Progress in modeling quality in aquaculture: an application of the Self-Organizing Map to the study of skeletal anomalies and meristic counts in gilthead seabream (Sparus aurata, L. 1758)

Introduction

Artificial life conditions, such as those imposed by aquaculture, can produce changes on cultured fish in term of morphology, physiology and behaviour (Kihslinger and Nevitt, 2006; Almeida et al., 2008). One of the most common drawbacks in aquaculture is the quite high presence of skeletal anomalies (SAs) in reared fish, which are generally associated with a general lowering of performances (i.e. swimming ability, conversion index, growth rate, survival, and susceptibility to stress, pathogens), on marketing image and therefore on commercial value of the reared lots (Hilomen-Garcia, 1997; Boglione et al., 2001, 2003, 2009; Cahu et al., 2003; Matsuoka, 2003; Lall and Lewis-McCrea, 2007; Le Vay et al., 2007; Castro et al., 2008). SAs includes variation in shape (skeletal malformations) and in number of skeletal elements (meristic count variations, asymmetry in the number of elements of paired fins). Gilthead seabream (Sparus aurata L.) is a fish of high commercial value due to its desirable characteristics (aroma, taste, white flesh). At present, the reduction of the gilthead seabream market price due to overproduction have forced aquaculture industry to reduce their production costs and improve their larval rearing efficiency. Skeletal malformations and their incidence are one of the most important factors affecting fish farmer’s production costs. Thus, Boglione et al. (2001) reported a range of 15–50% gilthead seabream juveniles with deformities are culled out from the productive cycle at the end of the hatchery phase.

Traditionally, different rearing approaches are applied in the framework of Mediterranean aquaculture of gilthead seabream. These can be divided in three main categories: extensive, semi-intensive and intensive. The first two are actually limited to few small realities, whereas the other one, which include land-based and off-shore facilities, represent the majority of total production of gilthead seabream. Larval rearing, on the other hand, has been exclusively carried out under intensive conditions. In order to improve the juveniles quality, and/or to obtain juveniles to be used for sea/lagoon-ranching, also gilthead seabream extensive and semi-intensive larval rearing have begun to be used in the Mediterranean area in the last years (Kentouri et al., 1995; Dhert et al., 1998; Divanach and Kentouri, 2000; Boglione et al., 2003; Cataudella et al., 2003). In this scenario, a morphological quality assessment of reared juveniles which use the wild juveniles phenotype as standard reference to evaluate the suitability of the used rearing methodologies has been proposed (Boglione et al., 1993, 1995, 2000, 2001, 2003, 2009; Cataudella et al., 2003; Gavaia et al., 2009). Nevertheless, there are no studies that tested statistically at a commercial (and not experimental) level the effects of the different larval rearing approaches on SAs onset and deformation occurrence in gilthead seabream.

In the last years, a new generation of powerful computational tools, the non-supervised artificial neural networks (ANN), have been used for classification, pattern recognition, empirical modeling and for many other tasks (Kohonen, 1989, 1995). These tools are particularly effective and sound for the analysis of ecological and biological data, which are frequently characterized by non linearity, internal redundancy and noise (Recknagel et al., 2006). Among ANN, Self-Organizing Maps (SOMs) are a data visualization technique which reduces the dimensions of data and generates a model in which the information stored in the original data is exemplified by ‘prototypes’, but also allowing the analysis of correlations among the data used for training with other external source of information. The way SOMs go about reducing dimensions is by producing a map of usually two dimensions which plot the similarities of the data by grouping similar data items together. So SOMs accomplish two things, they reduce dimensions and display similarities. This kind of ANN is useful to represent complex dataset and explore correlations between different types of information.

In this paper, SOMs was applied to a large dataset containing data on skeletal anomalies and meristic counts on about 60 lots of both wild and cultured seabream juveniles. These latter outcame from all the three rearing methodologies above described for this species. The aim of this work was to: (i) to develop a model of the occurrence of SAs in seabream, both in terms of type and quantity. This model should represent the pattern of co-occurrence of single skeletal anomaly and of the correspondence with the rearing approach used; (ii) to investigate correspondence between occurrence of skeletal anomalies and variability of meristic counts; (iii) to integrate this approach in order to produce a method for a scientific evaluation of quality in cultured lots of this species on the base of a wild-like similarity; (iv) to assess if this tool could be used to recognize lots origin by pattern of SAs or, conversely, predict incidence of skeletal by rearing approach.

Material and methods

The analyses were carried out on 4217 juveniles of gilthead seabream, belonging to 60 lots characterized by different origin (Table 1). In the following sections, the different sources of lots were abbreviated as: WI (wild), IN (intensive rearing), LV (Large Volumes rearing sensu Cataudella et al., 2002) and ME (Mesocosm rearing, sensu Divanach and Kentouri, 2000; modified). Specimens were anaesthesized, fixed in 10% formalin buffered with phosphate buffer (pH 7.2, 0.15 m) and double-stained in toto for cartilage and bone, according to Dingerkus and Uhler (1977) and Park and Kim (1984). Juveniles larger than 6 cm were X-rayed (Picker X-Ray cat.6191 – 805-E Control 599 Head). The data, collected from each individual by two different and independent operators, were: total and standard length, SAs and meristic counts. SAs were recorded with respect to their type and the body region in which they occurred. The skeletal terminology used is according to Harder (1975). The seven meristic counts (MC) considered were the numbers of vertebrae, number of rays of the left pectoral fin, of the anal fin, of upper and lower caudal lepidotrichia, and of dorsal spines and soft rays. The final form of the data was represented by a matrix of the frequency of each considered SA on each lot, computed from the binary (presence/absence) matrix of occurrence for each specimen. A SOM was then trained to display the high-dimensional datasets of SA in a 2-dimensional space: this implies a non-linear projection onto a lattice of hexagons. The Kohonen neural network consists of two layers: the input layer, connected to each vector of the dataset, and the output layer, consisting of a two-dimensional network of neurons (the units of the map). Each unit of the map is associated with a vector of weights, one for each input variable. In this study, the weights were represented by the frequency of each SA for each lot. For a detailed introduction to SOM, and for a complete description of training procedure, see Kohonen (2001), Park et al. (2003, 2004). The number of output units and the sizes of the map were chosen as 24 and 4 × 6, respectively, at the end of a calibration procedure based on computation of topological and topographic errors. These are two criteria usually used to evaluate the quality of the trained SOMs and to identify the optimal map in terms of size for a given input data. Indeed, maps of different sizes were trained and the size corresponding to the lowest values of these two error measures was selected (Park et al., 2003). However, it is important to stress that the size of the map did not alter the results, but simply affected the level of detail in the output. The City-Block (Manhattan distance) was chosen as distance. The number of epochs was set to 5000 in all cases. Subsequently, the fine-tuning epochs were set to 1000. The output of the SOM training procedure was represented in different ways. Labels were used to indicate the position of each lot at the end of training procedure. Differences in the border thickness of each hexagon was used to represent the Manhattan distance between adjacent hexagons, as computed by the vector prototypes. To represent the model of SAs occurrence, differential occurrences were represented with respect to the nine body regions used to classify SAs.

In this way, the probability value of anomalies occurrence was computed for each region and visualized as an image of seabream shape in which the different regions were identified by boundaries and filled in grey, whose darkness was proportionate to the probability of anomaly occurrence.

Range and median value were computed through the whole matrix for each meristic character. This way, a number of m discrete class was defined for each meristic character (MC). The value of frequency of individuals was then calculated for each discrete class of each MC, for each lot. The vectors of frequency obtained at the end of this procedure were preliminary compared using a Kolmogorov-Smirnov’ test to assess the homogeneity of the empirical distribution of MCs. Furthermore, the Hellinger distance was computed between the wild condition, which were considered as an unique lot, and each lot from captive conditions, for each MC. The seven values of distance obtained for each lot were assembled in an unique measure defined as the sum of these seven Hellinger distances. This measure was used to represent the dissimilarity from the wild condition as defined by the variability of meristic counts. Finally, both single values of distance and the total sum were plotted as external variables on the trained SOM.

To understand relationships between SAs and rearing approaches, we generated four binary vectors corresponding to the origin of each lot. Then, we calculated the mean value of each binary vector in each neuron of the trained SOM. If the output neuron was not occupied by input vectors, the value was replaced with the mean value of neighbouring neurons. These mean values assigned on the SOM map were visualized in grey scale, and then compared with maps of SAs.

The SOM was trained using 50 of the 60 lots. The resting randomly selected 10 lots were used to perform a test of the ability of the trained SOM to predict the characteristics of each lot (in terms of rearing approach used) using only the data about SAs. The test was carried out by detecting, for each of these 10 lots, the corresponding unit of the trained SOM: each lot was assigned to a SOM unit by simply computing his distances from all units and then selecting the smallest one. Then, the rearing approach corresponding to each lot was estimated using the pattern of occurrence of the different rearing approach in the corresponding SOM units, as obtained at the end of the training procedure. Finally, the SOM predictions were compared to the real information about lots origin.

Results

The trained SOMs classified lots according to the variation observed in the different weights of SAs. Values of quantization error (0.1) and topographic error (0.007) indicated that the SOM was smoothly trained in topology for the selected size. The panel in Fig. 1a shows the maps with the code labels corresponding to the lots assigned to each unit. The distribution evidenced a general high heterogeneity with either empty hexagons or hexagons with only few lots, but also hexagons with many ones. In particular, many lots populate the regions corresponding to the borders of the map, whereas the central area is less populated. Globally, the map showed a progressive increase of inter-hexagon distance moving from the down-right corner to the upper-left one, without particular discontinuities or full-blown groups (Fig. 1b). The panel in Fig. 1c shows the gradients for the nine seabream body regions. It is apparent that each anomaly was characterized by a pattern of occurrence in the neurons, and that while some of these patterns are in accordance, other ones are different from each others. However, the down-right corner corresponds to the lowest levels of SA incidences, whereas the opposite (upper-left) corner corresponds to the highest one. It was particularly clear specially for cephalic, hemal and caudal regions. For instance, an high level of occurrence (0.6–1) for anomalies of cephalic vertebrae characterized the first two units, whereas these anomalies are almost completely absente in the lower map region. A similar pattern could be observed for the anomalies occurring on the pre-hemal, hemal and caudal vertebrae, respectively. However, the right region of the map identify condition with high frequency of SAs on the hemal region. Finally, SAs affecting the caudal fin can be detected throughout the map. Figure 2(a) shows the result of the superimposition of the cumulative Hellinger distance on the trained SOM: it seems that data about SAs and MC are in complete agreement. In fact, a progressive increase of Hellinger distance can be observed along the diagonal connecting the first and the last hexagon. It descends that the lots associated to the first upper two rows of the map are the fairest ones from the wild conditions. Figure 2(b) shows the distribution gradients of wild and different rearing methodologies. Lots from Intensive rearing are associated to units located on the upper and left border of the map, while lots from Large Volumes correspond to units along the diagonal connecting the upper-right corner to the lower-left one. Mesocosms generated lots assigned to the units located on the lower-right area of the map, which is the same of wild lots. Figure 2(c) report the result of the predictive test carried out with the resting 10 lots not used for training. Each unit of the SOM tends to aggregate observations (lots) that are homogeneous with respect to the rearing approach used. When an hexagon contains lots with different origins, the most frequent condition identifies the most probable assignation to that hexagon, for a i-esime observation. In this way, the results evidenced a remarkable ability of the trained SOM to detect the characteristics of each lot in terms of rearing approach used: in eight of the 10 cases explored, the real condition corresponds to the highest frequency observed in the hexagon to which the lot is assigned. In the other two cases, however, the real condition corresponds to the second rearing category of the hexagon.

Discussion

In this study, we demonstrated the feasibility of modelling the occurrence of SAs in reared and wild-cached seabream by using SOM. The main goal of ordination is to achieve a well interpretable low dimensional visualization of pattern hidden in the raw data. Ideally, this pattern is purely empirical and data adaptive, and also allow for direct comparisons with different kind of information. The model obtained in this study represents the original information stored in the matrix of SA occurrence, but also allows the detection of a series of correspondence, namely between: (i) the patter of SAs occurrence and the different rearing approach currently used in seabream aquaculture (Fig. 2a,b); and (ii) the total SAs incidence and the variability of MCs. In fact, SOM represent the overall information by evidencing that wild lots correspond to the ‘extreme condition’, that is characterized by the substantial absence of SAs, while the different captive approaches deal with lots which are progressively far from this ideal situations. In this way, mesocosms seem to be the best choice to produce wild-like fish, whereas lots from intensive rearing are characterized by the large presence of SA. Moreover, a remarkable degree of homogeneity can be observed between lots from the same rearing approach, since they tend to segregate one near other in the trained SOM. Interestingly, this signal is in complete agreement with the information gained by the overall Hellinger distance computed on MCs. It is important to stress that data from MCs were not used in the trained procedure, and then represent a completely independent point of view. Globally, results suggested that this approach is reliable to be used for estimate the distance between aquaculture products and the wild-like phenotype used as quality reference (Fig. 1c). This seems of particular interest because data about SAs are easy to collect, not expensive, and also integrate information from different source of epigenetic and genetic disturbances which act during larval rearing. In reared fish, the identification of single causative factors for skeletal anomalies onset is impossible (and soundness), as there are so many factors (forced swimming, stressing stocking density, inadequate quality of administered food, reduced availability of space, handling practices,....) interacting on bone modelling and remodelling. In this framework, this approach seems to be a good method for integrating many and different types of informations and giving back a complete, clear and rapid result from analysis of data.

Finally, the approach presented here could be used to identify the best rearing methodology to produce juveniles with wild-like phenotype to be used for stocking or restocking actions in confined water areas (extensive and semi-intensive aquaculture in lagoons, earth ponds).