Prediction of Carcass Lean Yield Traits From Live Animal Ultrasound Measurements

J.P. Gibson, K. Nadarajah, C.A. Aker and R.O. Ball

Background

Ultrasound probes have been used for many years in Canada and elsewhere to measure backfat depth on live pigs for use in genetic evaluation programs. One objective of the OPCAP was to compare existing measurements of backfat depths with a wider range of measurements of fat depth, muscle depth and muscle areas provided by more advanced "real time" ultrasound machines. Real time ultrasound (also known as B mode ultrasound) machines provide accurate images of fat and muscle that should convey more information about the amount of lean and fat in the carcass. Apart from the equipment used, there is also the question of which are the best sites to take measurements, and whether measurements at different sites provide information about different aspects of carcass composition. Another issue is ease of use in the field. Current probes are simple to use and require no interpretation of images. Real time machines are slightly more complex to use and provide stored visual images that require detailed interpretation after scanning. Correct interpretation is particularly important with cross-sectional images where the correct location and angle to measure fat and muscle depth must be determined along with the area of the muscle. Longitudinal real-time measurements provide an image of fat and muscle depths between 10th and 15th ribs. Such images provide no information about the cross-sectional shape or area of the muscle and fat, but they can be interpreted relatively rapidly with the assistance of computer software such as AUSKey Auto-D developed by Dr. J. Stouffer.

Methods

A summary of scanning sites and methods is given by Aker (these proceedings). Cross sectional scans at 3-4th last rib, last rib and loin were interpreted manually by three different interpreters over the course of the experiment. Interpreters A, B and C interpreted images from 577, 495 and 1383 pigs. Only interpreter C interpreted the images from pigs which had shoulder and ham measurements. Fat and muscle depths from the longitudinal scans were obtained using the AUSKey Auto-D software (J. Stouffer, Animal Ultrasound Services, Ithaca, NY) with interpreter intervention to check positions of fat and muscle boundaries. All longitudinal images were interpreted by a novice interpreter (interpreter C) in Guelph and by an experienced interpreter at Animal Ultrasound Services, Ithaca.

We explored the possibility of predicting the following carcass traits using various combinations of ultrasound measurements:

Lean content of three primal cuts (%)
Lean content of shoulder (%)
Lean content of loin (%)
Lean content of ham (%)
Shoulder lean (% of lean in three primal cuts)
Loin lean (% of lean in three primal cuts)
Ham lean (% of lean in three primal cuts)

The first four traits are measures of overall leanness, whereas the last three are measures of where the lean occurs. Details of statistical methods are given in Appendix 3. A large number of alternative analyses were performed and only brief summaries of key results are presented here. Lean in three primals, lean in shoulder, lean in loin and lean in ham are all very closely related to each other and similar results were obtained for each. We only give results for lean content in three primals here.

Predicting Lean Content

Some key results are summarized in Table 1. The R2 in Table 1 is the proportion of variation in lean content between animals that is explained by the prediction equation. An R2 of 1 would be a perfect prediction while an R2 of 0 would indicate no predictive power. For each situation the accuracy of prediction equations that include breed and sex is shown. This would generally be the equation recommended for use in practice. Also shown is accuracy of the prediction that only uses ultrasound information. This indicates the minimum accuracy when breed and sex effects are not known, which should apply if predictions were made for breeds other than those studied in the OPCAP.

The predictions from A-mode measurements of fat depth gave an R2 of 0.62 (see Table 1). This is based on the average fat depth at the two scanning sites; but use of only one site caused little loss of accuracy, because measurements at the two sites were very highly correlated. There was also little difference in ability to predict leanness between measurements at 90 kg versus 100 kg liveweight (results not shown).

Using real-time cross-sectional measurements at a single site (last rib), the best interpreter achieved an R2 of 0.74. The other two interpreters had R2 0.1 to 0.15 lower than this (results not shown). The best interpreter was the most experienced, and produced more consistent and accurate interpretations than the two less experienced interpreters. However, the difference in accuracy between interpreters was reduced when information from several sites was included in the prediction of carcass lean (eg. compare best interpreter with interpreter C when multiple sites were used, Table 1, R2 = 0.78 versus 0.72 and 0.74).

There was essentially no difference in accuracy of predictions between longitudinal measurements interpreted in Guelph or Ithaca, both giving an R2 of about 0.64. This accuracy was about 0.1 lower than from cross-sectional measurements at a single site with the best interpreter, and about 0.14 lower than when two or more sites were used, regardless of interpreter.

Other than a difference in average leanness, there was no evidence of differences between breeds in the relationship between leanness and ultrasound measurements. This means that a single set of regression equations can be used for all breeds, which simplifies use on farm.

Predicting Distribution of Lean

Results for predicting lean distribution are shown in Table 2. The R2 were generally quite low, even when breed, sex and measurements from several sites were included and predictions were based on the best interpreter. When only ultrasound measurements were included, R2 dropped substantially. For shoulder and loin there was still some potentially useful predictive ability (R2 = 0.19 and 0.21), while for ham there was very little predictive ability (R2 = 0.09).

Discussion and Implications

The results show that cross-sectional images produced by real-time ultrasound can be used to predict carcass leanness with much higher accuracy than fat depths produced by a conventional A-mode machine. Prediction was also much more accurate than when using longitudinal real-time ultrasound images. Using information from two cross-sectional sites (10th rib and last rib) gives considerably higher accuracy than the best single site (last rib), even when an experienced interpreter is involved. Using two (or possibly more) sites becomes even more important when a less experienced interpreter is involved.

Longitudinal images can be interpreted with the aid of computer software (AUSKey Auto-D) which requires less time and skill than interpretation of cross-sectional images, especially if muscle area is used. Nevertheless, given the much higher accuracy possible from cross-sectional images, the recommendation is that real-time cross-sectional images at two sites should be used whenever possible.

Accuracies of prediction of distribution of carcass lean in the primals are probably too low at this stage to recommend application on farm. Accuracies are sufficiently encouraging, however, that it may be possible to find better predictors in the future. It should be possible to explore various other carcass measures in the OPCAP data to determine the limits to prediction of lean distribution. It is likely, however, that further trials will be required, perhaps using more advanced or more detailed ultrasound measurements, before prediction of lean distribution becomes available on farm.

Table 1. Accuracy of prediction of percentage lean in three primals from various ultrasound measurements.
Types of Measurements and
Model Fitted

R2
Types of Measurements and
Model Fitted

R2
A-Mode Fat Depth Cross-Sectional Real-Time
Breed, sex, ultrasound1
0.62
- Interpreter C2
Ultrasound only1
0.60
a. Measurements at all 3 back sites
Breed, sex, ultrasound
0.72
Carcass Fat Depth Ultrasound only
0.69
Breed, sex, fat depth
0.63
b. Measurements at all 5 sites
Fat depth only
0.60
Breed, sex, ultrasound
0.74
Ultrasound only
0.73
Cross-Sectional Real-Time
- Best Interpreter Longitudinal Real-Time
a. Measurements at last rib - Interpreter C2
Breed, sex, ultrasound
0.74
Breed, sex, ultrasound
0.63
Ultrasound only
0.71
Ultrasound only
0.59
b. Best two locations (last rib
and 10th rib)3 Longitudinal Real-Time
Breed, sex, ultrasound
0.78
- AUS Interpretations4
Ultrasound only
0.75
Breed, sex, ultrasound
0.64
Ultrasound only
0.59

1 Ultrasound includes all ultrasound measurements at that location plus all quadratic and interaction terms among measurements within each site.
2 Interpreter C interpreted all the longitudinal images at Guelph and was not as good as the best interpreter with individual cross-sectional images. When images at several locations were included, however, predictions were as good as with the best interpreter.
3 Addition of measurements at loin site did not improve accuracy.
4 Interpretations undertaken by Animal Ultrasound Services, Ithaca using the AUSKey Auto-D software.

Table 2. Accuracy of prediction of distribution of lean in carcass.
R2 for Prediction of
Measurements and Model
Shoulder Lean1
Loin Lean1
Ham Lean1
All Five Locations2
Breed, sex, ultrasound
0.24
0.27
0.19
Ultrasound only
0.20
0.17
0.08
Last Rib - Best Interpreter
Breed, sex, ultrasound
0.22
0.28
0.19
Ultrasound only
0.15
0.17
0.04
All Locations - Best Interpreter3
Breed, sex, ultrasound
0.26
0.32
0.24
Ultrasound only
0.19
0.21
0.09

  1. Lean in primal as percentage of total lean in three primals.
  2. Five locations were shoulder, 10th rib, last rib, loin and ham.
  3. Three locations were 10th rib, last rib and loin.