Statistical classifier

Automated data interpretation will usually be done using some statistical or AI technique. Because statistical classifiers are similar in their use to neural networks [Sarle, 1994] we will not discuss them separately. 

Neural network classifiers. The neural network or other statistical classifiers impose strong requirements on the data and the inspection, however, when these are fulfilled then good fully automatic classification systems can be developed within a short period of time. This is for example the case if the inspection is a part of a manufacturing process, where the inspected pieces and the possible defect mechanisms are well known and the whole NDT inspection is done in repeatable conditions. In such cases it is possible to collect (or manufacture) as set of defect pieces, which can be used to obtain a training set. There are some commercially available tools (like ICEPAK [Chan, et al., 1988]) which can construct classifiers without any a-priori information, based only on the training sets of data. One has, however, always to remember about the limitations of this technique, otherwise serious misclassifications may go unnoticed. 

Expert systems. In situations where the statistical classifiers cannot be used, because of the complexity or inhomogeneity of the data, rule-based expert systems can sometimes be a solution. The complex images can be more readily described by rules than represented as simple feature vectors. Rules can be devised which cope with inhomogeneous data by, for example, triggering some specialised data-processing algorithms. 

Case-based reasoning. The main advantage of CBR systems for NDT data interpretation is that they can cope with data coming from inspection of varying constructions under varying conditions with various system settings due to their ability to learn from the data classified by the operator. In such situations no reliahle statistical classifier can be designed, and the rule-hased classifiers would be either very inefficient or unpractically complex. 

Methods without models such as quantitative process history based methods (neural networks (Venkatasubramanian, et ah, 2003), statistical classifiers (Anderson, 1984)), or qualitative process history based methods (expert systems (Venkatasubramanian, et ah, 2003)),... 

Egmont-Petersen M, Talmon JL, Hasman A. Robustness metrics for measuring the influence of additive noise on the performance of statistical classifiers. Int J Med Inform 1997 46 103-12. 

In a further large-scale analysis on the same database, gene expression and biological screening data were used to identify a correlation between gene expression and cell sensitivity to compounds [83]. Sixty cancer cell lines were exposed to numerous compounds at the National Cancer Institute, and were determined to be either sensitive or resistant to each compound. Using a Bayesian statistical classifier, Staunton et al. 

These features are then input to a statistical classifier, such as an artificial neural network. The classifier is trained to distinguish benign from malignant lesion (CADx) or true from false detected lesions (CADe). The output of the statistical classifier is a... 

The output for CADx results can be more complex. A simple method is to report the output of the CADx scheme to the radiologist, which is related to the likelihood that the lesion is malignant. This output value can be converted to be an exact likelihood of malignancy (Jiang et al. 1999), and further, it can be adjusted to account for the differences in the prevalence of cancer in the dataset used to train the statistical classifier and the radiologists patient population (Horsch et al. 2008). 

Finally we remark that the majority of the parameters in the STCA filters have direct physical or mechanical interpretation, and that the transparency of the classification process is an important component in assuring the safety case for STCA. However, whether tuned by hand or optimised by a machine algorithm, the operational parameters are inferred from data. An alternative to direct physical modelling is to employ purely statistical classifiers, for example Ic-nearest neighbour classifiers or neural networks, for which there is no ready interpretation of the parameters. Nonetheless, these methods are highly effective in other areas and the machine optimisation of STCA parameters blurs the distinction between physical models on one hand and statistical black boxes on the other. We look forward to the construction of safety cases for purely statistical classifiers whose operational parameters are inferred from data and which have no ready physical interpretation. 

The final detected polyps are obtained by application of a statistical classifier based on the image features to the differentiation of polyps from false positives. Investigators use parametric classifiers such as quadratic discriminant analysis (Yoshida and Nappi 2001), non-parametric classifiers such as artificial neural networks (Jerebko et al. 2003b Kiss et al. 2002 Nappi et al. 2004b), a committee of neural networks (Jerebko et al. 2003a), and a support vector machine (Gokturk et al. 2001). In principle, any combination of features and a classifier that provides a high classification performance should be sufficient for the differentiation task. 

After 60 years, Fisher s LDA method is still the workhorse of the parametric methods. Of all the statistical classifiers, we understand LDA s theoretical properties best. We can derive it from ordinary (linear) least-squares (OLS) regression. LDA creates those linear combinations of the M features that best separate the classes. For each class k there is one of these discriminant functions The decision surfaces = dy - = 0 between any pair j,k) of... 

SHINN A p, KAY J w and soMMERViLLE c (2000),The use of statistical classifiers for the discrimination of species of the genus Gyrodactylus (Monogenea) parasitizing sahnonids. Parasitology, 120,261-269. 

---