During recent decades, European eels (Anguilla anguilla) have decreased to historically low numbers. This follows a combination of overexploitation, loss of habitats, and transfer of parasites and diseases.1–3 Since 2010, A anguilla is considered critically endangered on the International Union for Conservation of Nature Red List of Endangered Species.4 European eels spend most (5 to 25 years) of their lives in freshwater or brackish coastal areas and then move to European seawaters to start the long migration to the Sargasso Sea for breeding. They are found in virtually all coastal areas, rivers, estuaries, and lakes in Europe and are believed to reproduce only once during their lifetime.5 Their life cycle is highly complex and features an intricate hormonal control mechanism that inhibits sexual maturation when the eels are in European waters.6,7 Successful reproduction of captive eels has been limited by variability in responsiveness to hormonal treatment, poor offspring quality, and extensive challenges in larval rearing.8–11 Thus, commercial farming of this valuable species is still based exclusively on wild-caught eels of the first juvenile stage (ie, glass eels).3,12,13
The high demand for aquaculture production in combination with the poor condition of eel stock has led to various research initiatives aimed at conserving the species, monitoring and regulating the yield, and developing a self-sustaining eel culture.5,8,14,15 For both ecological and economic reasons, there is a pressing need to acquire a better understanding of reproductive physiology and the response to hormonal interventions in European eels that can lead to sustainable eel production. Recent advances in assisted reproduction and rearing techniques have provided valuable information about the early life history of eels,16–18 which is a promising step toward sustainable aquaculture. Methods for hormonally assisted reproduction have been developed and established for Japanese eels (Anguilla japonica). Those methods include repeated treatment with salmon pituitary extract to induce vitellogenesis and maturation-inducing hormone to induce follicular maturation.19,20 During follicular maturation, oocytes substantially increase in size because of hydrolysis of the yolk protein, which results in a prominent increase in ovarian volume and body weight. At ovulation, the eggs are manually expressed from the female (ie, stripping) and artificially fertilized.16 However, the responsiveness to treatment varies substantially among females, thereby affecting reproductive outcome success.
Ultrasonographic imaging is sometimes used in aquaculture for determining sex,21–24 gonadal volume,25,26 and stage of sexual development.27 Ultrasonographic examinations are noninvasive and have no known long-term adverse effects.28 The technique is inexpensive and with suitable protocols can be performed when fish are restrained for routine sample collection. Information available from ultrasonographic images includes tissue area and echotexture.
Echotexture reflects the pattern of grayscale pixels that comprise an image. Terms such as fine or coarse are used to indicate an ultrasonographer's interpretation. Determining echotexture in clinical ultrasonography is a subjective process and requires skilled and experienced observers. Texture analysis refers to quantification of image features perceived as textural by an observer. These quantities can then be used objectively in statistical models.29 Texture analysis is widely used in human medical imaging.30–33 In addition to being objective, the generation of texture variables and their use in predictive statistical models can be fully automated or may be semiautomated (eg, an operator is required to trace an outline of ovarian tissue on the screen).
The purpose of the study reported here was to examine ultrasonographic predictors of reproductive outcome success derived from images acquired at 2 time points during assisted reproduction of female European eels. We hypothesized that a GSIus would differ among females with various degrees of response to hormonal treatment and that variables describing echotexture could be used to predict females with a good response to hormonal treatment and discriminate them from those that respond inadequately (ie, did not enter the final phase of maturation within 21 weeks after the start of treatment). Successful ultrasonographic prediction of reproductive outcome would directly enhance the efficiency of assisted reproduction in this species by eliminating slow-responding and nonresponding females, which would allow producers to focus resources on the best responders and most promising breeders.
Materials and Methods
Animals
A total of 83 female European eels, raised from the glass eel stage at a commercial fish farm,a were included in the study. Before the start of the experiment, the eels were transferred to a research facility.b Eels were anesthetized; they were placed in an aqueous solution of ethyl p-aminobenzoatec (20 mg/L) for 2 minutes. Anesthetized eels were weighed and measured, and a permanent passive integrated transponder tag was inserted IM approximately 1 cm ventral to the cranial attachment of the dorsal-caudal fin. Eels (mean ± SD length, 73 ± 4 cm; mean body weight, 852 ± 152 g) were maintained at low light intensity (20 lux), salinity of 36 g of solutes/L, and temperature of 20°C. Eels were allowed to acclimatize to saltwater over a 14-day period.
The experimental protocol for the study was approved by the Danish Animal Experiments Inspectorate, Ministry of Food, Agriculture and Fisheries. All fish were handled in accordance with the European Union regulations concerning the protection of experimental animals (Dir 86/609/EEC).
Experimental procedures
Each eel received weekly IM injections of salmon pituitary extractd (18.75 mg/kg [0.05 to 0.5 mL]/female)8,19 to induce ovarian development (first injection was designated as week 1). The extract was injected into the epaxial muscles approximately 1 cm ventral to the base of the dorsal fin. Weekly treatments continued until females that responded well entered the final maturation phase (weeks 14 to 20); females that did not respond adequately were treated for a maximum of 21 weeks. Females that responded to treatment, indicated by a weight increase > 10% of initial body weight, received an additional injection of salmon pituitary extract followed by a single treatment with maturation-inducing hormonee (2 mg/kg [0.5 to 1.5 mL]/female).8,20 The maturation-inducing hormone was injected into the cranial, middle, and caudal portions of the ovaries, which were identified by palpation. This method of administration distributed the maturation-inducing hormone throughout the ovaries to induce final follicular maturation and ovulation.16 Timing of the maturation-inducing hormone injection was adjusted on the basis of the stage of oocyte development (oocyte stage 4 or 5 [considered ideal] on a scale of 0 to 5 reported elsewhere34). Development stage was determined on the basis of an ovarian biopsy specimen obtained through the hypaxial muscle approximately 5 cm cranial to the urogenital pore and 2 cm lateral to the ventral midline34 from anesthetized eels (eels were anesthetized as described previously).
Image acquisition
Ultrasonographic images of the ovaries of the eels were obtained at 2 time points (weeks 7 and 11). Ultrasonographic imaging was performed at the research facility. One investigator (FJM) obtained all images and optimized the images during ultrasonography. A portable ultrasound machinef with an 18-MHz linear-array transducer was used. The transducer surface was covered in coupling gel and placed in a thin plastic bag. The bag was sealed and fixed to the transducer cable with elastic bands. Eels were anesthetized as described previously, placed in dorsal recumbency on a moist towel, and kept moist during imaging to allow cutaneous respiration. The transducer was then placed on the ventral body wall 3 cm cranial to the urogenital pore. Transverse ultrasonographic images of the body cavity were obtained, and the largest possible cross-sectional area at this location was recorded. Typical penetration depth was 4 cm. Gain varied between 52% and 79%. All recordings were grayscale (256 shades of gray) B-mode images. Imaging time was approximately 40 seconds. Both static images and video clips were saved on the hard disk of the ultrasound machine. Images were transferred to CD-ROMs and were later accessed by use of open source software,g sorted, and exported as DICOM (ie, Digital Imaging and Communication in Medicine) files. After eels were examined, they were transferred back to their home tanks to recover from anesthesia.
Collection of ovarian samples
Eels were euthanized at the completion of the study. Those that responded well were euthanized after ovulation and stripping, and those with an inadequate response were euthanized after 21 weeks of injections. Eels were anesthetized (placed in an aqueous solution of ethyl p-aminobenzoatec [20 mg/L] for 15 minutes) and then decapitated. Ovaries were then harvested and ovarian tissue samples preserved in phosphate-buffered 4% formaldehyde solution for subsequent histologic examination.
Histologic examination and classification of response
Ovary samples were dehydrated in ethanol and xylene by use of standard procedures, embedded in paraffin, cut in 4-µm-thick sections, and stained with H&E stain. Microscopic evaluation was used to classify eels into 3 groups (fast responders, slow responders, and nonresponders) on the basis of the most developed follicular stage. Fast responders had final maturation oocytes in the hydration stage and postovulatory follicles, slow responders had vitellogenic oocytes, and nonresponders had previtellogenic oocytes (Figure 1).
Photomicrographs of ovarian tissues obtained from representative European eels (Anguilla anguilla) treated with IM injections of pituitary extract and that were classified as a fast responder with signs of completed vitellogenesis and ovulation (A), a slow responder with evidence of progression of vitellogenesis (B), or a nonresponder with previtellogenic stages only (C). Notice that the magnification for panel C is higher than for panels A and B to enable identification of the cell types in the undeveloped ovarian tissue. FOM = Final oocyte maturation. PG = Primary growth oocytes in previtellogenic stage. POF = Postovulatory follicles. VT = Vitellogenic oocytes. H&E stain; bars = 500 μm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Photomicrographs of ovarian tissues obtained from representative European eels (Anguilla anguilla) treated with IM injections of pituitary extract and that were classified as a fast responder with signs of completed vitellogenesis and ovulation (A), a slow responder with evidence of progression of vitellogenesis (B), or a nonresponder with previtellogenic stages only (C). Notice that the magnification for panel C is higher than for panels A and B to enable identification of the cell types in the undeveloped ovarian tissue. FOM = Final oocyte maturation. PG = Primary growth oocytes in previtellogenic stage. POF = Postovulatory follicles. VT = Vitellogenic oocytes. H&E stain; bars = 500 μm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Photomicrographs of ovarian tissues obtained from representative European eels (Anguilla anguilla) treated with IM injections of pituitary extract and that were classified as a fast responder with signs of completed vitellogenesis and ovulation (A), a slow responder with evidence of progression of vitellogenesis (B), or a nonresponder with previtellogenic stages only (C). Notice that the magnification for panel C is higher than for panels A and B to enable identification of the cell types in the undeveloped ovarian tissue. FOM = Final oocyte maturation. PG = Primary growth oocytes in previtellogenic stage. POF = Postovulatory follicles. VT = Vitellogenic oocytes. H&E stain; bars = 500 μm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Data analysis
For all analyses, 1 investigator (AVM) performed manual segmentation of the ovaries on the ultrasonographic images. This investigator was not aware of the classification of each particular eel. An area as large as possible that included only ovarian tissue (excluded vessels, nonovarian tissues, and obvious artifacts) was identified and used for analysis. Cross-sectional area of the ovary was determined from the number of pixels in the identified area. Each area was then normalized by dividing the area of the ovary by the size of the fish (estimated as fish length3). This created a GSIus. An ANOVA was then performed to compare the mean GSIus for the 3 groups. Test assumptions were met.
Patterns of pixel values between adjacent pixels on ultrasonographic images of the ovaries were examined. This was performed by creating GLCMs.35 These matrices can be used to derive a number of texture features. Texture features in ovarian ultrasonographic images were derived from GLCMs by use of an established method that is based on gray-level frequency between adjacent pixels with well-defined spatial relationships.35–37 A GLCM is a method for measuring interpixel relations in an image. It is a square matrix (G) of order N, where N is the number of grayscale values (256). The (i,j)th entry of G represents the number of times a pixel with intensity i is adjacent to a pixel with intensity j. Features for 22 texture variables were calculated from the GLCMs. These features have been described in detail in studies involving evaluation of 13 features,35 5 features,36 and 4 features.37 Statistical analyses were performed with GSIus and texture features alone and in combination.
The relationships of texture features to each other and to reproductive success (ie, classification of response to hormonal treatment) were included in a PCA.38 This allowed an analysis of variability to determine whether certain image features or groups of image features were related to treatment response (outcome classification group). In a separate analysis, the GSIus, texture features, and treatment response were included in a PCA.
Finally, a PLS-DA (a machine learning classifier system based on logistic regression) was performed39,40 by use of all texture features and reproductive outcome groups (fast responder, slow responder, or nonresponder). Leave-one-out cross-validation was used to test the PLS-DA model.
Data processing was performed with a commercial computing software package.h Significance was set at values of P < 0.05.
Results
Histologic classification
Fish were classified among the 3 groups on the basis of ovarian histologic findings. There were 57 fast responders, 18 slow responders, and 8 nonresponders.
Ultrasonography
The ovaries were successfully identified ultrasonographically in all eels (Figure 2). Contrast with adjacent tissue was sufficient to allow manual segmentation (Figure 3).
Ultrasonographic images of the ovaries of a fast responder eel after IM administration of salmon pituitary extract for 7 weeks (A) and 11 weeks (B). The eel was positioned in dorsal recumbency, and the transducer was placed on the ventral body wall 3 cm cranial to the urogenital pore. Skin of the ventral body wall is at the top of the image. The ovaries are the relatively hypoechogenic structure in the middle of the image and are surrounded by fat and muscle tissue. Bar = 1 cm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Ultrasonographic images of the ovaries of a fast responder eel after IM administration of salmon pituitary extract for 7 weeks (A) and 11 weeks (B). The eel was positioned in dorsal recumbency, and the transducer was placed on the ventral body wall 3 cm cranial to the urogenital pore. Skin of the ventral body wall is at the top of the image. The ovaries are the relatively hypoechogenic structure in the middle of the image and are surrounded by fat and muscle tissue. Bar = 1 cm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Ultrasonographic images of the ovaries of a fast responder eel after IM administration of salmon pituitary extract for 7 weeks (A) and 11 weeks (B). The eel was positioned in dorsal recumbency, and the transducer was placed on the ventral body wall 3 cm cranial to the urogenital pore. Skin of the ventral body wall is at the top of the image. The ovaries are the relatively hypoechogenic structure in the middle of the image and are surrounded by fat and muscle tissue. Bar = 1 cm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Image analysis of the ultrasonographic image of the ovaries of the European eel in Figure 2 after hormonal treatment for 11 weeks. The ovarian tissue was segmented manually to identify the ovarian cross-sectional area (green region). The number of pixels in this region was used to determine the cross-sectional area. The cross-sectional area was then used to calculate both the GSIus and texture features. Bar = 1 cm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Image analysis of the ultrasonographic image of the ovaries of the European eel in Figure 2 after hormonal treatment for 11 weeks. The ovarian tissue was segmented manually to identify the ovarian cross-sectional area (green region). The number of pixels in this region was used to determine the cross-sectional area. The cross-sectional area was then used to calculate both the GSIus and texture features. Bar = 1 cm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Image analysis of the ultrasonographic image of the ovaries of the European eel in Figure 2 after hormonal treatment for 11 weeks. The ovarian tissue was segmented manually to identify the ovarian cross-sectional area (green region). The number of pixels in this region was used to determine the cross-sectional area. The cross-sectional area was then used to calculate both the GSIus and texture features. Bar = 1 cm.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Ultrasonography at week 7 typically revealed small ovaries located centrally in the body cavity. Ovarian area ranged from 7,882 to 31,007 pixels (0.57 to 2.27 cm2), with a mean of 16,802 pixels (1.23 cm2).
Ultrasonography at week 11 revealed wide variation in size of the ovaries. In 10 eels, the complete cross-sectional area of the ovaries was larger than the ultrasonographic image. Thus, the subsequent calculation of the cross-sectional area was underestimated in these cases. Ovarian area ranged from 6,007 to 64,194 pixels (0.44 to 4.69 cm2), with a mean of 27,825 pixels (2.01 cm2).
GSIus
Results for the ANOVA indicated that the mean GSIus differed significantly among the 3 groups at both week 7 (P = 0.006) and week 11 (P < 0.001). A large GSIus was associated with a positive response to hormonal treatment. However, there was overlap of the GSIus for the 3 groups at week 7 (Figure 4). By week 11, the difference in GSIus between the groups was more apparent. Mean GSIus of fast responders, compared with that of the combined slow responders and nonresponders, also differed significantly at both week 7 (P = 0.002) and week 11 (P < 0.001).
Notched box-and-whisker plots of the GSIus at week 7 (A) and week 11 (B) of hormonal treatment for European eels that were fast responders (n = 57), slow responders (18), or nonresponders (8). Each box represents the interquartile range, the horizontal line in the box represents the median, the upper and lower boundaries of the notch represent the 95% confidence interval for the median, the asterisk represents the mean, the whiskers represent the most extreme data points that are not considered outliers, and circles represent outliers. There was a significant difference among groups at week 7 (P = 0.006) and week 11 (P < 0.001).
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Notched box-and-whisker plots of the GSIus at week 7 (A) and week 11 (B) of hormonal treatment for European eels that were fast responders (n = 57), slow responders (18), or nonresponders (8). Each box represents the interquartile range, the horizontal line in the box represents the median, the upper and lower boundaries of the notch represent the 95% confidence interval for the median, the asterisk represents the mean, the whiskers represent the most extreme data points that are not considered outliers, and circles represent outliers. There was a significant difference among groups at week 7 (P = 0.006) and week 11 (P < 0.001).
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Notched box-and-whisker plots of the GSIus at week 7 (A) and week 11 (B) of hormonal treatment for European eels that were fast responders (n = 57), slow responders (18), or nonresponders (8). Each box represents the interquartile range, the horizontal line in the box represents the median, the upper and lower boundaries of the notch represent the 95% confidence interval for the median, the asterisk represents the mean, the whiskers represent the most extreme data points that are not considered outliers, and circles represent outliers. There was a significant difference among groups at week 7 (P = 0.006) and week 11 (P < 0.001).
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Echotexture and PCA
The GLCM records the frequency of occurrence of specific gray-value pairs. Gray-level values for 2 adjacent pixels were measured in each of 4 directions, which resulted in 4 GLCMs derived for each image (Figure 5). Evaluating adjacent pixels (the pixel horizontally to the right, the pixel vertically above, and the pixels diagonally vertical to the right and left, respectively) ensured that almost all pixel pairs were considered and no pairs were considered twice.
Plots of the frequency of gray-level values for adjacent pixels in spatial arrangements (the pixel horizontally to the right [A], the pixel vertically above [B], the pixel diagonally vertical to the right [C], and the pixel diagonally vertical to the left [D]) as calculated from the ovarian cross-sectional area shown in Figure 3. The colored region in each plot represents the combinations of adjacent pixel values in the image. The color bar (bottom right corner) indicates the frequency of the possible combinations of adjacent pixels.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Plots of the frequency of gray-level values for adjacent pixels in spatial arrangements (the pixel horizontally to the right [A], the pixel vertically above [B], the pixel diagonally vertical to the right [C], and the pixel diagonally vertical to the left [D]) as calculated from the ovarian cross-sectional area shown in Figure 3. The colored region in each plot represents the combinations of adjacent pixel values in the image. The color bar (bottom right corner) indicates the frequency of the possible combinations of adjacent pixels.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Plots of the frequency of gray-level values for adjacent pixels in spatial arrangements (the pixel horizontally to the right [A], the pixel vertically above [B], the pixel diagonally vertical to the right [C], and the pixel diagonally vertical to the left [D]) as calculated from the ovarian cross-sectional area shown in Figure 3. The colored region in each plot represents the combinations of adjacent pixel values in the image. The color bar (bottom right corner) indicates the frequency of the possible combinations of adjacent pixels.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Twenty-two texture features were calculated for each GLCM for each image. Texture features of all images, together with the classification of response to hormonal treatment, were included in a PCA. In this procedure, data were transformed into sets of linearly uncorrelated variables or principal components. The PC1 was obtained by projecting the data onto a linear plane selected to maximize variance of the projected data; PC1 accounted for some, but not all, variance in the data. The PC2 and subsequent principal components were selected with the purpose of maximizing variance of the projected data with the constraint that each new principal component was orthogonal to (not correlated with) the previous components. For the data obtained for the eels, combined results for the first 4 principal components explained 93.6% and 93% of the variance in the data for week 7 and week 11 images, respectively. The PCA for both imaging dates revealed that the 4 GLCM directions resulted in extremely similar results. Therefore, only 1 direction for adjacent pixels (the pixel to the right) was chosen for further analysis. Analysis of images obtained at week 7 revealed that PC1 explained 46.75% of the variance and PC2 explained 25.39% of the variance. Similarly, at week 11, PC1 explained 47.11% of the variance and PC2 explained 24.31% of the variance. The PCA scores in this analysis represented the samples (ie, the images analyzed), and the PCA loadings represented the texture features. Results were plotted (Figure 6); proximity on either plot indicated similarity. Even though there was large overlap between the scores, there was some separation among the 3 classification groups. Loadings were distributed similarly on both dates of ultrasonography. This implied that the overall change in texture was small between the 2 ultrasonographic images.
Plots of results of PCA (scores for ovarian ultrasonographic images [A and C] and loadings for texture features [B and D]) for data obtained at week 7 (A and B) and week 11 (C and D) from European eels that were fast responders (gray diamonds), slow responders (black triangles), or nonresponders (gray squares) to hormonal treatment. The top and bottom percentages in each plot indicate the variation explained by the principal component on the y- and x-axis, respectively. Loadings in panels B and D represent texture features and are described elsewhere.35–37
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Plots of results of PCA (scores for ovarian ultrasonographic images [A and C] and loadings for texture features [B and D]) for data obtained at week 7 (A and B) and week 11 (C and D) from European eels that were fast responders (gray diamonds), slow responders (black triangles), or nonresponders (gray squares) to hormonal treatment. The top and bottom percentages in each plot indicate the variation explained by the principal component on the y- and x-axis, respectively. Loadings in panels B and D represent texture features and are described elsewhere.35–37
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Plots of results of PCA (scores for ovarian ultrasonographic images [A and C] and loadings for texture features [B and D]) for data obtained at week 7 (A and B) and week 11 (C and D) from European eels that were fast responders (gray diamonds), slow responders (black triangles), or nonresponders (gray squares) to hormonal treatment. The top and bottom percentages in each plot indicate the variation explained by the principal component on the y- and x-axis, respectively. Loadings in panels B and D represent texture features and are described elsewhere.35–37
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
In addition, a separate PCA was performed for both dates of ultrasonography that included texture features and GSIus. The addition of the nontexture variable (ie, GSIus) did not cause significant changes to the PCA plots, which suggested that the GSIus did not add additional information to this texture model because it influenced all texture features equally.
PLS-DA
The PLS-DA calculates a prediction probability and classification threshold by use of a multivariate discriminant method that models the covariance between 2 matrices, × and Y.39–41 In this case, matrix × represented the texture features in the ultrasonographic images and matrix Y represented the classifications of response to hormonal treatment. The analysis involved at least 4 latent variables40 and a leave-one-out cross-validation. The model was used to classify eels as fast responders, slow responders, or nonresponders on the basis of texture features (Figure 7). For week 7, classification error in the cross-validation test ranged from 0.304 to 0.329, whereas classification error for week 11 ranged from 0.172 to 0.387 (Table 1).
Plots of PLS-DA (scores for ovarian ultrasonographic images [A and C] and loadings for texture features [B and D]) for data obtained at week 7 (A and B) and week 11 (C and D) from European eels that were fast responders, slow responders, or nonresponders to hormonal treatment. The top and bottom percentages in each plot indicate the variation explained by the latent variable on the y- and x-axis, respectively. Loadings in panels B and D represent texture features and are described elsewhere.35–37 See Figure 6 for remainder of key.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Plots of PLS-DA (scores for ovarian ultrasonographic images [A and C] and loadings for texture features [B and D]) for data obtained at week 7 (A and B) and week 11 (C and D) from European eels that were fast responders, slow responders, or nonresponders to hormonal treatment. The top and bottom percentages in each plot indicate the variation explained by the latent variable on the y- and x-axis, respectively. Loadings in panels B and D represent texture features and are described elsewhere.35–37 See Figure 6 for remainder of key.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Plots of PLS-DA (scores for ovarian ultrasonographic images [A and C] and loadings for texture features [B and D]) for data obtained at week 7 (A and B) and week 11 (C and D) from European eels that were fast responders, slow responders, or nonresponders to hormonal treatment. The top and bottom percentages in each plot indicate the variation explained by the latent variable on the y- and x-axis, respectively. Loadings in panels B and D represent texture features and are described elsewhere.35–37 See Figure 6 for remainder of key.
Citation: American Journal of Veterinary Research 77, 5; 10.2460/ajvr.77.5.478
Results for a PLS-DA by use of 4 latent variables40 and leave-one-out cross-validation for texture features of ovarian ultrasonographic images obtained from 83 European eels (Anguilla anguilla) on the basis of reproductive outcome after IM administration of salmon pituitary extract.
| Time | Variable | Method | Fast responder | Slow responder | Nonresponder |
|---|---|---|---|---|---|
| Week 7 | Sensitivity | Calibration | 0.836 | 0.579 | 0.750 |
| Sensitivity | Cross-validation | 0.800 | 0.579 | 0.625 | |
| Specificity | Calibration | 0.630 | 0.841 | 0.730 | |
| Specificity | Cross-validation | 0.593 | 0.810 | 0.716 | |
| Classification error | Calibration | 0.267 | 0.290 | 0.260 | |
| Classification error | Cross-validation | 0.304 | 0.306 | 0.329 | |
| Week 11 | Sensitivity | Calibration | 0.911 | 0.500 | 0.750 |
| Sensitivity | Cross-validation | 0.875 | 0.667 | 0.750 | |
| Specificity | Calibration | 0.615 | 0.812 | 0.905 | |
| Specificity | Cross-validation | 0.615 | 0.703 | 0.919 | |
| Classification error | Calibration | 0.237 | 0.344 | 0.172 | |
| Classification error | Cross-validation | 0.264 | 0.387 | 0.172 |
Fast responder, slow responder, and nonresponder represent reproductive outcome determined on the basis of histologic evaluation of ovarian tissues to determine the most developed follicular stage. Fast responders had final maturation oocytes in the hydration stage and postovulatory follicles, slow responders had vitellogenic oocytes, and nonresponders had previtellogenic oocytes.
Discussion
The study reported here revealed that ultrasonography was a feasible modality for use in aquaculture and, specifically, that it could be used to evaluate the ovaries of European eels.26 In addition, we found that ultrasonographic imaging and analysis could contribute to improved procedures for assisted eel reproduction. A particular strength of this study was that predictions were based on objective texture features derived from image analysis, rather than on subjective observations.
Gonadosomatic index is a variable commonly used in aquaculture research and production. Results for the present study indicated that the GSIus could indicate the response to hormonal treatment by week 7 after the start of hormonal treatments. The difference in GSIus among groups was more apparent by week 11, which was approximately the midpoint of the maximum treatment period of 21 weeks.
The ultrasonographic images obtained at week 11 did not provide the complete cross-sectional view of the ovaries in 10 eels because the ovaries were larger than the ultrasonographic image. Thus, the predictive value of the GSIus for week 11 likely was underestimated. However, even with estimated values, the objective area measurements contributed to a predictive model.
With regard to texture features, we determined that there were signals in the ovarian images that could be useful in identifying females that respond inadequately to treatment. Use of 22 texture features will yield a model that by definition will be multivariate. In the model reported here, classification error rate was between 17% and 39% (Table 1). Interpretation of these rates must take into account the error of the reference method, and we estimated the error of the histologic classification was < 5%. Error rates in the discrimination analysis were clearly larger than the error of the histologic classification.
Ultrasonography can provide 2 important contributions to assisted eel reproduction. First, the classification of the eels during treatment can be predicted noninvasively. Second, the prediction of the response can be made early in the treatment period. Elimination of nonresponders at an early stage in particular would save resources by ensuring that personnel and expensive hormones are not expended on fish that eventually will not successfully reproduce.
The prediction method described in the present study was complex but affordable, and its application was semiautomated; the user only provided the images and manually segmented the ovaries. For the discrimination to be useful, a cost-benefit assessment must be performed to determine whether the exclusion of predicted nonresponders and slow responders is worth the risk of also excluding eels with false-negative results.
The main limitation of the present study was the limited size of the study population. The results indicated a good probability that texture features can be used as part of a classification process. Because of the relatively small number of eels used, compared with the number of variables, larger studies involving texture features and reproductive outcomes are needed to provide a classification model that can be used as a practical tool. It is not unusual in machine learning scenarios to allow continuous training by providing feedback to the model so that the model's performance improves with each additional iteration.
In the present study, the classification slow responder encompassed a wide span of developmental stages from almost no response to almost complete vitellogenesis. We considered the possibility that this class should be subdivided but thought that such subdivision would not be helpful in this case because it would result in unacceptably small group sizes.
We did not use standardized settings for the ultrasound machine because the nature of ultrasonography is that settings are often adjusted to yield optimal image acquisition. It is doubtful the use of variations in settings affected the texture analysis because the analysis methods were robust for differences in settings such as depth, image gain, and display brightness.
In the study reported here, ultrasonography was feasible for use in monitoring ovarian development in European eels and potentially other fish species in breeding programs. The ovarian cross-sectional area increased with progressive maturation, and the GSIus was used to indicate at an early stage of hormonal treatment whether an eel would be a fast responder, slow responder, or nonresponder. The study also revealed that texture analysis of the ultrasonographic images yielded a predictive value for discrimination among the 3 groups. Large data sets are required to build robust predictive models by use of the methods described for the study reported here. We suggest that texture information and morphometric variables should both be considered in future ultrasonographic applications for assisted reproduction of eels and potentially other fish species. The study reported here was, to our knowledge, the first in which a rigorous objective texture analysis was performed on the gonads of fish. The described approach to texture analysis could be applied more generally to standard veterinary imaging scenarios with ultrasonography and other imaging modalities in which image texture is of potential interest.
Acknowledgments
Presented as a poster at the Visionday Conference at the Technical University of Denmark, Kgs. Lyngby, Denmark, May 2014.
Supported by the 7th Framework Programme of the European Commission under Food, Agriculture and Fisheries, and Biotechnology (grant agreement No. 245257) and the Chemometric Analysis Center (CHANCE) at the University of Copenhagen, Copenhagen, Denmark.
The authors declare that there were no conflicts of interest.
ABBREVIATIONS
| GLCM | Gray-level co-occurrence matrix |
| GSIus | Ultrasonographic-derived gonadosomatic index |
| PCA | Principal component analysis |
| PC1 | First principal component |
| PC2 | Second principal component |
| PLS-DA | Partial least squares discriminant analysis |
Footnotes
Stensgård Eel Farm A/S, Randbøl, Denmark.
Lyksvad Fish Farm, Vamdrup, Denmark.
Benzocaine, Sigma-Aldrich Corp, Munich, Germany.
SPE, Argent Chemical Laboratories, Redmond, Wash.
DHP, 17α,20β-dihydroxy-4-pregnen-3-one, Sigma-Aldrich Corp, St Louis, Mo.
Mylab 30 vet, Esaote SpA, Genoa, Italy.
OsiriX, version 5.0, Pixmeo, Geneva, Switzerland.
MATLAB, version 2012a, The Mathworks Inc, Natick, Mass.
References
1. International Council for the Exploration of the Sea (ICES) Advisory Committee. Report of the Joint EIFAAC/ICES/GFCM Working Group on Eels. ICES CM 2014/ACOM:18. Copenhagen: ICES, 2014.
2. International Council for the Exploration of the Sea (ICES) Advisory Committee. Report of the Joint EIFAAC/ICES Working Group on Eels (WGEEL). ICES CM 2013/ACOM:18. Copenhagen: ICES, 2013.
3. Dekker W Did lack of spawners cause the collapse of the European eel, Anguilla anguilla? Fish Manag Ecol 2003; 10: 365–376.
4. Jacoby D, Gollock M. Anguilla anguilla. The IUCN red list of threatened species. Available at: dx.doi.org/10.2305/IUCN.UK.2014-1.RLTS.T60344A45833138.en. Accessed Nov 25, 2015.
5. Moriarty C, Dekker W. Management of the European eel. Irish Fisheries bulletin No. 15. Dublin: Marine Institute, Fisheries Research Centre, 1997.
6. Dufour S, Burzawa-Gerard E, Le Belle N, et al. Reproductive endocrinology of the European eel, Anguilla anguilla. In: Aida K, Tsukamoto K, Yamauchi K, eds. Eel biology. Tokyo: Springer, 2003; 373–383.
7. Pasquier J, Lafont AG, Leprince J, et al. First evidence for a direct inhibitory effect of kisspeptins on LH expression in the eel, Anguilla anguilla. Gen Comp Endocrinol 2011; 173: 216–225.
8. Tomkiewicz J, Tybjerg L, Støttrup JG, et al. Reproduction of European Eel in Aquaculture (REEL): consolidation and new production methods. DTU Aqua report 249–2012. Charlottenlund, Denmark: National Institute of Aquatic Resources, Technical University of Denmark, 2012.
9. Pedersen BH. Fertilisation of eggs, rate of embryonic development and hatching following induced maturation of the European eel Anguilla anguilla. Aquaculture 2004; 237: 461–473.
10. van Ginneken V, Vianen G, Muusze B, et al. Gonad development and spawning behavior of artificially-matured European eel (Anguilla anguilla L.). Anim Biol 2005; 55: 203–218.
11. Pedersen BH. Induced sexual maturation of the European eel Anguilla anguilla and fertilisation of the eggs. Aquaculture 2003; 224: 323–338.
12. ICES Advisory Committee on Fisheries. Report of the 2006 session of the Joint EIFAC/ICES Working Group on Eels. ICES CM 2006/ACFM:16. Copenhagen: ICES, 2006.
13. Kamstra A. Eel culture. In: Tesch FW, ed. The eel. 3rd ed. Oxford, England: Blackwell Science Ltd, 2003; 295–306.
14. International Council for the Exploration of the Sea (ICES) Advisory Committee. Report of the ICES Advisory Committee 2013. ICES Advice, 2013. Book 9. Copenhagen: ICES, 2013.
15. Dekker W. A Procrustean assessment of the European eel stock. ICES J Mar Sci 2000; 57: 938–947.
16. Butts IAE, Sørensen SR, Politis SN, et al. Standardization of fertilization protocols for the European eel, Anguilla anguilla. Aquaculture 2014; 426–427: 9–13.
17. Politis SN, Butts IAE, Tomkiewicz J. Light impacts embryonic and early larval development of the European eel, Anguilla anguilla. J Exp Mar Biol Ecol 2014; 461: 407–415.
18. Sørensen SR, Skov PV, Lauesen P, et al. Microbial interference and potential control in culture of European eel (Anguilla anguilla) embryos and larvae. Aquaculture 2014; 426–427: 1–8.
19. Kagawa H, Tanaka H, Ohta H, et al. The first success of glass eel production in the world: basic biology on fish reproduction advances new applied technology in aquaculture. Fish Physiol Biochem 2005; 31: 193–199.
20. Ohta H, Kagawa H, Tanaka H, et al. Changes in fertilization and hatching rates with time after ovulation induced by 17, 20β-dihydroxy-4-pregnen-3-one in the Japanese eel, Anguilla japonica. Aquaculture 1996; 139: 291–301.
21. McEvoy FJ, Tomkiewicz J, Støttrup JG, et al. Determination of fish gender using fractal analysis of ultrasound images. Vet Radiol Ultrasound 2009; 50: 519–524.
22. Martin-Robichaud DJ, Rommens M. Assessment of sex and evaluation of ovarian maturation of fish using ultrasonography. Aquacult Res 2001; 32: 113–120.
23. Kohn YY, Lokman PM, Kilimnik A, et al. Sex identification in captive hapuku (Polyprion oxygeneios) using ultrasound imagery and plasma levels of vitellogenin and sex steroids. Aquaculture 2013; 384–387: 87–93.
24. Novelo ND, Tiersch TR. A review of the use of ultrasonography in fish reproduction. North Am J Aquaculture 2012; 74: 169–181.
25. Bryan JL, Wildhaber ML, Papoulias DM, et al. Estimation of gonad volume, fecundity, and reproductive stage of shovelnose sturgeon using sonography and endoscopy with application to the endangered pallid sturgeon. J Appl Ichthyol 2007; 23: 411–419.
26. Bureau du Colombier S, Jacobs L, Gesset C, et al. Ultrasonography as a non-invasive tool for sex determination and maturation monitoring in silver eels. Fish Res 2015; 164: 50–58.
27. Newman DM, Jones PL, Ingram BA. Sexing accuracy and indicators of maturation status in captive Murray cod Maccullochella peelii peelii using non-invasive ultrasonic imagery. Aquaculture 2008; 279: 113–119.
28. Fowlkes JB. Non-thermal effects of diagnostic ultrasound. In: ter Haar G, ed. The safe use of ultrasound in medical diagnosis. 3rd ed. London: The British Institute of Radiology, 2012; 166.
29. Allison JW, Barr LL, Massoth RJ, et al. Understanding the process of quantitative ultrasonic tissue characterization. Radiographics 1994; 14: 1099–1108.
30. Chen DR, Chang RF, Kuo WJ, et al. Diagnosis of breast tumors with sonographic texture analysis using wavelet transforms and neural networks. Ultrasound Med Biol 2002; 28: 1301–1310.
31. Tenorio V, Bonet-Carne E, Botet F, et al. Correlation between a semiautomated method based on ultrasound texture analysis and standard ultrasound diagnosis using white matter damage in preterm neonates as a model. J Ultrasound Med 2011; 30: 1365–1377.
32. Häberle L, Wagner F, Fasching PA, et al. Characterizing mammographic images by using generic texture features. Breast Cancer Res 2012; 14: R59.
33. Gao S, Peng Y, Guo H, et al. Texture analysis and classification of ultrasound liver images. Biomed Mater Eng 2014; 24: 1209–1216.
34. Palstra AP, Cohen EGH, Niemantsverdriet PRW, et al. Artificial maturation and reproduction of European silver eel: development of oocytes during final maturation. Aquaculture 2005; 249: 533–547.
35. Haralick RM, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern 1973; SCM-3: 610–621.
36. Soh LK, Tsatsoulis C. Texture analysis of SAR sea ice imagery using gray level co-occurrence matrices. IEEE Trans Geosci Remote Sens 1999; 37: 780–795.
37. Clausi DA. An analysis of co-occurrence texture statistics as a function of grey level quantization. Can J Remote Sens 2002; 28: 45–62.
38. Wold S, Esbensen K, Geladi P. Principal component analysis. Chemometr Intell Lab Syst 1987; 2: 37–52.
39. Nocairi H, Qannari EM, Vigneau E, et al. Discrimination on latent components with respect to patterns. Application to multicollinear data. Comput Stat Data Anal 2005; 48: 139–147.
40. Wold S, Sjöström M, Eriksson L. PLS-regression: a basic tool of chemometrics. Chemometr Intell Lab Syst 2001; 58: 109–130.
41. Barker M, Rayens W. Partial least squares for discrimination. J Chemom 2003; 17: 166–173.
