Bayesian mixture model: Bloodless dissection of a sheep

Project Investigators:

Professor Kerrie Mengersen
Dr Clair Alston

(pdf 435kB)

Brief

Australia is a world leading producer and exporter of sheepmeat, with the estimated off-farm value of approximately $4.3 billion. While this is a valuable export market, the domestic expenditure of sheepmeat is also important. According to Meat and Livestock Australia, more than half of Australian fresh meat buyers pur-chase lamb and with estimated value of $2.4 billion. The international and domestic requirements on sheepmeat quality are very different. For instance, Australian consumers often favour lean meat, while consumers in the Asian market pre-fer sheepmeat with a high fat composition. Therefore, factors that impact on the tissue composition are important for sheepmeat industry.

In the past, to obtain the tissue compositions for experiments such as nutrient trials or management plans required the dissec-tion of slaughtered animals. However, this is often time consum-ing, expensive and suffers from significant measurement error. CT scanning in animal experimentation has increased over the last decade due to the benefits of studying the tissue composi-tion in livestock over time. CT scanning does not measure tissue type directly. After image processing, the scan provides an estimate of the denseness of the tissues at each of the given pixels. A pixel can be mixed, that is, consisting of two or even three tissue types in the area de-fined as a pixel. In addition, the density of tissue types varies de-pending on location on carcase. As a result, the same CT num-ber can represent either fat or muscle, at the lower boundary, or muscle or bone at the denser boundary.

A common approach in the literature for allocating the pixel to a tissue type is the use of fixed boundaries. However, at the accuracy of this varies between sheep and carcase location and the boundaries of pixel values are not determined in a rigorous scientific manner. At this stage, there is no gold standard way to measure the proportion of lean, fat and bone in a live sheep.

Statistical methodologies, such as a mixture model, can help to estimate the area of each of three tissue types of interest present in individual CT scan slices. Our team has been working on this project since 2005. During this time we developed different mixture models and implemented these on the CT scanning data. The results of our approaches are more accurate than the conventional approach and provide meas-ures of uncertainty. We are continually developing our techniques for analysing this data.
Future directions of the project include validation of the model using chemical measurements, work in which we are collaborating with colleagues at CREST (Paris). We are also focussing on computational im-provements in several areas, firstly in the area of using Graphic Processing Units in which we have achieved a speed-up factor of 30 for the computations and secondly by using approximation methods such as Variational Bayes.