Advances in On-line Monitoring and Control of the Morphological and Polymorphic Forms of Organic Crystals Grown from Solution †

Organic crystals grown from solution are known to exhibit multiple morphology as well as habits which are of great importance to the end-use property of the product such as the bioavailability and the downstream processing such as in filtration and drying. The crystal morphology can also dictate other quality measures such as the size. Compared with the great amount of research work that has been done on the on-line measurement of other quality measures such as the size and concentration using various spectroscopy techniques, the literature on the on-line measurement and manipulation of crystal morphology is scarce. Attempts were made in the past to use laser diffraction and ultrasound spectrometry for shape monitoring. These methods have not proved to be very successful due to the difficulty in extracting detailed shape information from the signals corrupted by noise and multiple scattering. In this paper, we describe a new approach for on-line crystal morphology measurement and control which is based on the integration of online imaging, multi-scale image analysis and crystal morphology modelling, and present the results obtained on applying the approach to the batch crystallisation of (L)-glutamic acid. Online imaging proved capable of capturing high fidelity crystal shapes and polymorphic transitions in real time. A multi-scale image analysis method was proposed to extract the crystals from the image background and to calculate shape descriptors which were then used for shape recognition and to derive monitoring charts showing the ratios of different polymorphs in real time as well as the relative average growth rates of facets of crystals. Calculating crystal growth rates and estimating kinetics parameters for needle-shaped crystals was also investigated. Finally, a methodology called ‘camera model’ for integrating on-line imaging and crystal morphology modelling was presented.


Introduction
implying variability in product performance, e.g. via a different solubility and hence bioavailability for a drug compound.
In the past, a significant effort has been devoted to the development of on-line analytical techniques for monitoring deterministic operating variables including the concentration using attenuated total ref lectance Fourier transform infrared spectroscopy (ATR-FTIR) [1][2][3] , to monitor the crystallographic structure using X-ray diffraction (XRD) and Raman spectroscopy 4,5) , and the crystal size distribution (CSD) using laser diffraction 5) , ultrasonic spectroscopy 6) and light scattering techniques [7][8][9] . In contrast, the literature on monitoring crystal morphology is scarce 10) .
In addition to the ultimate objective of morphology control in manufacture, the on-line monitoring of crystal morphology can also help the development of crystal morphology modelling and prediction techniques. This has not yet come to the level of being able to consider the simultaneous effect of all engineering factors or operating conditions critically related to product morphology such as solvents and cooling rates or heat transfer effects 11) , despite the significant progress and the availability of commercial software systems such as HABIT 12) and CERIUS2 13) . It is known that minor changes in supersaturation, cooling rates, reactor hydrodynamics, pH and the presence of impurities can have a significant impact on the crystal product. In addition, there are unexpected factors which can become deterministic such as wall effects on heat transfer, baff les, and impeller materials and types 14) . As a result, on-line measurement and realtime classification of crystal polymorphous is important because it can partially compensate the inability in modelling the morphology and shape. On the other hand, being able to measure morphology on-line in real time will no doubt provide a tool that can help the development and validation of new and more sophisticated crystal morphology prediction models. This paper reviews the previous efforts on the online sensing of crystal morphology and presents recent advances in developing an enabling technique for the real-time measurement and manipulation of the morphology of growing crystals through integrating imaging, image analysis and morphology modelling.

Previous Work
Laser diffraction techniques were investigated previously for the recognition of non-spherical particles, but only limited success has been achieved in this area. The main difficulty has been in obtaining a single-particle pattern in mixtures due to problems associated with partial scattering of the particles [15][16][17] . Mougin et al. 6,18,19) attempted to interpret the ultrasound attenuation spectra in order to identify dynamic changes in polymorphs during the crystallisation of an organic compound. A correct interpretation was found possible though some additional information needs to be provided which is not always readily available. Studies have also been made to produce crystals with different size and shape properties through supersaturation control using ATR-FTIR 3) . The modification of temperature to maintain supersaturation values during nucleation and crystal growth helps produce uniform size and shape properties, nevertheless no information on particle shape is obtained during the experiment. This practice requires concentration and supersaturation models for the compound to crystallise obtained from calibrations, and the temperature range for control may be limited.
In laboratory studies, photo-microscopy has been the most widely used method for analysing crystal morphology. Samples are taken periodically from the reactor and observed using a photo-microscope imaging system. Software systems are also often available for quantitatively analysing the images. Patience and Rawlings 20) reported an interesting approach for automating this process using a f low cell in a laboratory study of the crystallisation of sodium chlorate. The f low through the cell was periodically stopped to allow a photo-microscope to capture images after the crystals had settled down at the bottom of the flow cell. In this way, the crystals of the compound were lying on a habit face and showing a particular shape to be characterised by a commercial image analysis software. They also developed an automatic control system by manipulating the impurity in the feed. The system has some limitations. Firstly, the sampling period is from 20 to 30 seconds, whereas significant morphological events in crystallisation might take only a few seconds. In addition, the use of a flow cell is clearly not ideal for many industrial processes. The f low cell configuration requires a pump to circulate the solution, and this might cause crystal breakage, as discovered by Mougin 18) . The on-line image analysis approach reported by Patience and Rawlings employs boxed-area and aspect-ratio to describe the crystal shape along with predefined cluster regions for the classification of the two studied shapes, square and triangle. Their strategy for crystal recognition does not allow detection of any new crystal shapes existing within the reactor, and requires the assump-tion of some shapes and calibration to predefine the corresponding classification regions for the shapes. In addition, the regions are fixed throughout the whole process, meaning that changes in morphology due to, for example, the crystal growth of each face or unexpected impurities may lead to a low amount of crystals being detected or may mislead the process control. Lasentec Inc, USA, developed a Particle Vision Measurement (PVM) probe which can be used for visualisation of the process in-situ. The PVM uses a charged coupled device camera with light sources fitted into the probe 21) . Barrett and Glennon 10) used the PVM probe to visualise the crystals in the determination of the meta-stable zone of an inorganic compound. Patience 21) made an attempt to analyse images obtained by the PVM probe in the crystallisation of sodium chlorate. He encountered difficulties in the analysis of the images using a commercial image analysis software tool. More recently, Wilkinson et al. 22,23) reported an approach using video imaging and image analysis for the on-line measurement of particle length and circularity in the crystallisation of drug compounds. Their approach uses a threshold method for the segmentation of the crystals from the images, and a low lens magnification for achieving a high measurement rate of the crystals. (b) obtaining quantitative information from images through image segmentation and image analysis to calculate shape descriptors and perform polymorph and morphology recognition; and (c) calculating crystal growth rates and estimating growth kinetics parameters.

Experiments
Experiments were carried out using a glass-jacketed reactor of 500 ml, a data interface board and a PC running WinISO process control software provided by Hazard Evaluation Laboratory Ltd. (Fig. 2) 24,25) . To control the temperature, a Julabo FP50-HD thermostatic bath was employed. A ref lux condenser was used to circulate water at 6°C to minimise solvent loss at the higher reactor temperatures. Reactor stirring was performed using a pitched-blade stirrer rotating at a constant speed of 200 rpm. The temperature was measured using a platinum resistance thermometer (PT100), turbidity was measured with an in-housebuilt turbidimetric fibre-optic probe system. Both signals along with pH values were logged into the computer system. Observation and recording of the process were carried out in real time using on-line imaging with an instrument developed by the pharmaceutical manufacturer GlaxoSmithKline, UK. For off-line image analysis, the PVS830 described below KONA  Shape descriptors, classification, monitoring charts, growth rates Fig. 1 The framework was employed, using a pipette to take samples from the reactor solution at several instances during the process. The samples were quickly placed and covered on the sample glass slide of the previously calibrated PVS830, and images were captured and analysed. The time from taking a sample until the offline image analysis was from 10 to 20 seconds. The on-line imaging instrument employed in the experiments is a high-speed CCD camera system which has a maximum frequency of thirty images per second with a typical pixel resolution of 480҂640 and a field of view from 140 µm to 16 mm. The camera is situated just outside the reactor wall and an imaging window is attached to the external reactor wall to avoid convexity effects on the images (Fig. 3). To provide illumination, a strobe light source is used and the light is conducted using a fibre-optic guide. Camera acquisition and f lashing light are synchronised to freeze the moving particles by using a camera interface box developed by scientists at the pharmaceutical manufacturer GSK 22,23) . The captured images are sent to a PC running VideoSavant software (IO Industries, Inc.) for acquisition, storage and management of frames.
To enhance the contrast of particles from the image background, two fibre-optic light guides can be used, both adjustable in angle and distance. The system allows the visualisation and recording in real time of every event occurring throughout the process, thus capturing the complete history of the crystallisation.  For an estimation of the particle size, the camera is calibrated with a glass micrometer scale (Agar Scientific Ltd.), using Mitutoyo objective lenses with Navitar Precise Eye lenses for different magnifications, allowing observations down to 5 µm. An off-line imaging system, the PharmaVision System 830 (PVS830) from Malvern Instruments Ltd. 26) , was also used in the study. It is an automatic vision system for the analysis of the size and shape of particles. A zoom lens allows particles between 0.5 and 2000 µm to be analysed. The instrument automatically calibrates itself prior to an analysis by the measurement of both light intensity and precision grating with a known number of lines per millimetre. To perform the analysis, the sample is placed on a sample tray under a CCD video camera. A linear actuator connected to a PC moves the camera across the sample tray and the camera takes digitised video images. Particles from the images are automatically segmented by embedded computer software, thus obtaining a variety of shape parameters such as length, width, mean diameter and roundness, and supported by images of all particles for further visual understanding. The performance of the instrument is verified by reference slides which are certified and traceable to standards issued by the National Institute of Standards and Technology.
Solutions of 33.3 g of L-glutamic acid (purchased from Aldrich Chemicals) dissolved in 500 ml of fresh distilled water were prepared. The solutions were heated up to 95°C and kept constant at this temperature until everything was dissolved, then linearly cooled down to 15°C, keeping this temperature until the end of the experiment. Different linear cooling rates of 1, 0.5 and 0.25°C/min were investigated. The concentration and temperatures of the experiments were chosen to be close to those used in an industrialscale operation. It was found that for the experiment employing a cooling rate of 0.25°C/min, the images revealed that the cr ystals were at first growing with a rhombic shape, an α-form ( Fig. 4 (a) and 4 (b)). Transformation into a β-form was observed at a temperature of 60°C, as illustrated in the real-time image of Fig. 5 (a) and (b). For the experiment using a cooling rate of 0.5°C/min, α-form crystals were also first found and the transition to a β-form was also observed, but at a much later stage corresponding to a temperature of 30°C. No transition from α to β was observed when a cooling rate of 1°C/min was used. The results confirmed and helped explain the crystallisation mechanisms 24) .
During the onset stage, the crystallisation experi- ments were carefully followed by the on-line imaging system. The onset point is said to occur when the first crystals start appearing in the solution, which is sometimes called the 'cloud point'. Looking at the real-time images, we were able to identify a number of appearing particles instants before the fibre-optic turbidity probe could show a pronounced reduction in light transmittance at the cloud point. The temperature at the onset points obtained by on-line imaging corresponding to the three cooling rates are summarised in Table 1. The results were found to be either earlier or at the same time as turbidity results 24) .

Image Analysis, Shape Classification and Monitoring Charts
To obtain quantitative information from the images for subsequent monitoring and control purposes, a major challenge is the availability of methods for image analysis that need to be not only tolerant of the quality of on-line images, but also accurate, fast and robust. A literature survey of the various techniques and tools led to the same conclusion as Patience 21) had come to, i.e. that the existing methods do not work effectively for the on-line images obtained from crystallisation reactors. Images of slurries with particles suspended in a solution are much more complex than images of purely solid particles, and the major challenges lie in the fact that the slurries in a stirred reactor are in continuous motion, and that the variation of the distance between the camera lens and the particles captured in a snapshot makes the edges of some particles more vague than others. In addition, the light effect and temporal changes of hydrodynamics within the reactor may lead to varied intensity in the background.
As a result, and as part of the research, a new multiscale approach based on the Canny 27) and wavelet functions 28,29) was developed which proved to be very promising in extracting the crystals from the image backgrounds 25,28,30) . The image segmentation method involves several steps: edge detection over two scales (so called multi-scale), mathematical morphological closing and opening to remove noises and small objects, to ultimately obtain the segmented crystal objects. Fig. 6 shows an example of analysing a poorquality image (to choose a low-quality image to demonstrate the capability of the method). Fig. 7 shows the analysis results for images obtained using the Malvern Instruments PVS830 and the PVM system of Lasentec. The processing time for a typical image with a resolution of 640҂480 obtained by the GSK system is between seven and ten seconds using a PC of 2 GHz. No significant difference in processing time was observed when analysing images from the PVS830 with a resolution of 756҂548.
Following image segmentation, the next step is to classify the shapes and extract morphological data 31) . The shape can be defined by the object features that maintain invariance under similarity transformations, the transformations including rotation, translation and size scaling 32,33) .
There are two major groups of methods for the use of descriptors to represent a shape. The first group includes techniques that provide a detailed representation and preserve information of the object boundary and thus can enable reconstruction of the original shape from the descriptor values. These types of descriptors usually consider a shape as being a periodic mathematical function which can be represented, for example, as an expansion or as a high-order polynomial. The second group includes global shape descriptors which can also be used to characterise the region boundary, albeit not preserving details of the original shape for reconstruction. These types of descriptors include, for example, critical distances, ratios or shape factors. For the latter, we use the software of a PVS830 imaging system from Malvern Instruments Ltd. to make the calculations, which give the key shape descriptors, while for the former, we use Fourier descriptors.
An extensive review on shape description techniques was made by Zhang and Lu 34) . They concluded that Fourier descriptors based on spectral transforms provide robust performance, accuracy, compact features and low computation complexity, as well as being the most promising method to overcome noise. Fourier descriptors have been used by several researchers for the contour analysis of particles 35,36) . This technique is popular due to its invariance under two-dimensional transformations, i.e. rotation translation and scaling, and its ability to describe a region  Table 1 Crystallisation onset values obtained from on-line imaging boundary in detail, preserving shape information for reconstruction. The calculated Fourier shape descriptors were then used by a shape classification algorithm, the adaptive resonance theory (ART2) 37,38) for classification purposes. ART2 is a neural network that adopts a learning mechanism that is both unsupervised and recursive. Since it is unsupervised learning, unlike back propagation neural networks that require training using pairs of predefined shape clusters and descriptors 36) , ART2 automatically determines the number of clusters and the assignment of data pat-terns in a way that patterns in a cluster are more similar than those in a different cluster. The recursive learning feature is a mechanism that can continuously update the knowledge with new data available, without corrupting the existing knowledge already learned using previous data and without the need to make up the new data with previous data for re-training or re-learning, also very important for on-line use (in some literature it is called incremental learning). Being recursive is extremely useful for on-line use since new data are continuously made available.
The methodology for image segmentation, descriptor calculation and shape recognition was applied to the on-line raw images obtained during cooling crystallisation of (L)-glutamic acid. The number and volume of the crystals captured in the frame over a fixed time interval were used to estimate the number and volume fractions between the two polymorphs, alpha and beta.
For estimation of the ratios between the prismatic and needle shapes, each time point in Fig. 8 and 9 represents 60 images taken in the last 10 minutes. Fig. 8 shows the relative percentage of the number of the two polymorphic forms, alpha and beta, plotted against time, from crystallisation onset throughout the polymorphic transition. The trends illustrate that the number of crystals with prismatic alpha shape became a minority at around 30 minutes from crystallisation onset. This type of chart is very useful to process engineers and operators to obtain the quantitative information about the polymorphic transitions inside the reactor. Due to the natural progress from onset through growth of the crystals, images of the first instants during onset contain only a few crystals compared to images of subsequent stages. As a result, clearer and less f luctuating shape trends are obtained as the time progresses and the amount of crystals detected increases. Nevertheless, the broad features of the polymorphic transformation are still evident through detecting the changes in crystal shape.
Due to the differences in particle size, a volume fraction can provide a better idea of the amount of compounds in solid state with a particular shape. The crystal shape patterns in volume fraction based on the last ten minutes is shown in Fig. 9. It is worth noting that in the volume fraction plot, the crossingover is observed earlier than in the number fraction 76 KONA No.23 (2005) plots, meaning that the amount of (L)-glutamic acid beta-form becomes predominant in the solid state at the crossing-over in Fig. 9, albeit at the same timepoint the amount of crystals with prismatic shape, alpha-form, still constitute the majority. The dynamic change in crystal shape can also be depicted in Fig. 10, which illustrates the values of shape roundness of alpha-and beta-form crystals as crystallisation in the reactor progresses. Each point represents a single crystal, and its corresponding roundness value can be read from the vertical axis. As expected, alpha-form crystals have higher values of roundness, while beta-form crystals have low roundness values, and the crystals are clearly separated into two clusters due to the difference in round-ness values, with the upper cluster representing alpha-form crystals and the lower cluster corresponding to beta. The plot reveals that at the very early stages of crystallisation, many more crystals have high roundness values than low roundness values, a reflection of the fact that there are more alpha-than beta-form crystals. As time progresses (accompanied with reactor cooling), more crystals have low roundness values (beta-form), and few have high roundness values (alpha-form), indicating the transition from alpha-to beta-form. Fig. 11 (a) and (b) show the evolution of length, width and equivalent diameter of the crystals for alpha-and beta-forms, respectively. For both cases, a clear relationship is observed between the patterns of crystal length and width. Nevertheless, differences can be observed. For instance, the beta-form crystals show a low difference in crystal width from the first to the last minutes, whereas the crystal length values rise from around 110 µm with more pronounced fluctuations to values of nearly 250 µm during the period of phase transition. In contrast, for the alpha-form crystals, although there are also differences in the crystal length and width patterns, they present more similarities compared with the needle-shaped cr ystal patterns. This reflects the actual differences in growth rate for the different crystallographic faces. In the beta-form crystals, the [101] crystallographic face clearly has a predominant growth rate compared with the other faces of the crystals, therefore leading to the characteristic needle shape. The crystal habit of the alpha-form, on the other hand, maintains the prismatic ratio as a result of similar growth rates of the crystallographic faces. In contrast to other on-line techniques for measuring the crystal size that can only provide a value similar to the equivalent diameter, the present monitoring charts obtained from image analysis show the trends of changes in crystal size in more detail. For instance, in our studies, the image analysis results make evident that oscillations in crystal size in needle-shaped crystals mainly correspond to crystal length with minor changes in crystal width, whereas the morphological dimensions of the prismatic crystals oscillate with similar trends.

Cr ystal Growth Rates and Kinetics
The information obtained from imaging and image analysis was also used to estimate the crystal growth rates 25) . This was conducted for cooling (0.1°C/min) crystallisation of LGA which produced needle-shaped β LGA. As an example, Fig. 12 shows the mean values of length along with supersaturation, temperature and turbidity. Each point is the average of the previous 60 seconds containing 300 images. Table 2 gives growth rate values estimated for both the length and width, and the literature result for the same chemical.
Comparing the estimated growth rate values with the literature, the growth rates of different faces of beta-form crystals of glutamic acid have been investigated previously by Kitamura and Ishizu 39) using the single crystal method. The literature reported that the growth rate value at the relative supersaturation of 0.5 is lower compared with the values obtained here in the same direction, the crystal length, in the supersaturation range from 0.47 to 0.51. This difference is consistent with the difference in temperature. Clearly, higher growth rates are obtained at higher temperatures. Other factors within the crystalliser vessel are also associated such as the existing hydro-dynamics or collisions that promote crystal surface dislocations which enhance the growth, compared with the conditions of measurement in a single crystal. Although it is expected that at higher temperatures the values of growth rate will also be higher, it is observed in Table 2 that, for length, the growth rate tends to increase under the (simultaneous) effect of supersaturation increase.
Kitamura and Ishiza 39) also reported that for the width, the growth rate was too small to be measured. Here, the measurements of crystal width involve a combination of the faces and the estimated growth rates for the width are 4 to 6 times smaller than those for the crystal length.
The capability to measure crystal growth rates could, in principle, be used in the estimation of para-   Table 2 Estimated growth rate values for length and width of needle-shaped crystals meters associated to the kinetics of growth, in this case image analysis providing the possibility of estimating the kinetics in different morphological directions of the crystals. Here, although the conditions are of cooling temperature, it is tempting to tr y to fit the growth rates obtained from imaging data to a kinetic model. Due to the low relationship found between width growth rates and supersaturation it was considered that for the case of width, it would not lead to a sensible estimation. Therefore, only the estimation of kinetic parameters associated with the crystal length was attempted.
To estimate the parameters, that is the growth kinetic constant k and the order n, a semi-empirical kinetic model was used R҃ks n which can be re-written in logarithmic form to estimate the parameters from a linear regression (Fig. 13) ln R҃n lnsѿln k Strictly speaking, the value of k is dependent on temperature following the Arrhenius equation 40) . Here, although the variation of temperature is relatively low for the studied points, the estimation was made above all to investigate the potential of using imaging in the estimation of growth kinetics, with the possibility of using the technique in future research in isothermal experiments. Table 3 summarises the obtained values.
Looking at the obtained value of the exponent, the estimation suggests that the growth in length is close to a dislocation controlled BCF mechanism. However, one needs to bear in mind that the semi-empirical kinetic model used is in fact describing the growth in the direction of a (100) face, which for this particular morphology is non-existent. The growth in the crystal length is mainly driven by the {101} family of faces, being aware of the existence of an angle between the growth rate in length and the growth rate of the (101) face. Due to the rotation of the crystals in the threedimensional space in the solution, it is difficult to provide a precise estimation of the growth rates for the face (101), R (101) , which would be given by the equation R (101) ҃R L cos 36°(6.3) Nevertheless, for the crystal morphology of the beta-form crystals, it is reasonable to assume that the growth rate in length is very close to the growth rate of the faces {101}.
The present results are encouraging. They clearly show that in the current state, the technique can provide a very close approximation of the growth rates and kinetics associated to some particular crystal faces. And they also show that if the rotation of the crystal in the solution can be quantified, the results provided by the imaging technique can lead to estimation of the same parameters of crystal growth of facets with high accuracy.

The Camera Model: Integration of Imaging and Cr ystal Morphology Modelling
The integration of cr ystal morphological modelling with in-situ shape measurement using on-line microscopy provides a potentially useful instrument for crystallisation study, for example, for validation of morphology prediction models, for constructing 3-D shapes from on-line 2-D images, as well as for classification of the polymorphic and morphological forms.   Table 3 Estimated kinetic parameters of growth in length for needle-shaped beta-form crystals An approach for the integration called camera model was proposed by Li et al 41,42) . It involves a means of generating a library of 2-D images through rotating a 3-D crystal in the multidimensional space, and a method for matching the 2-D images in the library to those 2-D images obtained on-line. The first step involves generating a 2-D library by rotating the 3-D crystals, which can be explained using Fig. 14. If the camera is placed at a point on the X axis, as shown in Fig. 14, the rotation around X axis will not result in different 2-D shapes, but only change the orientations of the same 2-D shape. As a result, the rotation around the X axis can be ignored. Since the crystal is a facetted object, when rotating around an axis, either Y or Z, from one angle to another at a very small step, for example, around the Y axis, the 2-D shapes corresponding to the two positions may only change in a very minor way. Fig. 15 shows some 2-D shapes when rotating a crystal at different angles, i.e. 1°, 2°, 3°, 4°, 5°, 6°, 7°, 8°, 9°, 10°, and 11°around the Y axis, from which it is clear that even when rotating the crystal 11°from the initial position, the 2-D shape does not change very much. This means that the number of 2-D shapes with a significant difference to each other is limited. Here, the words "significant difference" though are vague, and can be defined by the users. For instance, it can be defined as that as long as the number of edges in the projected 2-D images changes (increases or decreases), then it is considered a new 2-D image with a significant difference from its last image. It is worth pointing out that the minimum degree at which the crystal needs to rotate is dependent on the complexity of the 3-D crystal and the initial position of rotation.
Two methods of rotation can be used. One approach is to use a constant rotation angle to continuously rotate the crystal. In this case, the value of the angle must be small enough so that the similarity distance between two 2-D shapes corresponding to any two neighbouring orientations is smaller than a predefined threshold value. An alternative approach of rotation is to use variant rotation angles. In this approach, when rotating the crystal around an axis from one position to the next, the angle between the two positions is chosen so that the similarity distance of the two 2-D shapes corresponding to the two positions is larger than a threshold value sd 1 and smaller than another threshold value sd 2 (sd 1 sd 2 ). The advantage of this rotation method is that it can reduce the size of the 2-D image library generated. Fig. 16 shows some examples of 2-D images gen-  Fig. 17 shows an example of online images which has undergone the segmentation procedure. Table 4 shows the identification of crystals in Fig. 17 by matching the 2-D crystals with those in the 2-D library generated by the camera model.

Final Remarks and Future Work
In the past, pharmaceutical manufacturers have been hesitant to introduce new equipment, instruments and control techniques due to regulatory uncertainties which could lead to longer approval times and high costs 43) . As part of the FDA initiative to encourage innovation and efficiency of manufacturing process design and control and quality assurance 44) , the Process Analytical Technology (PAT) framework 45) is an important move decided on jointly by the regulatory authority and the industry, which could potentially lead to a revolutionary change in the manufacture of pharmaceuticals that requires all materials properties including concentration, size, shape, morphology and surface properties of feedstocks, intermediates and products at nano-, micro-and macro-scales, to be measured on-line or at-line. In addition, the measurements will not only be used for display purposes, but will also be analysed using advanced sensor and PAT data mining techniques to gain a fundamental understanding of the process and product behaviour   Table 4 The image matching results for crystals of Fig. 17 needed to achieve product quality control and assurance.
The results presented in this paper demonstrate that the integration of advanced on-line imaging sensors and multi-scale image analysis techniques provides great potential for implementing real-time techniques to manipulate the crystal growth and control the morphology of crystals. The work also can potentially lead to full integration of morphology modelling, measurement, control and optimisation techniques. The work reported in this paper was obtained with a laboratory-scale reactor. In the future, it is planned to extend the study to an industrial-scale 200litre reactor (Fig. 18)