Evaluation of an automatic lean meat percentage quantification method based on a partial volume model from computed tomography scans

Abstract

The quality of a pig carcass is mainly measured by the lean meat percentage (LMP), which can be virtually estimated from computed tomography (CT) scans. Different strategies exist to classify the CT voxels into tissues such as fat, lean and bone, being the thresholding-based methods the most commonly used. However, these methods are usually affected by the partial volume effect, and also by data variability, which is implicit from different CT scanners and protocols, since no standard behaviour has been defined. The aim of this paper is to extend an LMP quantification method which uses a partial volume model by adding a new step to detect the animal skin, and thoroughly evaluate the new approach by analysing each of its steps. The evaluation is performed by comparing the whole pipeline of the proposed approach with a simple thresholding method and a thresholding method with bone filling and skin detection, which is an intermediate step of the new pipeline. Five experiments have been designed to test how accurate are the results of the method regarding the LMP values computed from the manual dissection, as well as the robustness to data variability. Two different manual dissection methodologies have been tested: the partial dissection, which estimates the LMP using the lean of the four main cuts of the carcass plus the tenderloin, and the total dissection, which uses the lean of the twelve main cuts. A total of 146 half carcasses have been used for this study (105 using the partial dissection methodology, and 41 using the total dissection one). To evaluate the experiments, the LMP values virtually obtained from the three methods have been compared mostly with the LMP values from the manual dissection, computing the coefficient of determination R2 from the correlations, as well as the root mean square error of prediction by means of leave-one-out cross-validation. A statistical analysis is performed to resolve if two correlations are significantly different. The experiments’ results confirm the high accuracy of the proposed approach for the LMP estimation, and mainly its high robustness to data variability. The experiments also disclose that the detection of the animal skin and its classification as a new tissue, instead of classifying it as lean, improve the results. The evaluated method has demonstrated to be as effective as the thresholding method with bone filling and skin detection, and more robust to data variability than the other evaluated methods.