Improved classification of crystallization images using data fusion and multiple classifiers

Abstract

Identifying the conditions that will produce diffraction-quality crystals can require very many crystallization experiments. The use of robots has increased the number of experiments performed in most laboratories, while in structural genomics centres tens of thousands of experiments can be produced every day. Reliable automated evaluation of these experiments is becoming increasingly important. A more robust classification is achieved by combining different methods of feature extraction with the use of multiple classifiers.

Time-Dependent Multi-Light-Source Image Classification Combined With Automated Multidimensional Protein Phase Diagram Construction for Protein Phase Behavior Analysis

“…An additional approach that would be of particular interest to enhance the identification of uncertain classifications in protein phase diagrams is the use of 2 or more parallel classification algorithms as presented by Buchala and Wilson. 22 The combined probability of multiple classifiers could enrich the classification outcome by a more accurate representation of classification uncertainty.…”

“…Incorporation of more classes to cover the wide variety of possible morphology types has shown a decrease in classification accuracy compared to simpler class systems. [19][20][21][22] However, high accuracy for more advanced approaches is required to fully capture the complexity of protein phase behavior. A strategy to improve the accuracy of classification models is the use of different data sources, which have been explored to find more distinctions between protein phase behavior (sub)types.…”

Using textons to rank crystallization droplets by the likely presence of crystals

“…Although previous studies set out to classify crystallization drops into discrete human-assigned categories, it is not clear that they demonstrated that this is achievable or indeed that the underlying premise of one-drop-one-score is even useful: individual drops routinely exhibit multiple precipitation behaviours that may inform one another but which are nevertheless only very loosely defined by the community (Newman et al, 2012). Unsurprisingly, even human classification of droplets yields poor agreement rates (Buchala & Wilson, 2008), and using such variable opinions as ground truths severely undermines the training of learning algorithms and not only reduces accuracy, especially for multi-class classifiers, but makes it unmeasurable. The increased rate of false negatives is particularly pernicious since the formation of crystals is in general a rare event yet is experimentally crucial to detect.…”

“…Early work used edge detection solely to detect crystals (Ward et al, 1988;Zuk & Ward, 1991), which was later extended to the autoclassification of experimental outcomes into various numbers of classes, for example 'clear', 'precipitate' or 'crystalline behaviour', with information derived from a diverse range of texture-analysis methods and/or edge detection, using off-theshelf machine-learning algorithms. While Wilson (2002) and Bern et al (2004) used only edge-based features, most work focused on texture analysis, either on its own or in combination with edge-based features, including grey-level occurrence matrices (GLCMs; Spraggon et al, 2002;Zhu et al, 2004;Cumbaa & Jurisica, 2010) and spectral methods such as Fourier transform (Walker et al, 2007) and wavelet modelling (Buchala & Wilson, 2008;Liu et al, 2008;Watts et al, 2008). More recently, Lekamge et al (2013) used time-series information by calculating GLCMs from the difference images of consecutive inspections.…”

“…There was also a general trend that started with deriving features globally from the whole droplet (Spraggon et al, 2002;Wilson, 2002), which shifted to identifying local features from subsections of an image (Bern et al, 2004;Zhu et al, 2004;Kawabata et al, 2006;Pan et al, 2006;Liu et al, 2008), but later returned again to global features (Walker et al, 2007;Buchala & Wilson, 2008;Watts et al, 2008;Cumbaa & Jurisica, 2010;Lekamge et al, 2013). A variety of machine-learning techniques have been used, including the naïve Bayes classifier (Wilson, 2002), support vector machines (Kawabata et al, 2006;Pan et al, 2006;Buchala & Wilson, 2008) and neural networks (Spraggon et al, 2002;Walker et al, 2007;Buchala & Wilson, 2008;Watts et al, 2008), but the majority of recent work has favoured decision tree-based methods (Bern et al, 2004;Liu et al, 2008;Cumbaa & Jurisica, 2010;Lekamge et al, 2013). It is notable that in all of this body of work it has not been much explored how to present the output of the computations to experimenters most effectively.…”