Automated evaluation of crystallisation experiments

Research output: Contribution to conferencePaper


The gradient of a pixel intensity is used to detect image discontinuities and, as a vector quantity, provides not only the magnitudes of sharp changes in the image but also the direction in which the change is greatest. This direction is perpendicular to the important edges in the image and can be used, for example, to identify circular objects. When applied in crystallisation screening, this allows a mask to be applied around a crystallisation drop and the analysis to be restricted to this region. Artefacts within the crystallisation drop are then considered as individual objects and each is evaluated in terms of a number of attributes related to its size and shape, curvature of the boundary and variance in intensity. Other more obvious crystal-like characteristics such as straight sections of the boundary and straight lines of constant intensity within the object are also calculated and a set of values, or feature vector, is assigned to each object. A combination of self-organising maps and decision trees are used for classification. The weights are obtained from the feature vectors associated with a training data set consisting of objects from each class that have been pre-classified by eye. Consistent patterns in the combination of the variables allow new objects to be assigned to various classes and given a related score, in this case, ranging from 0 for insignificant objects, such as those due to shadows at the edge of the drop, to 6 for good single crystals. With each object in an individual image classified in this way, an overall score can be given for that image, reflecting its likelihood of containing crystals.
Original languageEnglish
Number of pages12
Publication statusPublished - 1 Jan 2004


  • Crystallization;
  • Image Processing;
  • Pattern Recognition;
  • Automation

Cite this