1-s2.0-S0928098702001896-main.pdf

(1494 KB) Pobierz
European Journal of Pharmaceutical Sciences 17 (2002) 217–227
www.elsevier.nl / locate / ejps
D
oes a powder surface contain all necessary information for particle size
distribution analysis?
*
a,
a
a,b
Niklas Laitinen
, Osmo Antikainen , Jouko Yliruusi
a
Pharmaceutical Technology Division , Department of Pharmacy , P . O . Box 56, University of Helsinki , 00014 Helsinki , Finland
b
Viikki Drug Discovery Technology Center ( DDTC ), P . O . Box 56, 00014 University of Helsinki , Helsinki , Finland
Received 4 April 2002; received in revised form 8 August 2002; accepted 13 September 2002
Abstract
The aim of this study was to utilise a new approach where digital image information is used in the characterisation of particle size
distributions of a large set of pharmaceutical powders. A novel optical set-up was employed to create images and calculate a stereometric
parameter from the digital images of powder surfaces. Analysis was made of 40 granule batches with varying particle sizes and
compositions prepared with fluidised bed granulation. The extracted digital image information was then connected to particle size using
multivariate modelling. The modelled particle size distributions were compared to particle size determinations with sieve analysis and
laser diffraction. The results revealed that the created models corresponded well with the particle size distributions measured with sieve
analysis and laser diffraction. This study shows that digital images taken from powder surfaces contain all necessary data that is needed
for particle size distribution analysis. To obtain this information from images careful consideration has to be given on the imaging
conditions. In conclusion, the results of this study suggest that the new approach is a powerful means of analysis in particle size
determination. The method is fast, the sample size needed is very small and the technique enables non-destructive analysis of samples.
The method is suitable in the particle size range of approximately 20–1500 mm. However, further investigations with a broad range of
powders have to be made to obtain information of the possibilities and limitations of the introduced method in powder characterisation.
2002 Elsevier Science B.V. All rights reserved.
Keywords : Image analysis; Powder characterisation; Particle size distribution; Digital image information; PLS modelling
1 . Introduction
1995; Eriksson et al., 1997; Andres et al., 1998; Bosquillon
et al., 2001; Passerini et al., 2002).
`
Digital images contain a significant amount of infor-
Ros et al. (1997) define the characterization of granular
mation and this information should be utilised much more
products with image analysis to be complex since the
effectively in material characterisation. Recently, Grasa
defining of the sample size is difficult and because of the
and
Abanades
(2001)
have
pointed
out
the
growing
diversity of the data that may be extracted from digital
importance of quantitative analysis of digitalised images in
images.
In
analysis
of
particle
sizes
and
shapes,
only
industries handling bulk powders. Within pharmaceutical
characteristics of individual particles are usually measured.
technology,
images
are
produced
with
various
optical
It would be advantageous to examine powders as larger
techniques, e.g. to evaluate the surface morphology of
populations by examining powder surfaces, which create
particles and coatings (Cartilier and Tawashi, 1993; Ken-
descriptors
for
the
analysed
material
in
question.
This
nedy and Niebergall, 1997; Pons et al., 1999; Andersson et
approach removes the problems related to the dispersion,
al., 2000). Furthermore, digital images are used in pharma-
which often adds a time-consuming step to sample prepa-
ceutical powder technology when employing image analy-
ration and can be very problematic for many particles, e.g.
sis (IA) techniques in measurements of size distributions
due to cohesiveness of the material. The difficulty and
and different shape parameters (Brewer and Ramsland,
importance of sample preparation and dispersion with IA
techniques is well known and it has been discussed in
literature (Allen, 1988). Traditional image analysis, where
individual particles are measured can also be very time
* Corresponding author. Tel.: 1358-9-191-59746; fax: 1358-9-191-
consuming. Pons et al. (1999) have reviewed and illus-
59144.
E - mail address : niklas.laitinen@helsinki.fi (N. Laitinen).
trated the use of IA in routine analysis and bring up issues
0928-0987 / 02 / $ – see front matter
2002 Elsevier Science B.V. All rights reserved.
PII: S0928-0987(02)00189-6
964250559.016.png 964250559.017.png 964250559.018.png 964250559.019.png 964250559.001.png 964250559.002.png 964250559.003.png 964250559.004.png 964250559.005.png 964250559.006.png 964250559.007.png 964250559.008.png 964250559.009.png 964250559.010.png 964250559.011.png 964250559.012.png
 
218
N . Laitinen et al . / European Journal of Pharmaceutical Sciences 17 (2002) 217–227
concerning comparisons of different sizing methods, the
made from two images, which are illuminated from two
number of analysed particles and elements in shape and
different directions.
texture quantification of particles. They also discuss the
Optical techniques that are used for measurement sur-
issue of reluctance of the use of IA techniques in analysis
face characteristics are widely used in different industrial
of particle morphology. It seems that the limited routine
inspection applications. Cielo (1988) has thoroughly de-
use
of
IA
is
related
to
the
relative
slowness
of
the
scribed
the
field.
Often
these
industrial
techniques
are
characterisation process, the large size of an image as a
operations that measure the roughness characteristics of
dataset and the variety of the choice of methods to use in
different surfaces in process control and quality control
problem solving.
environments. Russ (1999) has described the concept of
Probably the most common methods in particle size
surface imaging comprehensively.
analysis besides image analysis are sieve analysis and laser
The aim of this paper is to introduce a new approach
diffraction. Washington (1992) lists a number of problems
where digital image information is used in the characterisa-
and errors that prevalent in sieving, e.g. errors in the
tion of particle size distributions of a large set of pharma-
sieves, influence of the sieve load and sieving time, and
ceutical granules. We describe and utilize a new optical
errors due to the properties of the material. The choice of
set-up
to
create
images
and
calculate
a
stereometric
sieving time can have effects on the results and the amount
parameter from the digital images of powder surfaces,
of
material
sieved
may
affect
reproducibility.
Material
which can be connected to particle size using multivariate
cohesiveness may also cause errors in measurements and
modelling.
result in false size distributions. The amount of sample in
sieving is relatively large, therefore, it is not suitable for
expensive
materials
or
materials
of
which
only
small
2 . Materials and methods
quantities are available. The use of laser diffraction is very
popular in particle size measurement of pharmaceutical
2 .1. Materials
powders. A disadvantage is that the amount of sample is
usually relatively large when the particle-in-air method is
In total, 40 granule batches with varying particle sizes
used. There are many advantages with light diffraction
and compositions prepared with fluidised bed granulation
analysis, e.g. the ease of operation and reproducibility.
were analysed. The granulations were made in a bench-
Previously, we have introduced a concept of comparing
scale fluidized bed granulator in varying process conditions
image
information
of
pharmaceutical
powder
surfaces
(Glatt WSG 5, Glatt, Binzen, Germany). The sampling for
(Laitinen et al., 2000). This concept is based on the idea
the
granules
was
made
using
a
rotary
sample
divider
that image information of powder surfaces with similar
(Fritsch laborette 27, Idar-Oberstein, Germany). In the text
visual appearances could be linked with related material
the granule batches will be referred as granules G1–G40.
properties.
We
used
a
content-based
image
retrieval
(CBIR)
system,
which
has
been
developed
to
extract
2 .2. Methods
feature information from images in large image databases.
We proved that there is substantial information in digital
2 .2.1. Particle size measurements with sieve analysis
images of powder surfaces that can be linked to particle
and laser diffraction
size. Huang and Esbensen (2000, 2001) have introduced an
The particle size distributions were measured with sieve
approach in powder characterization that uses global image
analysis (Fritsch analysette, Germany) using the following
information. They applied angle measure technique (AMT)
sieves: 0.045, 0.071, 0.090, 0.125, 0.180, 0.250, 0.355,
in
image
analysis
of
a
wide
variety
of
powders.
The
0.500, 0.710, 1.000, 1.400, and 2.000 mm. The sample size
method simplifies traditional IA and does not deal with
in sieve analysis was 20 g (5 min with amplitude 6). A
individual
particles.
In
the
AMT
approach,
a
special
Fraunhofer laser diffraction instrument (Malvern 2600C,
camera with red, green and near infrared channels was
Malvern Instruments, UK) with a dry powder feeder was
used with low-angle unilateral illumination. AMT has been
also
used
to
measure
the
particle
size
distributions
of
developed to describe signal complexity as a function of
granules G1–G40. The focal lens length used was 800
geometrical scale from local to global. In the described
mm. Three samples were measured with both techniques
application, the images of powders are unfolded to produce
( n 53).
one-dimensional measurement series, which AMT trans-
forms to multivariate scale characterisations. They were
2 .2.2. Particle size measurements with the optical
able to create models between different powder charac-
technique
teristics such as, particle, size, wall friction angle and angle
of repose, and digital image information. In the present
2 .2.2.1. Sample preparation
study we also exploit global image information, but the
The samples were prepared by pouring powder to a
imaging environment and the data generation is different.
conical sample cup. An aluminium plate was then used to
A monochrome camera is used with unilateral illumination
level the powder surface with the upper edge of the sample
and calculations to create a particle size distribution are
cup. The powder surface forms a circular surface with a
N . Laitinen et al . / European Journal of Pharmaceutical Sciences 17 (2002) 217–227
219
10-mm
diameter.
Three
samples
of
each
powder
were
prepared ( n 53).
2 .2.2.2. The imaging setup
The optical instrument consists of the following ele-
ments. The imaging unit, which has a light source and a
monochrome CCD camera (JAI, CV-M50, Copenhagen,
Denmark) with a lens objective, is connected to a frame
grabber (WinTV, Hauppauge Computer Works, Haup-
pauge, NY, USA) and a Personal Computer. The symmetri-
cally positioned, bilateral, light sources, on opposite sides
of the sample, stand on rails, on which they can be
accurately positioned. The illumination system includes
two lamp housings, 100 W quartz tungsten halogen lamps,
and two collimating lens assemblies (Oriel Instruments,
Stratford, CT, USA). The collimated output beam can be
turned 908 with a beam turning assembly. The light sources
are connected to stabilized DC power supplies (Oriel
Instruments, Stratford, CT, USA). The imaging system is
presented Fig. 1.
In this study the following imaging settings were used.
A 50-mm lens objective with additional 40 mm extension
tube was used. The dimmer was set to F8. The light source
Fig. 2. A graph of rough (a) and smooth (b) synthetic surfaces. The
distance
from
the
sample
was
20
cm.
The
angle
of
formation of shades on the surface using unilateral illumination.
illumination was 308. The used power source voltage was
5.5 V and the image resolution in the frame grabber was
6003800 pixels. The dimensions of each sample surface in
monochrome camera is used. Therefore each image pixel
the taken images were 8.236.1 mm. All images were taken
gets grey scale values between 0 and 255, where 0 is
in
a
dark
room
with
no
disturbing
light
sources.
The
totally black and 255 completely white. By using lateral
settings were chosen as a result of extensive preliminary
illumination
with
a
substantial
illumination
angle
the
studies
on
optimising
the
imaging
environment.
The
shading
effects
expose
the
topography
of
the
surfaces
calibration of the imaging conditions was made with a
examined. The shading effect and the consecutive effects
smooth
non-reflecting
white
calibration
board
(Xerox
on the grey scale variations are illustrated graphically with
Premier, batch 11 / DD/ YKD/ 1, Xerox Corporation, USA).
examples of synthetic surfaces in Fig. 2.
2 .2.2.3. Image information and shade formation
2 .2.2.4. Imaging and calculation of
the difference dis -
The
digital
image
consists
of
a
matrix
of
picture
tribution
elements or pixels and each pixel corresponds to a surface
Two images of each sample were taken. The two light
element in an image. The principal idea behind this method
sources are used to illuminate the sample. One digital
is, that smooth surfaces will have small variations in grey
image of the sample is first taken by illuminating the
scale values, and the rougher the surfaces are, the larger
sample with light source 2a (Fig. 1). Then, another image
are the variations in the grey scale values. In this study a
is taken by illuminating the sample with light source 2b
Fig. 1. A scheme of the imaging setup. 15CCD camera with optics, 2a and 2b5light sources on rails with collimated light beams, 35sample in sample
cup, 45PC and frame grabber, 55image storage.
964250559.013.png 964250559.014.png
220
N . Laitinen et al . / European Journal of Pharmaceutical Sciences 17 (2002) 217–227
(Fig. 1). Consequently, two digital images are received and
2 .2.2.6. The PLS model for laser diffraction
two matrices of their grey scale values are formed. The
A PLS model that corresponds particle size analysis
difference of these two matrices is then calculated. For a
with laser diffraction was also created using the 511 values
theoretical completely smooth surface the difference of the
from
the
distributions
of
the
difference
matrix
of
the
two
matrices
consist
of
zeros.
For
a
real
surface
the
granules G1–G40 as explanatory variables. The percent
difference matrix gets values between 2255 and 1255. In
volume proportions of 31 size fractions of the analysed
the next step a distribution of the difference matrix is
granule batches were use as the response variables. Data
formed, i.e. how many numbers represent each of the
from three parallel samples of the 31 batches were used as
possible 511 values (Fig. 3). The calculations were per-
model data and data from three parallel samples of the nine
formed with Mathcad 2001 Professional software (Math-
batches were used as test data. The same test batches that
Soft, USA). Fig. 3 presents two different example granule
were used for the PLS modelling for sieve analysis was
sample surfaces with two images illuminated from the
used for laser diffraction. These specific test batches were
opposite sides for each material. Subsequently, the forma-
chosen to get a representative test data that covers the
tion of the difference distributions from the difference
particle size range of all batches as well as possible. The
matrices
is
shown.
One
can
notice
that
the
difference
selection was made according to the particle mean size
distribution
is
characteristic
for
the
different
kind
of
results and after a visual comparison of the particle size
surfaces.
distributions of the granule batches.
The size distribution is then derived from the difference
matrix received from the two digital images. The concept
of
the
measurement
is,
that
a
sample
with
a
specific
3 . Results
particle size distribution forms a characteristic distribution
for its grey scale difference matrix. In order to generate a
3 .1. Particle size measurements
size distribution a model between the real particle size
distribution and the distribution of the grey scale value
The mean particle sizes measured with sieve analysis
difference matrix has to be created. This model transforms
and laser diffraction are presented in Table 1. The mean
the grey scale difference matrix to a particle size dis-
granule size measured with sieve analysis varies from 91
tribution. In this study, we evaluated the model that was
to
553 mm
and
the
mean
sizes
measured
with
laser
created using the particle size measurement results from
diffraction vary from 90 to 924 mm.
sieve analysis and laser diffraction. We used multivariate
modelling and created a PLS model (partial least-squares)
3 .2. Modelled particle size distributions
with Simca-P version 8.0 software (Umetrics, Umea,
Sweden). PLS relates two data matrices, x and y , to each
˚
Pairs of images of batches G1, G9, G21, G28 and G35
other by a linear multivariate model. In our case, x is the
are shown in Fig. 4a–e. In the following text, these specific
grey scale difference matrix and y
is the particle size
test batches are used as representative examples to show
measurement results. The PLS method enables modelling
the results of the modelling. These example batches shown
of data in which the number of variables exceeds the
were chosen in order to demonstrate characteristic types of
number of observations (Wold, 1995). A Pearson correla-
models that were created.
tion analysis was made between the modelled and mea-
sured particle median size values.
3 .2.1. The model for sieve analysis
The
modeled
particle
size
distributions
of
the
test
2 .2.2.5. The PLS model for sieve analysis
batches G1, G9, G21, G28 and G35 with respect to sieve
For the forming of a model for sieve analysis the 511
analysis are presented in Fig. 5a–e as examples of the
values from the distribution of the gray scale difference
modeled test batches. The figures indicate that the created
matrix
of
granules
G1–G40
were
used
as
explanatory
models
corresponded
fairly
well
with
the
particle
size
variables and the percent mass proportion of each 12 sieve
distributions measured with sieve analysis. The correlation
fractions of the analysed granule batches were use as the
coefficient ( r ) between the median particle size of the
response variables. The data from three parallel samples of
results from the optical measurements and sieve analysis
31 granule batches were used as model data and the data
was 0.82 ( P ,0.0001). The distributions are more similar
from three parallel samples of nine granule batches as test
for particles in the size range from 100 to 200 mm (mean
data. The batches G1, G5, G8, G9, G19, G21, G25, G28
granule size with sieve analysis) (G9, G21, and G28).
and G35 were used as test data. These specific test batches
Granules
with
smaller
mean
sizes
and
their
size
dis-
were chosen to get a representative test data that covers the
tributions are over-represented in the training data com-
particle size range of all batches as well as possible. The
pared to granules with larger granule mean sizes. This
selection was made according to the particle mean size
explains the better models for smaller particles. A size
results and after a visual comparison of the particle size
difference
around
200 mm
in
granule
mean
size
of
distributions of the granule batches
modelled data and the size distribution obtained with sieve
Fig. 3. An example of the formation of the difference matrix and the difference distribution from images of two different granule batches. The images Ex1 and Ex2 are a pair of images and images
Ex3 and Ex4 also form a pair from which the difference matrices are calculated.
964250559.015.png
 

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika:

Zgłoś jeśli naruszono regulamin