterial toxins entering the blood stream. When untreated, sepsis can progress

to severe sepsis and finally septic shock, a life-threatening condition. Symptoms such as elevated

heart rate, increased respiratory rate, and impaired blood flow are commonly associated with septic

shock. Septic patients are usually treated in the intensive care unit (ICU) with intravenous antibiotics

and fluids to maintain normal blood oxygen levels and organ functions (5). In a clinical study by

Kumar et al, 2006, it was shown that the survival rate decreases by 7.6% for every hour that

treatment is delayed for patients with septic shock (6). The Surviving Sepsis Campaign (SSC) is a

collaboration between the Society of Critical Care Medicine and the European Society of Intensive

Care Medicine to reduce the mortality of sepsis. The SSC have estimated that 18 million people

are affected by sepsis every year, and economically the cost of treating an ICU patient with sepsis

is six times higher than treating an ICU patient without sepsis (5). Delays in effective treatment are

associated with prolonged hospital stay of 2 days (7).

Clinical laboratory routines

The standard routine for handling patient samples varies between different hospital laboratories

and countries, but this section describes the routines based on Uppsala University Hospital. When

blood samples from sepsis patients arrive at the clinical laboratory, the number of bacterial

colony-forming unit (cfu) per ml is normally very low or even undetectable, approximately 1-100 cfu/ml

in adults (8). An initial incubation step is therefore required before bacterial typing and subsequent

traditional phenotypic tests can be carried out. The sample is incubated with nutrient rich media in

the automated microbial detection system BacT/ALERT® 3D (bioMérieux). In the BacT/ALERT

system, bacterial growth is detected by a colorimetric indicator at the bottom of the nutrient bottle

that changes color when reacting with CO

₂

, a metabolic product from growing microorganisms.

cultivate the pathogen over night (fig. 2). Identification of the bacterial species can be made by

observation of the colony morphology after overnight incubation. A disk diffusion test is also

made, and for Gram-negative strains a rapid disk diffusion test is made as well. The rapid disk

diffusion test is preliminary evaluated after 4-6 hours to get an early indication of any resistance.

The inhibition zone diameter from the disk diffusion tests are measured and compared to EUCAST

breakpoints for determination of SIR to specific antibiotics. For some bacterial species traditional

fermentation test and biochemical tests can also be made for verification. Once the bacterial species

is known, an Etest is carried out for certain strains and antibiotics where a specific MIC value is

clinically important.

Species identification and determination of resistance profile normally require 3-5 days. While

waiting for the results, the treating physician is forced to rely on empirical information such as the

patient’s response to the drug, but the rising prevalence of resistant bacteria increases the risk of

choosing an ineffective treatment. Especially in cases of septic shock, the condition of the patient

can deteriorate quickly and it might be too late when the laboratory results arrive. Due to this

situation there is a need for new AST tools to handle the increasing antibiotic resistance, improve

patient care and reduce healthcare costs. In recognition of this need it was recently announced that

the Longitude Prize offers a £10 million reward to those who can develop a rapid point-of-care test

for bacterial infections that is cheap, accurate, and easy-to-use.

Figure 2. Routine of positive blood samples at clinical laboratories. The positive samples are first Gram-stained to determine if it is a Gram-positive or Gram-negative pathogen, and then plated on selective agar plates. A disk diffusion test is made, and for Gram-negative strains the zones of inhibition are preliminary measured already after 4-6 hours of incubation. The zones are measured again after 16-20 hours of incubation and compared to EUCAST breakpoints to determine SIR. The colony morphology on the selective agar plates are evaluated after 24-48 hours of incubation for identification of the pathogen. To determine the MIC value for a certain antibiotic, a colony is resuspended in liquid media and swabbed on a plate. An Etest is applied, and after 16-20 hours of incubation the MIC value can be obtained. Approximately 3-5 days are required before species identification and antibiotic susceptibility profile is obtained. However, recent technological advances such as the MALDI-TOF MS enables the bacterial species to be identified much earlier and the method is currently in evaluation at the clinical laboratory at Uppsala University Hospital.

Recent technological advances allow the positive blood samples to be directly analysed using

matrix-assisted laser desorption/ionisation time-of-flight mass spectrometry (MALDI-TOF MS)

to identify the species. This instrument is highly sensitive and specific, and is currently under

evaluation for implementation at the Uppsala University Hospital. Many clinical laboratories in

the US and Europe have implemented the automated system Vitek®2 (bioMérieux) for bacterial

identification and AST. The Vitek 2 system uses a 64-well plastic card containing 41 fluorescent

biochemical tests for AST and species identification. Results can be obtained within 5 hours, but

the system is not 100% accurate so it is not employed in clinical hospital laboratories in Sweden (9).

Novel antibiotic susceptibility testing methods

The use of rapid susceptibility tests have been shown to decrease the mortality rate and

hospitalisation costs (10). PCR-based methods and microarray methods are both genotypic analyses

that can detect the presence of resistance genes, the former using site-specific amplification and

the latter complementary oligonucleotide hybridisation. Both are highly sensitive methods that can

give an indication of the antibiotic resistance profile within a short time frame, but they cannot

exclude resistance to a certain antibiotic, provide detection of novel resistance mechanisms, detect

phenotypic resistance, or give a MIC value.

Advances in DNA sequencing have enabled whole genome sequencing (WGS) of bacterial

genomes, but similar to PCR-based and microarray methods, WGS can only detect the presence of

resistance genes. It cannot detect uncharacterised resistance mechanisms or give any estimates of

the level of resistance gene expression. Our understanding of the genotype-phenotype relationship

is insufficient to be able to predict how several resistance products interact to display one level of

phenotypic resistance. However, WGS would be very suitable for tracking outbreaks of clinically

important strains such as ESBL producing bacteria and MRSA in hospitals.

Other emerging new technologies that could potentially be used for AST and especially detection

of phenotypic resistance include flow cytometry, microbial cell weighing by a vibrating cantilever,

and microfluidics. In all future phenotypic methods the requirement of measuring bacterial growth

is the limiting factor for decreasing time to detection. MALDI-TOF is already implemented in many

routine laboratories for identification of species, but it is also under development for genotypic or

phenotypic AST, by either finding fingerprints of resistance gene products, or using the sensitivity

of MALDI-TOF to early detect growth in the presence of a test panel of antibiotics (11)(12).

One of the limitations of current AST methods is the large number of viable organisms required,

making a time-consuming incubation step necessary. Next-generation AST methods should aim

to provide alternatives to current methodologies that are rapid, do not require pre-incubation, and

can be highly automated.

A microfluidic assay for antibiotic susceptibility testing

greyscale intensity (fig. 3D) change is analysed over time and along the antibiotic concentration

gradient to determine MIC (fig. 3E). In a recent study performed in this laboratory, it was

demonstrated that CellDirector 3D could be used to rapidly and accurately determine the MIC

value directly from clinical S. aureus positive blood cultures (manuscript in preparation). Compared

to standard phenotypic AST methods, microfluidic devices for AST require smaller samples and

reagent volumes.

Detection principles in the microfluidic assay

Dark field microscopy is a very simple and useful technique for detection of unstained biological

Figure 3. Principles of the microfluidic assay for AST. (A) The microfluidic device developed by Gradientech AB (Uppsala, Sweden). (B) A chamber is filled with bacteria in an agarose mix. Two microfluidic channels, one containing antibiotic and the other medium with antibiotic, pass by the chamber. Due to diffusion, the fluids in the channels will diffuse into the chamber and create a linear antibiotic gradient in the chamber. (C) A camera is connected to capture time-lapse images of the bacterial growth in the assay under 5 hours. (D-E) Bacterial growth can be measured by analysing the change in intensity during the experiment.

Antibiotic concentration Medium Medium + Antibiotics

1 h

3 h

5 h

1 h

3 h

5 h

A B C

Antibiotic conc. (µg/ml) Antibiotic conc. (µg/ml) Antibiotic conc. (µg/ml) Rate of intensity change Rate of intensity change Rate of intensity change

samples that would otherwise be difficult to distinguish in bright field microscopy. Transmitted light

is blocked from the lens, and only light scattered when passing the sample is captured. An empty

field of view will appear dark, hence the name dark field microscopy. Cell growth in CellDirector

3D will appear as bright spots when dark field microscopy is used. So far, only intensity analysis

of images captured in a dark field microscope has been used to extract a MIC-value from the

CellDirector 3D assay. The disadvantage of dark field microscopy is the amount of illumination

needed to obtain a signal and, in addition, the system is not sensitive enough to capture single cells

in the assay.

A novel instrument known as oCelloScope (Philips BioCell A/S) uses a unique technology that

includes a tilted focus plane and scanning of a series of images to obtain highly detailed 3D images.

The oCelloScope enables tracking of cell clusters much earlier than to dark field microscopy. Using

the oCelloScope for detection of cell growth in CellDirector 3D, the time to detection could be

decreased.

Image analysis

Digital image processing can be applied to extract meaningful information from images. One of the

easiest ways of using image analysis to generate MIC data from the dark field microscopy images

is intensity analysis, where the average darkness of each column of the image is calculated. The

advantage of using intensity analysis is that it is a fairly simple algorithm, but there is a wide range

of image analysis tools available that also could be used for detection of cell growth. Thresholding

is a basic image segmentation tool to separate foreground from background, if there is a distinction

(14). Otsu’s thresholding method is a clustering-based method that calculates the optimal separation

based on maximal variance between the two classes, while Li and Yen’s thresholding methods are

based on entropy, a statistical measurement of randomness in the image (15). The output from a

thresholded image is a binary image, where each pixel either has the value 0 or 1. Depending on

which thresholding method that is applied the output will be different (fig. 4). Binary images can be

labeled to obtain descriptive properties for each region of interest (ROI) in the image.

Figure 4. Different image analysis tools, different result. Three different thresholding methods (Li’s, Otsu’s and Yen’s) are applied to the same original image. The result from thresholding is a binary image, where the image is divided into two groups, 0 (black, background) and 1 (white, object). The image analysis morphological operation erosion is then applied to remove areas that are too small. To compensate for the size reduction, the morphological operation dilation can be used on the remaining regions. Each region in the image can then be labelled and characteristics such as area, shape, and cavity can be computed.

Li’s method

Otsu’s method

Yen’s method

Morphological operations are basic logical algorithms performed on binary images that can be

applied to remove internal cavities, connect two regions, or remove small regions. A structuring

element that can have any shape such as a diamond, square, or disk, moves over each pixel in the

image. In morphological dilation, if the centre of the structuring element is located within a region

(white), the entire surface that the structuring element covers is expanded to become a part of that

region. Morphological erosion is the opposite of morphological dilation; it reduces a region instead

of expanding it. (16)

Image processing tools are available in different formats. Softwares like ImageJ, ICY, and

CellProfiler are all centred on an interactive graphical user interface (GUI), making them very

easy to use for simpler image processing applications. For development of more complex image

analysis algorithms, command line interfaces (CLI) like Image Processing Toolbox™ (MATLAB®),

OpenCV (Python), or Scikit-image (Python) are more suitable. However, using the latter requires

knowledge about programming and a deeper understanding of image analysis.

Aim of this study

To address the increasing problem with antibiotic resistance, Gradientech AB is developing a

multichannel assay, QuickMIC™. While the CellDirector 3D assay can only run one sample and

one antibiotic at a time, the QuickMIC assay will have 8 parallel microfluidic chambers. Manual

analysis of the time-lapse images is very time consuming so for the assay to be efficient in a clinical

setting, there is a need for an automated readout system. In this study, different software packages

were evaluated for the development of a software that could automatically perform the analysis.

The QuickMIC system will soon be evaluated in a clinical setting, studying the five most common

pathogens associated with sepsis. In previous studies made on CellDirector 3D, the standard

inoculum has been 10

6

_{cfu/ml or higher, but the bacterial concentration in the blood of septic}

patients is always much lower (1-100 cfu/ml). Therefore, this study also investigates the lowest

detection limit of bacterial inoculum size for CellDirector 3D, using the four pathogens that will be

used in QuickMIC study. Different image analysis approaches were also developed and evaluated

for improvement of the detection limit. Using the precise detection system oCelloScope, single-cell

images are generated to investigate the time to detection and if it could be decreased.

M

ATERIAL AND METHODS

Bacterial strains

The bacterial strains used in this study were E. coli (ATCC 25922), Klebsiella pneumoniae (ATCC

29665), Pseudomonas aeruginosa (ATCC 27853), and S. aureus (ATCC 29213). The American Type

Culture Collection (ATCC) is a nonprofit organisation that characterises and distributes standard

reference microorganisms. Gram-negative strains E. coli, K. pneumoniae, and P. aeruginosa were grown

on Müller-Hinton II (MHII) agar plates, while blood agar plates were used for the Gram-positive

strain S. aureus. All strains were incubated in a 37°C incubator room.

Etest assay

A bacterial colony was suspended in saline to 0.08-0.1 of OD

₆₀₀

, which corresponds to 0.5

McFarland standard (~10

8

_{cfu/ml). The McFarland scale is a standard reference of visual turbidity}

surface of a MHII agar was swabbed three times using three different angles by rotating the plate

approximately 120° each time. When the plate had dried, an Etest was carefully placed on the

agar using tweezers. After 16-20 hours incubation, the MIC value could be read from the strip.

Duplicate Etest were made for each strain.

CellDirector 3D assay preparation

A blister package containing CellDirector 3D (Gradientech AB) was put in a vacuum chamber for

30 minutes. Vancomycin was used for MIC testing of Gram-positive S. aureus, while ciprofloxacin

was used for MIC-testing of the Gram-negative strains E. coli, K. pneumoniae, and P. aeruginosa.

Stock solutions of antibiotics were made by dissolving 10 mg of the antibiotic compound in either

1 ml PBS (vancomycin) or 0.1 M HCl (ciprofloxacin). The stock solution was diluted to a final

concentration that was chosen based on results from Etest. To get a good antibiotic concentration

range in the assay, the maximum assay concentration was set so that the MIC value of the strain

would be in the middle of the assay chamber.

A bacterial colony was suspended in MHII to 0.08-0.1 of OD

₆₀₀

(0.5 McFarland). The suspension

was diluted either 5, 50, 500, or 5000 times in MHII-broth and then mixed 1:1 with 0.5 % TopVision

low melting point agarose (Thermo Fisher Scientific) to reach an estimated concentration of 10

6

_,

10

5

_{, 10}

4

_{, or 10}

3

_{cfu /ml. The original suspension was diluted 1:10 000, 1:100 000, and 1:1 000 000}

and plated on MHII agar plates to make a viable cell count after overnight incubation. A volume

of 8-8.5 µl of the bacteria-agarose mix was injected into the chamber of the CellDirector 3D using

reverse pipetting. The mixture was allowed to solidify in the fridge for 5-10 minutes.

A syringe was filled with 1 ml MHII medium, while a second syringe was filled with 1 ml MHII

medium containing antibiotics of the specified concentration. In a 37°C incubator room, the

syringes were fitted into a syringe pump (Harvard Apparatus Syringe Infusion Pump Model 22)

with a flow rate set to 2.0 µl/min. The media tube ends were connected drop to drop to the inlets

of the CellDirector 3D and the assay was placed upside down in an up-right microscope using

a 2x objective (Nikon Optiphot-2, Tokyo, Japan). A dark field condenser was used for capturing

dark field images. The syringe pump was started and the assay was carefully monitored in the

microscope. When the microfluidic channels of the assay were filled with media from the syringes,

a vacuum pump (SP 104 SA, Schwarzer Precision) was connected, and the outlet tube was removed.

A camera (Canon EOS 700D) connected to the microscope was set to capture images (2304×3456

pixels) of the assay every 10 minutes (shutter speed: 1/125 s). The experimental time was set to 5

hours, producing a set of 30 time-lapse images, showing the bacterial response towards antibiotic

over time. The images were automatically saved in jpeg-format. With the exception of E. coli, all

experiments were run in at least duplicates.

oCelloScope preparation

A CellDirector 3D assay was prepared according to normal protocol, but instead of being

placed under a microscope it was placed in the oCelloScope to generate high-quality single-cell

3D microscope images. The built-in software (UniExplorer 4.1.1) was set to capture images

every 10 minutes for 5 hours, creating a data set of 30 images in total. Since the sensor in the

oCelloScope is too small to capture the entire assay, the system captured 110 images that were

merged in UniExplorer to produce z-slice images of the entire assay. The merged images were

exported as bmp-files. The images were manually cropped in imaging software ICY to a size of

approximately 6750×2500 pixels. Single experiments of S. aureus and P. aeruginosa were conducted

Intensity analysis

Time-lapse images were converted to 8-bit gray-scale images, giving each pixel a value between 0

and 255. The mean of each column was computed, converting each image into a vector describing

the mean intensity along the gradient axis. The rate of intensity change was computed using second

order central difference describing the change in cell growth along the gradient.

∂f / ∂x = f (x + 1/2h) - f (x-1/2h)

{1}

The mean of the intensity change of the final five time points was plotted against antibiotic

concentration. The MIC-value was obtained where the observed growth rate was less than 0.01

intensity units (fig. 6A).

Blob analysis based on Otsu’s thresholding

The last image of the time-lapse dataset was used as template in the blob analysis to determine

the ROI and its boundary coordinates. Otsu’s thresholding was used to create a binary image. The

resulting regions were ROI and labelled to give each region a unique tag. The area and coordinates

for each ROI was computed and saved. The process of binarising and labelling ROIs was repeated

for the remaining images in the dataset. The area for each ROI was computed, which resulted in

a matrix that described how the area changes (colony growth) over time. The rate of change for

each ROI was computed using second order central differences {1} and the mean of the final five

time points was computed. A Savitzky-Golay smoothing filter was applied to reduce noise in the

output signal and to produce a growth rate curve. The boundary coordinates were translated into

concentration coordinates and the change in area size was plotted against concentration.

Blob analysis based on entropy thresholding

Images were compressed into entropy images before a blob analysis was performed. Based on the

entropy compression of the last image in a set of time-lapse images, a threshold value was chosen

manually and the remaining images were converted to binary images using this threshold value. The

blob analysis was then carried out as described in the blob analysis based on Otsu’s thresholding.

R

ESULTS

Choice of imaging software

Figure 5. Automatic image analysis software for CellDirector 3D. The software was built using the Python programming language and various libraries such as numpy, Pandas, scikit-image. The GUI was built using the application platform PyQt and guidata, which is based on the Qt platform. (A) The software consists of a number of widgets (QMainWindow, QPushButton, QSlider, FigureCanvas, DataSetEditGroupBox, DataSetEditShowGroupBox) and layout (QVBoxLayout, QSplitter, QGroupBox) objects. (B) Images loaded into the software for analysis can be viewed, cropped, and rotated before analysis. Different input paramters such as sampling interval, concentration range, and MIC threshold can be modified by the user. When the ‘calculate’-button is pressed by the user, the software will analyse the image data based on intensity and a plot showing the intensity change will be displayed.

A

B

QuickMIC data analysis

Input parameters

Lowest conc. (µg/ml) Highest conc. (µg/ml) Sampling interval (min) Savitzky-Golay window Rotaiton angle (degrees) MIC threshold Output parameters Images Time (s) Width Height MIC: Actions Read Images

Crop Growth Rate

Software development for intensity analysis

The QMainWindow class from Qt provided a main application window for the GUI of the software.

The in- and out parameter boxes were created from guidata elements DataSetEditGroupBox and

DataSetShowGroupBox respectively. Seven buttons (’Read Images’, ’Invert Image’, ’Crop’, ’Growth

Curve’, ’First Derivate’, ’Growth Rate’, and ’Calculate’) were added using the Qt QPushButton

class and positioned in a QGroupBox Widget. A QWidget containing a FigureCanvas object from

matplotlib and a QSlider from Qt was placed below the parameter box, for visualisation of loaded

images. Another QWidget containing a FigureCanvas object only was added for the display of

plots (fig. 5).

The image dataset, chosen by the user, is loaded into the software as 8-bit gray scale images (256

shades of gray), using the io and img_as_ubyte modules from the scikit-image library. The images

are displayed by the imshow-function from matplotlib, to allow the user to rotate or crop the

image if needed. In the upcoming multichannel assay, the camera will capture an area larger than

the chamber in the assay, so the images will need to be cropped before analysis. The user is further

allowed to change certain parameters, such as the antibiotic concentration range, threshold value

for defining the MIC, the sampling time interval, and the number of data points the smoothing

algorithm should take into account. When “Calculate” is pressed by the user, the program will

perform the intensity analysis and display a plot of the rate of intensity change against antibiotic

concentration (fig. 6B). The computed MIC-value is displayed in the output parameters-box, and a

red arrow indicates the position of the MIC in the growth curve. Accuracy of the semi-automatic

analysis algorithm was analysed using data from the previous clinical study. By comparison of the

MIC values obtained from manual analysis, it was concluded that the automatic analysis determined

slightly higher MIC values than the manual analysis but both produced MIC within the accepted

error range from the Etest results (fig. 6C, S1).

Determination of the antibiotic concentration range

The antibiotic concentration range used in CellDirector 3D in this study was chosen based on

the result from Etest (tab. 1). For S. aureus the MIC when using vancomycin was determined to

be 1.5 µg/ml and the antibiotic gradient range was set to 0-4 µg/ml. For P. aeruginosa the MIC

of ciprofloxacin was determined to be 0.25 µg/ml and the antibiotic gradient range was set to

0-0.5 µg/ml. The MIC of K. pneumoniae in combination with ciprofloxacin was 0.008 µg/ml and

the antibiotic gradient range was set 0-0.016 µg/ml. For E. coli the MIC of ciprofloxacin was

determined to be 0.012 µg/ml and the antibiotic gradient range was set to 0-0.025 µg/ml.

Detection limits using darkfield imaging and intensity analysis

Different inoculum sizes of S. aureus and P. aeruginosa were loaded into CellDirector 3D. Intensity

analysis and a viable cell count determined the optimal detection limit was to be ~ 4·10

5

_{cfu/ml, but}

the lowest detection limit was determined to be ~ 4·10

4

_{cfu/ml (fig. 7). Cell growth was detected}

Table 1. Determination of antibiotic concentration ranges used in the microfluidic assay.

Antibiotics vancomycin ciprofloxacin ciprofloxacin ciprofloxacin Concentration range low-high (µg/ml) 0 - 4 0 - 0.5 0 - 0.016 0 - 0.025 MIC from Etest

Figure 6. An overview of the automatic data analysis software. (A) The intensity in each image is analysed based on the mean intensity along the x-axis. The change in intensity was computed using second order central difference. The growth curve is obtained by computing the mean of the five final time points, and a Savitzky-Golay filter is applied to reduce noise. (B) The logical structure of the automatic data analysis software. The light orange boxes are event that occur without user interferance. (C) Comparision of the MIC value obtained by manual (green) and automatic (red) analysis. The result was compared with the MIC determined by Etest (blue) and the accepted error range of Etest (light blue). (VSSA = susceptible S. aureus, hVISA = heterogeneous vancomycin-intermediate S. aureus, VISA = vancomycin-vancomycin-intermediate S. aureus).

Start

Choose dataset

Read images as 8-bit gray scale images

View image dataset Update output parameters Compute calculations Rotate

images imagesCrop

Update input parameters

Plot growth rate

vs time vs concentrationPlot growth rate Plot intensity

vs time

Save plot

C

VSSA 1 VSSA 2 VSSA 3 hVISA 1 hVISA 2 hVISA 3 VISA 1 VISA 2

Figure 7. Evaluation of different inoculum sizes using S. aureus in combination with vancomycin. The microfluidic assay was loaded using inoculum sizes 5·103 _{(A), 4·10}4 _{(B), 5·10}5 _{(C), and 4·10}6 _{(D) cfu/ml and the antibiotic gradient}

was 0-4 µg/ml. Images of the chamber were captured every 10 minutes during 5 hours, producing an image dataset of 30 images. An image of the chamber at 300 min is displayed for each inoculum size (upper) in this figure. To determine the MIC value, the images were analysed based on average intensity along the antibiotic gradient (lower).

S. aureus (4.5 . ₁₀3 _{cfu/ml) - vancomycin}

Growth curve (time = 300 min)

Antibiotic concentration (µg/ml) 0.25 0.30 0.20 0.15 0.10 0.00 -0.05 -0.10 0.05 A

verage change in gray scale intensity

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

MIC = 0.13 µg/ml

S. aureus (4.3 . ₁₀4 _{cfu/ml) - vancomycin}

Growth curve (time = 300 min)

2.5 3.0 2.0 1.5 1.0 0.0 0.5 A

verage change in gray scale intensity

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

MIC = 0.88 µg/ml

Antibiotic concentration (µg/ml)

S. aureus (5.0 . ₁₀5 _{cfu/ml) - vancomycin}

Growth curve (t<ime = 300 min)

8 10 6 4 2 0 A

verage change in gray scale intensity

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

MIC = 1.30 µg/ml

Antibiotic concentration (µg/ml)

S. aureus (4.3 . ₁₀6 _{cfu/ml) - vancomycin}

Growth curve (time = 300 min)

6 7 5 4 3 1 0 -1 2 A

verage change in gray scale intensity

when using an even lower inoculum size of ~ 5·10

3

_{cfu/ml, but the growth curve produced by the}

intensity analysis was not quantifiable. The MIC determined in the CellDirector 3D was comparable

to the MIC determined by Etest (fig. 8), and the optimal inoculum size 10

5

_{cfu/ml was confirmed}

for E. coli and K. pneumoniae.

Detection limit of blob analysis using cluster-based thresholding

To decrease the detection limit in CellDirector 3D, a blob analysis approach that traced growth

within cell clusters rather than average intensity was developed (fig. 9). Otsu’s thresholding method

was used to obtain a threshold value for separation of foreground (bright pixels) from background

(dark pixels). The threshold value was applied to all images for the creation of binary images

and the last image in a dataset was used as template to find ROI. Each region was given a unique

label, which enabled computation of properties such as area and bounding box to be measured

for each region. The morphological image operations erosion and dilation were initially added

to reduce noise, but these operations were later replaced with a less computationally heavy filter

that discarded ROI with an area below 100 pixels and above 100 000 pixels. The coordinates of

the bounding box of each ROI in the template were computed in order to track cell growth in

the specified ROI over time. All remaining images from previous time points in the dataset were

labelled and the area of each ROI determined by the template was calculated. The change in cluster

Figure 8. Lowest inloculum size using different strains. Different inoculum sizes of S. aureus and P. aeruginosa were evaluated in the microfluidic assay and the results were compared to the MIC value determined by Etest (dashed line). The accepted error range of Etest is one dilution step above and below the determined MIC (shaded area). The inoculum size determined as the optimal inoculum size was also evaluated for E. coli and K. pneumoniae.

104 ₁₀5 ₁₀6 ₁₀7 0.05 104 ₁₀5 ₁₀6 ₁₀7 104 ₁₀5 ₁₀6 ₁₀8 103 ₁₀7 104 ₁₀5 ₁₀6 ₁₀8 103 ₁₀7 0.04 0.03 0.02 0.01 0.00 0.05 0.04 0.03 0.02 0.01 0.00 1.0 0.8 0.6 0.4 0.2 0.0 4.0 3.0 2.0 1.0 0.0 3.5 2.5 1.5 0.5

S. aureus - vancomycin E. coli - ciprofloxacin

K. pneumoniae - ciprofloxacin P. aeruginosa - ciprofloxacin

Etest: 1.5 µg/ml

Estimated MIC (µg/ml) Estimated MIC (µg/ml)

Estimated MIC (µg/ml)

Estimated MIC (µg/ml)

Inoculum size (cfu/ml) Inoculum size (cfu/ml)

Inoculum size (cfu/ml) Inoculum size (cfu/ml)

Etest: 0.012 µg/ml

size over time was computed using second order central difference. The bounding box coordinates

could easily be converted from x-axis position to corresponding antibiotic concentration in the

assay. A Savitsky-Golay filter was applied to produce a trend line, and the MIC value was obtained

where the change of area was less than 0.1 units.

The performance of the cluster-based thresholding blob analysis described above, was evaluated

using the data generated from the investigation of the lowest detection limit by intensity analysis.

Using the cluster-based thresholding blob analysis, the optimal inoculum size was determined to

be ~ 5·10

4

_{cfu/ml (fig. 10). The algorithm could detect cell clusters when the assay was loaded}

with 10

3

_{cfu/ml as well, but the number of data points and the signal-to-noise ratio was too low}

for a growth curve to be produced for the dataset. This blob analysis method could also produce a

seemingly true growth curve for 10

5

_{cfu/ml, despite the fact that the cell clusters detected actually}

were the background growth occurring out of focus behind the cells in the focal plane. A higher

Figure 9. Schematic figure of the blob analysis approach. A blob analysis was developed to investigate if it could lower the detection limit in the microfluidic assay. Images were converted to binary images by thresholding. The regions in the binary images were labelled to determine ROI. Based on properties such as area and coordinate that could be computed from each ROI, a plot showing the change in blob size along the antibiotic gradient axis could be produced. Image dataset Binary images ROI Area Template image Threshold value Boundary coordinates Area Binary images ROI Growth curve Smoothing

Use last image to find ROI Threshold Label Compute Filter: Remove area < 100

area > 100 000 _{ROI coordinates}

Calculate change in area size

Savitzky-Golay filter Rate of change in area

Rate of change in area

Label Threshold

Figure 10. Data analysis using cluster-based blob analysis. Using image datasets of S. aureus and vancomycin generated during the investigation of the lowest detection limit, the blob analysis was evaluated for different inoculum sizes 5·103 _{(A), 4·10}4 _{(B), 5·10}5 _{(C), and 4·10}6 _{(D) cfu/ml. The inverted image of the chamber at 300 min is}

displayed here (left) with the ROIs were automatically detected in the blob analysis (red boxes). In the corresponding growth curve (right), the size change in each ROI is represented by a blue dot. A Savitzky-Golay filter was added to produce a trend line to the data (green line) and the suggested MIC value is represented by a red dot.

MIC = 2.38 µg/ml

MIC = 3.38 µg/ml

MIC = 3.34 µg/ml

MIC = 3.49 µg/ml S. aureus (4.5 . ₁₀3 _{cfu/ml) - vancomycin}

S. aureus (4.3 . ₁₀4 _{cfu/ml) - vancomycin}

S. aureus (5.0 . ₁₀5 _{cfu/ml) - vancomycin}

S. aureus (4.3 . ₁₀6 _{cfu/ml) - vancomycin}

Growth curve (mean of last five time points) ROI (time = 300 min) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) 0 20 30 40 50 10 A

verage change in blob size

0 100 150 200 50 A

verage change in blob size

0 100 150

50

A

verage change in blob size

0 100 150 200 50 A

verage change in blob size

A

B

C

inoculum size of 10

6

_{cfu/ml was not compatible with this analysis at all, since the dense growth}

meant that very few individual blobs could be detected. The remaining problem with this method,

independent of inoculum size, is to design an algorithm that can produce a robust trend line from

the data points.

Detection limit of blob analysis using entropy-based thresholding

Since cluster-based thresholding failed to generate a satisfying segmentation a different approach

was tried. The gray-scale images were compressed to entropy images, describing the level of

randomness in an image. The entropy algorithm used in this analysis was computed using a base 2

logarithm to return the minimum number of bits needed to encode the local intensity distribution

(17). Background pixels all have very little contrast so the entropy is low, but where cell growth

occurs there is a high contrast difference so the entropy is high. The images were binarised based on

a thresholding value manually set based on the entropy (fig. 11). Labelling of binary images defined

Figure 11. Schematic figure of the entropy blob analysis approach. TeXT. To increase the sensitivity of the blob analysis, images were converted into entropy images that highlight regions that have high difference in intensity. The images were further analysed in a similar manner as the previously described entropy analysis.

Images Binary images ROI Area Template Threshold value Boundary coordinates Area Binary images ROI Growth curve Smoothing Find ROI Threshold Label Compute Filter: Remove area < 50 area > 1 000 000 ROI coordinates Calculate change in area size Savitzky-Golay filter Entropy Label Threshold Compute

Rate of change in area

Rate of change in area

the ROIs and the area within each ROI was computed. All ROI with an area below 50 pixels or

above 1 000 000 pixels in the template image were considered noise and discarded. A growth rate

curve was produced in a manner similar to the cluster-based blob analysis. This entropy-approach

was also evaluated using S. aureus data produced from the investigation of the lowest detection

limit. The algorithm worked well for inoculum sizes of ~5·10

3

_-10

4

_{cfu/ml, while it was more}

difficult to separate individual cell clusters for ~5·10

5

_{cfu/ml and ~5·10}

6

_{cfu/ml (fig. 12).}

Figure 12. Data analysis using entropy-based blob analysis. The entropy-based blob analysis was evaluated using image datasets of S. aureus and vancomycin generated from the investigation of the lowest detection limit. Datasets of different inoculum sizes 5·103 _{(A), 4·10}4 _{(B), and 5·10}5 _{(C) cfu/ml, were analysed to determine if}

entropy-based blob analysis could lower the detection limit. The image captured after 300 min is displayed together with the ROI found by the algorithm marked as red boxes (left). A growth plot was generated (right) by plotting the size of each ROI along the concentration gradient axis (blue dot). A trend line was computed by adding a Savitzky-Golay filter (green line) and the MIC was determined (red dot).

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) -50 0 50 100 150 200 250 A

verage change in blob size

-100 100 200 300 400 500 600 A

verage change in blob size

0 -150 -50 0 50 100 150 200 A

verage change in blob size

-100 S. aureus (4.5 . ₁₀3 _{cfu/ml) - vancomycin}

S. aureus (4.3 . ₁₀4 _{cfu/ml) - vancomycin}

S. aureus (5.0 . ₁₀5 _{cfu/ml) - vancomycin}

Single cell detection for shorter detection times

Image data from the oCelloScope have higher resolution and therefore allow observation of small

cell clusters or even single cells, so cell growth can be detected at a much earlier time point compared

to dark field microscope data. The entropy-based blob analysis was tested on data generated in the

oCelloScope on S. aureus and P. aeruginosa, using inoculum sizes 2·10

5

_{cfu/ml and 4·10}

5

_cfu/ml

respectively. From the growth rate curve an accurate MIC value for S. aureus and P. aeruginosa was

obtained (fig. 13). An investigation of individual blobs over time showed that cell clusters could be

detected within the first hour of the experiment (fig. 14). A comparison of the time to detection

showed that the time to detection was significantly improved when datasets produced from the

oCelloScope were analysed based on entropy rather than intensity (fig. 15). A MIC value could

be determined for P. aeruginosa exposed to ciprofloxacin within 2 hours when using entropy based

analysis, while the intensity based analysis could not determine a value until after 3 hours. For

images produced using dark field microscopy, there was no significant improvement of the time to

detection by using the entropy analysis.

D

ISCUSSION

The automatic software

For the CellDirector 3D assay and the upcoming QuickMIC assay to be easy to use, an automatic

data analysis software is necessary. Depending on the size of the image dataset the data analysis

processing time varies, but for the datasets used in this study the analysis only took a few minutes.

Besides a faster analysis, the automatic analysis also allows the user to not have any deeper knowledge

in mathematics and reduce the possibility of calculation errors. The manual and automatic analysis

both determined MIC-values that were within the accepted error range of Etest, which is

one-fold dilution step. The manual analysis returned slightly higher MIC than the automatic analysis,

due to different methods being used to determine the MIC. The manual analysis determined the

MIC value by computing where the second-derivate of the growth rate curve crosses the x-axis,

while the automatic analysis determined the MIC based on when the growth rate was below a

threshold value. The second derivate was not applied to the automatic analysis since it was difficult

to determine the MIC due to oscillations in the second derivate curve.

Issues of intensity analysis

Improvements of the cluster-based analysis algorithm

The main problem with the cluster-based thresholding approach was the calculation of the final

trend line. Noise, particularly from the bacteria-agarose injection site and a scar left from assay

manufacturing, were too prominent to go undetected by the algorithm, causing a shift in the trend

Figure 13. Data generated in the oCelloScope. New image datasets were produced in the oCelloScope for S.

aureus in combinaiton with vancomycin and P. aeruginosa in combination with ciprofloxacin. The datasets were

analysed using the entropy-based blob detection that was developed in this study. The image captured of the chamber at 290 min is displayed in this figure together with the ROIs that were detected by the algorithm (red box). The corresponding growth curve are also shown in the figure (blue dot - blob size of each ROI, green line - trend line, red dot - estimated MIC).

S. aureus (2.3 .₁₀5 _{cfu/ml) - vancomycin (time=290 min)}

P. aeruginosa (4.5 .₁₀5 _{cfu/ml) - ciprofloxacin (time=290 min)}

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0.0 0.06 0.13 0.19 0.25 0.31 0.38 0.44 0.5 Antibiotic concentration (µg/ml) 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Antibiotic concentration (µg/ml) -100 0 100 200 300 400 500 A

verage change in blob size

0.0 0.06 0.13 0.19 0.25 0.31 0.38 0.44 0.5 Antibiotic concentration (µg/ml) 0 400 600 800 1000 1200 200 A

verage change in blob size

S. aureus growth curve P. aeruginosa growth curve

line compared to the actual bacterial growth in the assay. Compared to the entropy-based analysis,

this cluster-based approach tends to generate a larger number of ROIs, and many of them have

large areas. However, many of the ROIs identified are actually noise, and the low resulting

signal-to-noise ratio will give problems when calculating a trend line.

The smoothing filter Savitsky-Golay that was used in this study is based on local least-square

polynomial approximation. The method uses convolution to maintain the height and shape of peaks

while reducing noise. Generally, this method is good for smoothing data with a large frequency

span, but it is less successful in rejecting noise (18). Different smoothing algorithms could be

investigated to examine if an improved trend line could be generated. For example, adjacent cell

clusters can be grouped together using k-means clustering and a smoothing filter can be applied to

these bigger clusters to produce a trend line.

Improvements of the entropy-based analysis algorithm

Although the entropy-based thresholding algorithm gave more accurate and reliable results

than the cluster-based thresholding algorithm, the main disadvantage of using entropy-based

thresholding is the computational intensity of the algorithm, resulting in a long time required to

process the data. The time required for an entropy-based analysis of three images of 1000x1500

pixels ( approximately 2% of a dataset) was almost twice (21 sec) as long as the time cluster- and

intensity-based analysis required (11 sec and 8 sec respectively). A performance test showed that

the computational time increases exponentially with image size. On the other hand, the results

showed that it is possible to determine an MIC-value earlier using entropy-based analysis instead

of intensity-based analysis using image data generated in the oCelloScope. Using CUDA, a parallel

computing platform, or optimising the code in Cython could improve the computational time

and should be considered for future development of the algorithm. If entropy-based analysis

Figure 14. Single-cell detection. Using blob analysis, each ROI size can be tracked in time to determine when detection is possible.This ROI is from a S. aureus image dataset generated in the oCelloScope.

S. aureus (2.29 . ₁₀5 _{cfu/ml) - vancomycin}

Figure 15. Comparision of different analysis and detection methods. The time required until a MIC value could be determined was compared when analysing data using intensity analysis and entropy-based blob analysis. Images captured in dark field microscopy were also subjected to comparision with images generated in the oCelloScope.

0 50 100 150 200 250 300

Time (min)

0 50 100 150 200 250 300

Time (min) 0 50 100 Time (min)150 200 250 300

0 50 100 150 200 250 300

Time (min)

0 50 100 150 200 250 300

Time (min) 0 50 100 Time (min)150 200 250 300

0 50 100 150 200 250 300

Time (min) 0 50 100 Time (min)150 200 250 300 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0.0 0.5 0 1 2 3 4 5 Estimated MIC (µg/ml) 0 1 2 3 4 5 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 S. aur eus (4.3 . 10 4 cfu/ml)

- vancomycin (dark field micr

oscope) S. aur eus (5.0 . 10 5 cfu/ml)

- vancomycin (dark field micr

oscope) S. aur eus (2.3 . 10 5 cfu/ml) - vancomycin (oCel loScope) P. aeruginosa (4.5 . 10 5 cfu/ml) - cipr ofloxacin (oCelloScope)

Intensity analysis Entropy -based blob analysis

can return a MIC value within a significantly shorter time than intensity analysis, the additional

computational time required could be worth it.

An alternative to entropy-based thresholding would be to use a high-pass filter on the Fourier

transform of images. The Fourier transform is a tool in image analysis for decomposition of a

function (image) into the trigonometric basic functions sines and cosines, that plots the amplitudes

and phases of these cosines and sines against their respective frequencies. One of the many uses

of a Fourier transform is to separate the low frequency (background) and high frequency (objects)

components of an image. Since high contrasts in an image represent high frequencies in the

Fourier domain, a high-pass filter that filters away the low frequency components of the image

will retain these high contrast objects while removing variations in the background. It remains

to be seen whether a high-pass filter on the Fourier transform of an image before thresholding

can give an equally good image segmentation as entropy, and more importantly if it would be less

computationally intense.

A critical improvement of the blob analysis algorithms developed in this study is the detection of

cell clusters at earlier time points. The current algorithms determine the ROI based on cell clusters

found at the end of the experiment, thus after 5 hours. Cell clusters might not be as prominent and

would be more difficult to find earlier on, but exactly how difficult remains to be evaluated. In the

future it would also be interesting to investigate if image analysis can be applied to separate samples

containing two different species.

Lowest inoculum size detection limit

The lowest detection limit for CellDirector 3D varies depending on the image analysis algorithm

used. Intensity analysis has so far been the method to produce a growth curve and determine a

MIC value in the CellDirector assay. Out of the three image analysis techniques presented in this

report, intensity analysis has generated the most robust results (fig. 16). It could compute a growth

curve for inoculum sizes between 10

4

_{and 10}

6

_{cfu/ml, while entropy-based thresholding worked}

well for 10

3

_{- 10}

4

_{cfu/ml but not at all for 10}

5

_-10

6

_{cfu/ml. The cluster-based thresholding method}

could identify cell clusters for inoculum sizes of 10

3

_-10

4

_{cfu/ml, and it could produce surprisingly}

accurate growth curves for 10

5

_{cfu/ml even though it identified out-of-focus background growth}

rather than cells in the focal plane as ROIs. Although blob detection could be used with inoculum

sizes as low as 10

3

_{cfu/ml, the main problem lies with the low bacterial number loaded into the}

assay. At inoculum densities of 3·10

3

_{cfu/ml, as few as approximately 30 bacteria will be in the}

assay since only a volume of 8 µl is injected into the chamber. Cell growth could only be detected

Figure 16. Detection ranges of the different analysis methods. The lower detection limit of the intensity analysis is approximately 5·104 _{cfu/ml, while the upper limit is the highest inoculum size that has been analysed in this study,}

~5·106 _{cfu/ml. For the cluster-based blob detection the detection range lies between ~ 8·10}3 _{- 2·10}5 _{cfu/ml, while the}

entropy-based blob detection has a slight wider detection range of ~ 5·103 _{- 5·10}5 _cfu/ml.

cfu/ml

103 ₁₀4 ₁₀5 ₁₀6 ₁₀7

Intenstiy analysis

Cluster-based blob detection

in one out of four experiments at this inoculum size. Because of the small size of the chamber,

loading the assay with a low cfu/ml will generate few data points, the resolution of the signal will be

low and stochastic noise will become more prominent. Assuming perfect ability to measure every

cell in the system and a perfect random distribution, a minimum of 50 cells in the chamber is what

we believe is required for sufficient concentration resolution. This would set the physically possible

lowest inoculum size detection limit to be approximately 50 cells (1 ml/8 µl = 6.25·10

3

_cfu/ml)

when using CellDirector 3D. A smaller chamber size would increase determined lowest detection

limit of this assay. Taking into account experimental variation, a practical lowest detection limit

of 1·10

4

_{cfu/ml is advisable based on the results in this study. However, the output from lower}

inoculum sizes to approximately 10

3

_{cfu/ml could still give qualitative information about resistance}

level, even if an exact MIC value could not be determined.

Evaluation of the shortest time to detection

Image data from CellDirector 3D was collected during 5 hours, but the actual time to detection

appears to be approximately 2-3 hours for the strains tested in this study (fig. 15). For vancomycin, a

large glycopeptide that is often used as the last resort for treatment of multiresistant Gram-positive

infections, the time until a linear antibiotic gradient is established by diffusion in CellDirector 3D

is around 3 hours. The results in this report suggest that for antibiotic compounds with lower

molecular weight, such as ciprofloxacin, it would be possible to determine a MIC within two hours

when image data is collected in the oCelloScope (fig. 15A). However, the shorter time to detection

needs to be statistically verified by conducting more experiments.

C

ONCLUSIONS

One of the main advantages of using CellDirector 3D is that it is a phenotypic method that cannot

only determine susceptibility or resistance but also return a MIC value. This study suggests that

the lowest inoculum size for CellDirector 3D is 10

4

_{cfu/ml and a MIC can be obtained within 2-3}

hours depending on detection algorithm and antibiotic (fig. 17). In comparison to other phenotypic

rapid AST methods, LifeScale (Affinity Biosensor) can give an antibiotic susceptibility result within

2 hours from positive blood culture, and Accelerate ID/AST (Accelerate Diagnostics, Inc) requires

1 hour for species identification and 5 hours for AST, also from a positive blood culture. ASTRID

(Q-Linea) and the automated system Iridica (Abbott) both require 5 hours for ID and AST from

whole blood. It is most likely not possible to apply CellDirector 3D directly on whole blood, but the

assay could be applied after a shorter incubation step (~4 hours). In the near future, the QuickMIC

assay will be evaluated in a clinical setting using the five most common sepsis pathogens. From a

positive blood culture, species identification can be determined using MALDI-TOF. Thus, AST

and ID could be obtained within a working day and before the second dosage interval during

treatment of patients.

Positive

blood culture MALDI-TOF MS CellDirector 3D

ID + AST

4-6 hours