An efficient data-driven tissue deformation model

In this paper we present an efficient data-driven tissue deformation model. The work originates in process automation within the pig meat processing industry. In the development of tools for automating accurate cuts, knowledge on tissue deformation is of great value. The model is built from empirical data; 10 pig carcasses are subjected to deformation from a controlled source imitating the cutting tool. The tissue deformation is quantified by means of steel markers inserted into the carcass as a three-dimensional lattice. For each subject marker displacements are monitored through two consecutive computed tomography images - before and after deformation; tracing corresponding markers provides accurate information on the tissue deformation. To enable modelling of the observed deformations, the displacements are parameterised applying methods from point-based registration. The parameterisation is based on compactly supported radial basis functions, expressing the displacements by parameter sets comparable between subjects. For modelling the tissue deformation, principal component analysis is applied, treating each of the parameter sets as an observation. Using leave-one-out cross-validation, marker displacements are estimated in all subjects from the mean parameters. This yields an absolute error with mean 1.41 mm. The observed lateral movement of the loin muscle is analysed in relation to the principal modes, and the results are compared to manual measurements of carcass composition. We find an association between the first principal mode and the lateral movement. Furthermore, there is a link between this and the ratio of meat-fat quantity - a potentially very useful finding since existing tools for carcass grading and sorting measure equivalent quantities.