The development of a semi-automated system of recognition for Gyrodactylus salaris based on statistical classifiers
Final report: Project FC1145
Andrew Shinn1, James Kay2 & Christina Sommerville1
1Institute of Aquaculture, University of Stirling, Stirling FK9 4LA
2 Dept. of Statistics, University of Glasgow, Glasgow G12 8QQ
Introduction
This project is funded under ROAME A for DEFRA funded R & D. It contributes to DEFRA’s policy objective which aims to ensure the health and diversity of wild fish stocks. ROAME A states that a priority of its scientific objectives is the development of rapid, sensitive and quantifiable diagnostic methods.
This project aimed to develop a method for the identification of Gyrodactylus salaris Malmberg, 1957, a List III pathogen of salmon. The objectives of the project set out to produce a semi-automated software programme requiring minimal expertise from the operator, designed specifically for use by the Fish Health Directorate. The programme aimed to execute a statistical classification system based on the morphology of characteristic structures comprising the parasite attachment organ.
Background to the current project
Previous work by Kay, Shinn & Sommerville (1999), McHugh, Shinn & Kay (2000), Shinn, Kay & Sommerville (2000) investigated the use of a number of different statistical classification methodologies to discriminate closely-related and morphologically similar species of Gyrodactylus von Nordmann, 1832 parasitising salmonids. These studies concentrated on the discrimination of three species, one of which was the notifiable pathogen Gyrodactylus salaris Malmberg, 1957. Morphometric data collected from light microscope images of the attachment hooks of these species were used to train and assess the performance of four different statistical classifiers. The performance of two methods, 9-nearest neighbours (9-NN) and linear discriminant analysis (LDA) gave confident discrimination of G. salaris from the two other species, G. derjavini Mikailov, 1975 and G. truttae Glaser, 1974. The 9-NN had the least number of misclassifications and so was tested further for the analysis of individual hooks. Five measurements made on the marginal hooks of Gyrodactylus measured under the light microscope gave near perfect discrimination of G. salaris (all G. salaris specimens correctly assigned but one specimen of G. derjavini was misidentified as G. salaris) (Shinn et al., 2000).
Having shown that discrimination of species was possible using statistical methods, it was questioned whether these methods could discriminate two taxonomically confused co-occurring species. Gyrodactylus thymalli Zitnan, 1960 from grayling (Thymallus thymallus) is not a notifiable species but it is commonly confused with G. salaris, and their accurate discrimination from one another is vital. When only these two species were considered, LDA provided high correct classification percentages (G. salaris as G. salaris 98.1%; G. thymalli as G. thymalli 99.9%) (McHugh et al., 2000). However, when the dataset was expanded to encompass additional species (e.g. G. truttae, G. derjavini), the percentage of correct classification dropped (G. salaris as G. salaris 94.6%; G. thymalli as G. thymalli 98.2%). To circumvent this problem, a two-stage classifier was used; the first round to discriminate morphologically similar pairings (G. salaris-G. thymalli from G. derjavini-G. truttae) followed by a second round of classification that then discriminated between G. salaris and G. thymalli. These studies formed the foundation for the development of the semi-automated system on which this project is based.
Objectives of the current project
The method of identifying G. salaris currently employed by CEFAS uses the classic approach of manually taking point-to-point measurements on the attachment hooks of Gyrodactylus using an eye-piece graticule and comparing the measurement ranges for the different variables. This approach is laborious and is dependent upon the operator making accurate length readings using a graticule, which at the light level is divided into coarse divisions (0.5 micron intervals only).
The objectives of the current project were to develop better, rapid, semi-automated diagnostic systems, which could be used by non-specialist operators to separate eight species of Gyrodactylus parasitising salmonids. Previous work (Kay et al., 1999, McHugh et al., 2000; Shinn et al., 2000) had shown that methods could be developed with a high accuracy of identification using a digitising tablet that permits accurate positional measurements, which are then used to construct a statistical classifier to identify the species of Gyrodactylus under analysis. This approach constituted Phase I of the project. Phase II of the project was less well developed but had greater potential for user accuracy without the loss of computer processing time. The objective of Phase II was to explore the potential of automating the entire process by exploring methods that extract the outline of the hook and shape-based information, that could be used to discriminate species. These shape analysis based methods could then be developed further in future projects to develop a second classifier that could then be used to discriminate species on the basis of hook shape.
Methodology
Phase I methodology
For identification, the parasite is mounted onto a glass slide and viewed using a light microscope linked to a computer. Image software is used to collect morphometric data on the attachment hooks, which are then fed automatically into a statistical classifier, which identifies the species under analysis. A number of statistical classifiers were trained and assessed. To train the classifier, the classifier was exposed to data collected from different species of Gyrodactylus that had been identified by an expert, and the system adjusted until it always gave the correct answer. The classification efficiency of the trained classifier was then assessed by presenting specimens to it whose identity was known only by the Gyrodactylus expert. For the purpose of this project, it was agreed that the species to be included should be those species of Gyrodactylus commonly found on British salmonids e.g. G. derjavini, G. thymalli, G. truttae and G. salaris, a species not occurring in the UK but made notifiable as a Class III pathogen in the UK in 1988. In addition, species that could accidentally parasitise salmonids e.g. G. arcuatus Bychowsky, 1933 and G. gasterostei Glaser, 1974 were included as well as other potentially problematic species e.g. G. sommervillae Turgut et al., 1999. The classifier was refined in order that it could be used by non-specialists following demonstrations of its use. This classifier is supported by a substantial manual. This classifier is referred to below as “PointR” (Fig. 1). It was proposed that personnel at the CEFAS, Weymouth Laboratory would test the software in its ease of use and its ability to discriminate certain species of Gyrodactylus.
Phase II of the project, explored the potential of using shape analysis to discriminate species based on reliable taxonomic criteria. Using image analysis, images of the attachment hooks of each species were processed semi-automatically and reduced to a wire-frame outline which were then further summarised by the application of shape descriptors to it (perimeters, feret measurements etc). These were then used to train a classifier in the same way. The preliminary results of the approaches tested for Phase II were so successful that a funded extension (4 months) was agreed in order to develop a classifier based on shape classification for non-specialists to use. This second classifier, "OGRE", is also supported by a substantial manual and has been tested by research staff at CEFAS, Weymouth Laboratory.
Results
Gyrodactylid classifiers based on point-to-point measurements
The classifier on which the semi-automated system was built utilized a large data set (owned by the Institute of Aquaculture (IoA)) of morphometric data of species of Gyrodactylus that parasitise salmonids in the UK and G. salaris from Scandinavia. However, new data was collected to increase the number of specimens and to incorporate specimens from wider geographic locations, habitats and seasons.
Species included within the point-to-point based classifier “Point-R”.
Eight species of Gyrodactylus, which included multiple populations, were collected, measured and entered into the database and can now be discriminated by the current version of “Point-R”. The species used were: G. arcuatus Bychowsky, 1933 from Gasterosteus aculeatus L.; G. derjavini Mikailov, 1975 from Oncorhynchus mykiss (Walbaum), Salmo salar L. and Salmo trutta L.; G. gasterostei Gläser, 1974 from G. aculeatus; G. kherulensis Ergens, 1974 from Cyprinius carpio L.; G. salaris Malmberg, 1957 from Salmo salar; G. sommervillae Turgut et al., 1999 from Abramis brama L.; G. thymalli Zitnan, 1964 from Thymallus thymallus L. and G. truttae Gläser, 1974 from Salmo trutta. A total of 1032 specimens were analysed (14,910 pieces of data), and used for the construction of the statistical classifier.
The morphometric variables on which the classifier is based and trained.
Malmberg (1970) described 16 morphometric variables to characterise the attachment hooks of Gyrodactylus. Later, Shinn et al. (2001) revised the list of morphometric parameters by adding another five parameters. To increase the potential discriminatory power of the classifier used in the study, a further five parameters were described and incorporated. The macro "Point-R" is based on 25 point-to-point measurements. An excerpt from the user manual describing some of the measurements used in “Point-R” is given in Figure 2.
The methods of statistical classification assessed for “Point-R”.
From the earlier studies by Kay et al. (1999), McHugh et al. (2000) and Shinn et al. (2000) two methods of statistical classification namely linear discriminant analysis (LDA) and k-nearest neighbours (k-NN) had performed well and these classifiers were assessed first in their ability to discriminate species based on the current dataset.
Cross-validation of the data
To gauge the performance of the classifiers, nine cross-validation experiments were conducted. Each experiment used a different subset of morphometric measurements e.g. the subset “all” includes all 25 morphometric measurements, “ham” includes only those measurements from the hamulus etc. The nine experiments are detailed in Table 1. In each case five-fold cross-validation and 100 random splits of the data using stratified random sampling were used. The tests were assessed using two statistical classification methodologies, linear discriminant (LDA) and k-nearest neighbour (k-NN) methods.
Figure 2: An excerpt from the user's manual on how to use "Point-R".
Stepwise discrimination
A stepwise discrimination using (a) stepwise linear discrimination in SPSS and (b) a classification tree in S-Plus were also ran. The aim of this approach was to sequentially pick and incorporate the individual measurements that provide the best discrimination, given the other measurements already built into the classifier. This procedure, of adding in the next best variable, continues until none can improve the classifier. The subsets of selected measurements used in each experiment are given in Table 1.
When using the stepwise approach, some caution is required as the performance of the classifier is assessed on the training data. While the addition of further morphometric variables ought be expected to improve the classification efficiency, it may in fact lead to the over-fitting of the training data, with the consequence that the fitted classifier may perform poorly when given new test specimens.
When using all 25 morphometric measurements, 21 of them were selected by the discriminant analysis and entered into the classifier. The measurements included were from the three major attachment hooks (hamuli, ventral bar and marginal hooks). The cross-validation experiments were then repeated using only the measurements available for each structure (e.g. subset “vb”, “ham” and “marg” (see Table 1)) and then again using the best measurements suggested by the stepwise discrimination tests (e.g. “first8”, “nine” etc). A summary of the cross validation results for each species using each subset of measurements are displayed in Table 2.
Table 1: Subsets of measurements used in the cross-validation experiments.
Subset / Hook measurements included in each experimentall / all 25
vb / all ventral bar only (six measurements)
ham / all hamulus only (11 measurements)
marg / all marginal hook only (8 measurements)
hm / all hamulus and all marginal hook measurements (19 measurements)
first8 / mhtl, hpl, mhdw, vbprol, mhsickl, hapert, vbtw,vbtl
nine / hapert, hpl, cosHPCA, mhtl, mhsickl, mhdw, vbtw, vbtl, vbprol
six / hapert, hpl, cosHPCA, mhtl, mhsickl, mhdw
eight / hapert, hpl, cosHPCA, mhtl, mhsickl, mhdw, hdsw, mhins
Abbreviations: cosHPCA = cosine of the hamulus point curve angle; hapert = hamaulus aperture distance; hdsw = hamulus distal shaft width; hpl = hamulus point length; mhdw = marginal hook sickle dorsal width; mhins = marginal hook instep height; mhsickl = marginal hook sickle length; mhtl = marginal hook total length; vbprol = ventral bar process length, vbtl = ventral bar total length; vbtw = ventral bar total width.
Table 2: Summary of the cross-validation results for each subset of morphometric measurements for the eight species of Gyrodactylus tested.
Median percent correct classification
Subset / arc / thy / der / gast / kher / sal / somm / truall / 100 / 100 / 100 / 100 / 100 / 100 / 96.8 / 93.3
vb / 100 / 56.7 / 76.7 / 100 / 76.7 / 100 / 69.8 / 53.3
ham / 100 / 93.3 / 100 / 89.3 / 90.0 / 100 / 87.3 / 90.0
marg / 100 / 93.3 / 100 / 100 / 96.7 / 93.3 / 92.1 / 93.3
hm / 100 / 100 / 100 / 100 / 100 / 100 / 96.8 / 96.7
first8 / 100 / 96.7 / 100 / 100 / 98.3 / 100 / 93.7 / 93.3
nine / 100 / 100 / 100 / 100 / 96.7 / 100 / 95.2 / 93.3
six / 100 / 100 / 100 / 100 / 96.7 / 100 / 93.7 / 93.3
eight / 100 / 100 / 100 / 100 / 100 / 100 / 95.2 / 93.7
Abbreviations: arc = G. arcuatus; der = G. derjavini; gast = G. gasterostei; kher = G. kherulensis; sal = G. salaris; somm = G. sommervillae; thy = G. thymalli; tru = G. truttae.
As can be seen from Table 2, the subset based on ventral bar measurements, "vb", produces the lowest correct classifications and can now be ignored. The results for subset "hm", which omit the ventral bar measurements, are as good as those obtained with all the measurements "all", and produces slightly better figures for G. truttae. The subset "eight" which contains four measurements from the hamulus and four measurements from the marginal hooks is also as good as the full set of measurements "all". To reduce processing time and effort, a diagnostician using this system could contemplate taking only those measurements identified within the “hm” (n = 19) or "eight" (n = 8) subset rather than all 25.
Summary of results for the subsets "all" and "eight" in detail for both methods of statistical classification
The significance of using, for example, these different two subsets, one using 25 measurements ("all") and the other using 8 ("eight"), can be compared for two different classification methods of choice (linear discriminant analysis (LDA) and k-nearest neighbours (k-NN)) (Tables 3 - 6).