| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
ANIMAL PRODUCTS |
Department of Animal Science, University of Padova, viale dell’Università 16, 35020 Legnaro (PD), Italy
Abstract
The aims of this study were 1) to apply computer image analysis to obtain measures of lean and fatty areas on the cross section of dry-cured hams, 2) to investigate variation of these measures, and 3) to evaluate reproducibility and repeatability of these techniques. Traits of concern were the cross-sectional area (SA), lean, or muscles, area (LA), and the fatty area (FA) centered on the cross section and surrounded by biceps femoris, semimembranosus, semitendinosus, and quadriceps femoris, as well as the FA-to-SA ratio (FESR). Hams were obtained from crossbred pigs (n = 279) slaughtered at 9 mo of age (mean BW of 169 ± 17 kg). Digital images of the cross section of dry-cured hams were captured using standardized procedures. Three replicated measures of areas were collected by three operators using three image analysis techniques (automatic, automatic-assisted, and manual). Variance components were estimated using a linear model that included slaughter group, gender, and gender x slaughter group as fixed effects and operators, pig, and operator x pig as random effects. Statistical analyses considered all measures (n = 7,533) or measures collected after reinstruction of all operators for spatial calibration of the analysis system (n = 4,428). Average SA, LA, FA, and FESR were 350 cm2, 220 cm2, 8.7 cm2, and 2.5%, respectively. Variability of FA (CV = 42%) and of FESR (CV = 39%) was four times greater than that of SA and LA. Slaughter group, pig, operator, and operator x pig effects were the most (P < 0.01) important sources of variation of measures. Correlations between measures obtained with different techniques were greater (P < 0.01) than 0.90, with the exception of LA measures. Coefficients of reproducibility for SA and LA ranged from 87 to 94%, whereas those for FA and FESR ranged from 88 to 98%. Coefficients of repeatability ranged from 92 to 99%. Automatic-assisted and manual methods provided more reproducible and repeatable measures than the automatic technique. Spatial calibration of the software system was a key issue affecting reproducibility and repeatability. Reinstructing the analysts for spatial calibration enhanced both reproducibility and repeatability of all methods of analysis. Computer image analysis is a technique suitable for measuring lean and fatty areas in cross-sectioned hams, providing reproducible and repeatable measures, and it might be used in large sample-based studies to investigate causes of defective fatty areas.
Key Words: Computer Image Analysis • Fat • Ham • Pigs • Repeatability • Reproducibility
Introduction
Dry-cured ham is the most valuable product of the pig industry in Italy, and raw hams account for more than 50% of carcass market value (Gigli et al., 1993
). Characteristics of fatty areas on the cross section of dry-cured hams have been recently considered in relation to enhancement of ham quality. Wide, visible fatty areas are a major defect that affects the acceptability of dry-cured ham by the consumer (Lo Fiego et al., 2000), and fat deposition within the ham affects the quality of processed meat (Gandemer, 2002; Virgili and Schivazappa, 2002).
Difficulties in carrying out studies on these traits arise from experimental costs due to the loss of product value and from the unavailability of standard techniques for measuring areas of interest. Computer image analysis (CIA) offers the opportunity of objectively assessing the visual appearance of the cross section of dry-cured hams (Cernadas et al., 2002) and has been already used for measuring carcass and meat traits in cattle and pigs (Swatland, 1995; Shackelford et al., 1998; Tan et al., 2000). Advantages of CIA are low experimental costs due to reduced loss in ham value, fast data acquisition, suitability for analysis of large samples, and applicability in commercial plants. Although measuring areas of the cross section of hams by CIA techniques is feasible, there is no current knowledge on the effects that replicated measures and multiple operators exert on reliability. Moreover, indications on the agreement between measures obtained with CIA procedures exhibiting different degrees of automation and processing time are not available, and the influence of the degree of automation on the precision of measures is unknown. Therefore, this study aimed to 1) apply CIA techniques to obtain measures of lean and fatty areas from digital images of cross-sectioned dry-cured hams, 2) investigate variation of these measures, and 3) evaluate reproducibility and repeatability of these techniques.
Materials and Methods
Animals
Raw hams were collected from 279 crossbred pigs (9 mo of age, with an average BW of 169 ± 17 kg) slaughtered on a single day each month (10 slaughter groups) at the same commercial abattoir (Montorsi, Correggio, Italy). Pigs (142 gilts and 137 barrows) were offspring of nine boars of the C21 Large White line (Gorzagri, Fonzaso, Italy) mated to 36 Large White-derived crossbred sows. Number of slaughtered pigs and the ratio of barrows to gilts within each slaughter group (SB) ranged from 24 to 30 and from 0.76 to 1.31, respectively. Hams were removed from carcasses according to a standard methodology (Associazione Scientifica di Produzione Animale, 1991), and all left hams were cured according to the San Daniele procedure (D.O.P. Prosciutto di San Daniele, 1996).
Image Acquisition
Before image capture, bones were removed from dry-cured hams with no fat removal, and hams were cross-sectioned by a cut made at approximately the same anatomical location. Relative to the head of femur, the transverse section was made at approximately one-third the length of the femur. The smallest portion of each cross-sectioned ham was placed in a ham clamp covered by a green cloth to create a homogeneous background. A nonglare 25.16-mm metallic black bar was placed marginally on the subcutaneous fat area and was used as reference object for spatial calibration of the computer image analysis (CIA) software.
Images were captured using a digital color camera (model Coolpix 950; Nikon Corp., Tokyo, Japan) equipped with a 3x aspheric zoom (from 7 to 21 mm) lens. The camera was mounted on a tripod such that the tip of the lens was approximately 70 cm from the upper margin of the clamp, and the camera was perpendicular to the cross-section surface. A bluish light (45 W) was placed on the tripod and directed with an angle of 45° to the cross-section being photographed to avoid irregular reflection on the surface of the ham. Images were captured using the autofocus function. Camera shutter speed and aperture were set at 1/8 s and f 3.1, respectively. Images were captured in JPEG format (1,600 x 1,200 pixels), stored on the camera card, and transferred to a workstation to create a library suitable for image processing.
Image Processing and Analysis
Digital images were analyzed using Image Pro Plus 4.1 (Media Cybernetics, Silver Spring, MD). Areas of interest included 1) the total area of the cross section; 2) the fat eye area (FA; a visible fatty area approximately centered on the cross section and surrounded by biceps femoris, semimembranosus, semitendinosus, and quadriceps femoris muscles); and 3) the lean, or muscles, area (LA; the area of the cross section that excluded the area of subcutaneous fat, fat eye, and skin). Furthermore, the ratio of the FA to the cross-sectional area (FESR) was computed and used in the data analysis.
To convert measures from pixels to millimeters, the spatial calibration of the software system (Media Cybernetics, Silver Spring, MD) was a mandatory step of the analysis of digital images. The reference object was traced by the operator, and its known length (25.16 mm) was fed into the software program preliminarily to each image processing. Image analysis was performed using three methods (automatic, automatic-assisted, and manual), which differed for the degree of involvement of the operator in outlining areas of interest and, consequently, the amount of time required for image processing.
For the automatic procedure, images (originally captured in RGB format) were converted using the monochromatic (grayscale) model with the aim of enhancing the contrast between fat and lean areas. Areas of concern were measured automatically after the area under computation was encompassed with a bounding box by the analyst. For the manual procedure, a contour line outlining the area of interest was traced manually by the analyst using a mouse and a graphic tablet, and areas were measured by the software program. For the automatic-assisted procedure, the outlining of areas was initiated by the analyst by positioning the mouse cursor on the edge of areas of interest and areas were automatically outlined by the computer program; however, during automatic tracing, adjustment or refinement by the analyst was possible.
Images were analyzed by three operators previously instructed and trained for CIA during 1 mo. Image analyses were carried out by each operator in 24 processing days, and only one operator was allowed to perform measurements in a specific day. Each operator performed three replicated measures per image and per method according to a full factorial design. Hence, the total number of analyses was 7,533 (279 images x three operators x three procedures x three replicates). After 10 processing days, operators were instructed and retrained for spatial calibration of the image analysis software program. Measures were partitioned in two sets of data: data set ALL, which included all available measures (n = 7,533), and data set 14D, which included only measures (n = 4,428) obtained after operators were instructed again for spatial calibration.
Statistical Analysis
Preliminary Analysis.
Sources of variation of measured areas and FESR were investigated separately for each method of CIA by analyzing data set ALL using the general linear model procedure of SAS (SAS Inst. Inc., Cary, NC), and the following linear mixed model:
 |
where y is a CIA measure, µ is the intercept, SB is the fixed effect of the slaughter group (10 groups), G is the fixed effect of gender (137 barrows and 142 gilts), PIG is the random effect of the pig (279 pigs) nested within SB x G interaction, OP is the fixed effect of the operator (three operators), REP is the fixed effect of the replicate (three replicates), SB x G, OP x REP, OP x PIG, and REP x PIG are interaction effects, and e is a random residual. The animal error term was used to test significance of SB, G, and SB x G effects. Relationships between measures provided by different CIA techniques were investigated by computing Pearson’s correlation coefficients using PROC CORR of SAS (SAS Inst. Inc., Cary, NC).
Estimation of Reproducibility and Repeatability of CIA Measures.
Estimation of variance components was accomplished, separately for measures provided by different CIA methods and for the two sets of data (data set ALL and 14D), using the VARCOMP procedure of SAS (SAS Inst. Inc.), and the following mixed linear model:
 |
where y, SB, G, PIG, OP, and e have the meaning previously described, but OP was considered a random effect. This is a standard model to estimate reproducibility and repeatability (International Organization for Standardization, 1994a,b). Animal (PIG), operator (OP), and PIG x OP effects, as well as residuals, were assumed to be independently and normally distributed with a mean of zero and variance
and 
, respectively. Restricted maximum likelihood was used as the method of estimation of variance components.
To evaluate the reproducibility of measures obtained using CIA techniques, two parameters were considered: reproducibility (RD), which is defined as the value below which the absolute difference between two single measures obtained with the same method of analysis on the same image under different conditions (different operators and different time) is expected to lie with a probability of 95% (International Organization for Standardization, 1994a), and coefficient of reproducibility (RD%), which is an indicator of the degree of agreement between two single measures made by different operators using a specific technique on the same image.
Functions of estimated variance components were
 |
where RD is the reproducibility (International Organization for Standardization, 1994a), and
 |
where RD% is the coefficient of reproducibility.
Two statistical parameters were estimated with the aim of evaluating the repeatability of measures provided by CIA techniques: repeatability (RT), which is defined as the value below which the absolute difference between two single measures obtained with the same method of analysis on the same image under the same conditions (same operator and a short interval of time) is expected to lie with a probability of 95% (International Organization for Standardization, 1994a); and coefficient of repeatability (RT%), which is an indicator of the degree of agreement between repeated measures made, using the same technique on the same image by the same operator, within a short period of time. These parameters were computed as
 |
where RT is the repeatability (International Organization for Standardization, 1994a), and as
 |
where RT% is the coefficient of repeatability.
Results and Discussion
Age and weight at slaughtering, as well as the average weight of fresh hams (14.0 kg), were within the ranges of characteristics of carcasses and cuts of heavy pigs available in Italy for dry-cured ham production (Lo Fiego et al., 2000). Heavy slaughter weights stem, primarily, from the need to generate raw hams weighing at least 11 kg as indicated by guidelines of San Daniele dry-cured hams production. Heavy weights of slaughter pigs have been traditionally considered indicators of meat maturity. Improvement of weight gain due to genetic selection and enhanced feeding and management practices have progressively weakened this association, and current guidelines for the production of dry-cured hams limit the age of slaughter pigs to a minimum of 9 mo.
Cross-section and fat eye areas were approximately 350 and 8.7 cm2, respectively, and differences between the mean size of areas due to different methods of image processing and analysis were limited (Table 1). Conversely, measures of the LA were, on average, 16% greater when obtained by the automatic-assisted or manual method than the automatic procedure. This discrepancy was due to a different area definition for the automatic method (small areas due to marbling and the empty area due to bone removal were not considered when the automatic procedure measured LA). Hence, comparisons between the automatic technique and other methods for image analysis were not feasible for the lean area, and separate statistical analyses for different methods of CIA were mandatory. Variation of the fat eye area (average CV = 41.9%) was markedly greater than that of the cross-section (average CV = 11.1%) and lean area (average CV = 11.5%). Variability of the fat eye area affects variation of FESR, which is probably the most important trait because it drives the consumer impression about a "lean" or "fat" ham. In this sample, FESR ranged from 0 to 6% and, on average, the fat eye area accounted for 2.5% of the cross-section area. It was not feasible to compare values obtained for areas of concern with other literature findings because no other study investigated the size of fat or lean areas in cross-sectioned dry-cured hams.
|
. Results for FESR (data not shown) were similar to those for measures of the fat eye area. Irrespective of the procedure of image analysis, effects included in the model accounted for a very large amount of variation of measured areas (R2 for the cross-sectional area, LA, and FA ranged from 95 to 99%).
|
). The SB was a composite source of variation, which included a number of effects due, for this study, to differences in the herd of origin, season, transport, and lairage conditions.
Only the measures of FA and FESR were markedly influenced (P < 0.01) by gender (Table 2). Hams from barrows exhibited, on average, a 10% increase in FA and FESR when compared to those from gilts. This result is consistent with those from studies that compared carcasses and hams of barrows and gilts in relation to fatness (Bittante et al., 1993; Gou et al., 1995).
The effect of pig was the most important source of variation in the size of areas (Table 2). A prominent role of genetic factors in affecting variation of measured areas and exploitation of individual variation in selection schemes aimed to improve quality of dry-cured hams might be argued. However, these hypotheses must be supported by specific genetic analyses, which will be the aim of future developments of the current research.
Effects due to replicates, and to the interaction between operator and replicate, did not (P > 0.05) influence measures of areas, irrespective of the method used for image analysis (Table 2). This observation is evidence of consistent operator technique in performing analysis of the same image more than once. Although more precise indications rely on the assessment of repeatability, it is likely that a unique measure per image ensures reliable assessment of size of areas under consideration, irrespective of the method of image analysis used.
Differences (P < 0.01) in measures for the same image were induced by different operators. A number of critical steps, which might contribute to interpret these differences, have been identified: spatial calibration of the analysis system (a mandatory step for all methods), placement of a reference box including the area under computation (mandatory for the automatic procedure), drawing of the contour line delineating the area (manual procedure), and manual intervention by the operator to refine or correct the contour line drawn by the software program to delimit the area (automatic-assisted procedure). The magnitude of differences across operators (Figure 1) was very small for the cross section and LA (maximum difference between least squares means for different operators in relation to average size of the area was less than 2% for both traits and all methods of image analysis). Maximum difference for the fat eye area was limited to 2% when the automatic-assisted procedure was used but showed a marked increase when the analysis of images was performed using other methods (6 and 9% for manual and automatic procedure, respectively). Hence, measuring by the automatic-assisted procedure minimizes the extent of differences across operators.
|
. All correlations were greater than 0.90 (P < 0.001), with the only exception of the correlations between measures of the lean area provided by the automatic procedure and those obtained with other methods. Differences in the definition of the lean area between the automatic procedure and the automatic-assisted or manual methods were responsible for the moderate correlations between measures of the lean area provided by different CIA techniques.
|
The Reproducibility of Image Analysis Measures
When variance components were estimated using all available measures (data set ALL), the reproducibility (RD) of measures of the cross-sectional area (Table 4) was similar across CIA procedures and ranged from 47 to 49 cm2 (14% of the average size of the area). Measures of the lean area were less reproducible (exhibited by higher estimates of RD) when using the automatic-assisted and manual procedures than when measured automatically. However, because of differences in the evaluated LA, which led to a lower average measure provided by the automatic method in comparison with the other techniques, ratios of RD estimates to average size of the lean area were similar across CIA techniques and ranged from 14.1 to 14.4%.
|
Reinstruction of the operators for spatial calibration of the software system greatly enhanced the RD of CIA measures of the cross-sectional and lean area, but its effects on the RD for the fat eye area and FESR were smaller and not consistent across methods of image analysis (Table 4). After reinstruction and a further training for spatial calibration, estimates of RD (data set 14D) for the cross section and LA were decreased 38 to 48%; however, the changes in RD for measures of the fat eye area and FESR (ranged from 4 to 22%) were not beneficial for measures of the fat eye area obtained with the automatic technique and for all measures of FESR. The enhancement of RD, which occurred after operators were further instructed for spatial calibration, was caused by a marked decrease in the magnitude of operator x pig and residual variance estimates. The amount of variation induced by operator effects was not reduced by reinstruction for spatial calibration. Hence, reinstruction led to a more homogeneous behavior of each operator, in relation to the calibration of the image analysis system, when the analysis was performed on different images.
Results obtained for RD% were consistent with those obtained for RD (Table 5). Ranking of CIA techniques based on RD% was identical to that based on RD estimates. The agreement between measures provided by different operators under the same technique was lower for areas of greater size (cross-sectional and lean areas) than for the fat eye area and FESR. The RD% ranged from 74 to 80% for the cross-sectional and lean areas, whereas the RD% for FA and FESR ranged from 88 to 98%. More-favorable estimates of RD% for the cross-sectional and lean areas were obtained from the analysis of measures collected after the reinstruction of operators for spatial calibration (data set 14D), and the increase in RD% was similar for the different techniques. The increase of R% estimates for measures of the cross-sectional area ranged from 16.8 (automatic method) to 19.4% (manual procedure), and the estimate of RD% for LA exhibited an increase that ranged from 17.4 (automatic method) to 21.2% (automatic-assisted method).
|
). Measures of the lean area provided by the automatic procedure exhibited a lower RT estimate than those obtained using the automatic-assisted or the manual procedures. When RT estimates were related to the average size of areas, the resulting ratio was very similar for different CIA techniques (from 11 to 12%).
|
Reinstruction of the operators for spatial calibration (data set 14D) reduced the magnitude of RT estimates obtained for the cross-sectional and lean areas but again demonstrated the limited effects on RT of other measures. The decrease in magnitude of RT estimates due to reinstruction of operators ranged from 37 to 45% for measures of the cross-sectional and lean areas (Table 6). Changes in RT estimates for the fat eye area and FESR were smaller but not consistent across procedures. Standardization of the behavior of operators when performing spatial calibration was a critical step in the CIA techniques applied to analysis of the cross section of dry-cured hams.
Results for RT% were consistent with those obtained for RT (Table 7). Ranking of procedures based on RT% was identical to that provided by RT and was not affected by reinstruction of operators for spatial calibration. Analysis of all measures provided higher estimates of RT% for the fat eye area and FESR (>90%) than for the cross-sectional and lean areas, irrespective of the technique used for measuring and did not exhibit meaningful changes after operators were further instructed for spatial calibration.
|
Implications
Results of this study indicate that computer image analysis is a highly repeatable and reproducible technique suitable for measuring the size of lean and fatty areas in cross-sectioned dry-cured hams. Of the methods of image analysis tested in this study, the automatic-assisted procedure was the most precise and least time consuming. The implementation of computer image analysis may be used for studies investigating the causes of defective size of fatty areas in dry-cured hams and related predictors useful in the screening of fresh hams before processing.
Footnotes
1 This research was funded by MURST (60%). The authors are very grateful to S. P. A. Montorsi and S. S. Gorzagri for their cooperation. The authors are indebted to G. Dallolio, D. Padoan, and G. Sandrolini for their support in organizing the experimental procedures at the ham factory, and the support of F. Fornasiero is also acknowledged for the use and adaptation of the image analysis software. 
2 Correspondence—phone: +39 049 8272664; fax: +39 049 8272633; e-mail: paolo.carnier{at}unipd.it.
Received for publication February 14, 2003. Accepted for publication October 16, 2003.
Literature Cited
Associazione Scientifica di Produzione Animale. 1991. Pages 48–56 in Metodologie Relative alla Macellazione degli Animali di Interesse Zootecnico e alla Valutazione e Dissezione della loro Carcassa. ISMEA, Roma, Italy.
Bittante, G., L. Gallo, and P. Montobbio. 1993. Estimated breed additive effects and direct heterosis for growth and carcass traits of heavy pigs. Livest. Prod. Sci. 34:101–114.
Cernadas, E., M. L. Durán, and T. Antequera. 2002. Recognizing marbling in dry-cured Iberian ham by multiscale analysis. Pattern Recogn. Lett. 23:1311–1321.
D.O.P. Prosciutto di San Daniele. 1996. Disciplinare della denominazione di origine protetta prosciutto di San Daniele. Regolamento CEE No. 1107.
Gandemer, G. 2002. Lipids in muscles and adipose tissues, changes during processing and sensory properties of meat products. Meat Sci. 62:309–321.
Gigli, S., M. T. Pacchioli, and D. Barchi. 1993. Valutazione della coscia per il prosciutto. Riv. Suinicol. 37:31–39.
Gou, P., L. Guerriero, and J. Arnau. 1995. Sex and crossbreed effects on the characteristics of dry cured ham. Meat Sci. 40:21–31.
International Organization for Standardization. 1994a. Accuracy (trueness and precision) of measurement methods and results—Part 1: General principles and definitions. ISO 5725-1. International Organization for Standardization, Geneva, Switzerland.
International Organization for Standardization. 1994b. Accuracy (trueness and precision) of measurement methods and results—Part 2: Basic method for the determination of repeatability and reproducibility of a standard measurement method. ISO 5725-2. International Organization for Standardization, Geneva, Switzerland.
Lo Fiego, D. P., R. Virgili, M. Belletti, F. Tassone, M. Pecoraio, M. Riverberi, and V. Russo. 2000. Caratteristiche delle carcasse e dei tagli di differenti tipologie di suino pesante attualmente presenti sul mercato. Riv. Suinicol. 41:143–148.
Shackelford, S. D., T. L. Wheeler, and M. Koohmaraie. 1998. Coupling of image analysis and tenderness classification to simultaneously evaluate carcass cutability, longissimus area, subprimal cut weights, and tenderness of beef. J. Anim. Sci. 76:2631–2640.
Swatland, H. J. 1995. Objective assessment of meat yield and quality. Trends Food Sci. Tech. 6:117–120.
Tan, F. J., M. T. Morgan, L. I. Ludas, J. C. Forrest, and D. E. Gerrard. 2000. Assessment of fresh pork color with machine vision. J. Anim. Sci. 78:3078–3085.
Virgili, R., and C. Schivazappa. 2002. Muscle traits for long matured dried meats. Meat Sci. 62:331–343.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
