Intramuscular fat percentage estimation through ultrasound images

Abstract

This work presents a new framework to estimate intramuscular fat percentage (%IMF) on live cattle based on ultrasound images. The %IMF measured in the longissimus dorsi muscle between the 12th and 13th rib is highly correlated with beef tenderness, one of the most determinant factors in meat quality pointed by consumers. Therefore, an automatic procedure for the estimation of this parameter is highly desirable. The proposed framework automatically determine a region of interest (ROI) in the acquired images dened by structures present in the image such as the subcutaneous fat and the ribs. A set of forty two features are extracted from each cropped ROI. These features are based on statistics and transformation of the ROI, for example, texture descriptors such as Local Binary Pattern, co-occurrence matrix, histograms, Fourier Transform coecients, among others. A feature extraction step is performed based in Principal Components Analysis, in order to reduce the number of dimensions and improve the computational performance. The new space of features triples the correlation with the real %IMF. As a result of this step, a feature vector of ten components is obtained, which accumulates 99% of the variance. The estimation of the %IMF is performed in this ten-dimensional space training a model based on Support Vector Regression (SVR), using a radial basis function as a kernel. For this kernel, the variance of kernel function and the tolerance parameters were optimized in the train stage. The framework is validated in a database of 283 ultrasound images obtained from 71 live steers. The acquisition was carried out by a trained professional in animal production. An estimation of the %IMF was obtained by an expert based in the ultrasound image aided with a commercial software. Also a standardized chemical analysis of the beef, with an error lower than 0.3,% was performed obtaining a ground-truth value for the %IMF. The database was divided into two sets randomly drawn, one to train the algorithm and compute the regression coecients and the other to test it. This procedure was repeated 100 times, varying the test and training set, and their average is presented here. The performance is measured using the Root-Mean-Square Error (RMSE), resulting in an improvement of 21% on the measurement compared with the estimation obtained by the expert with the software. The proposed framework shows promising results for a fully automatic procedure.