Integrated use of ultrasound imaging and multivariate image analysis for detecting bone fragments in poultry meat

Poultry meat industry requires intelligent systems for achieving non-invasive real-time detection of bone fragments. Therefore, the main aim of this study was to assess the feasibility of using ultrasound imaging and multivariate image analysis to detect bone fragments in boneless and skinless chicken breast fillets. Bone fragments of different sizes were inserted into the chicken and contact ultrasound images were acquired, following a pre-established pattern, in the control (C) and out-control (OC, with bone) samples, by scanning the breast's surface, using contact ultrasound sensors (1 MHz) working in through transmission. Energy-magnitude and energy-distribution ultrasound parameters were computed at pixel level in time (TDA) and frequency domain (FDA). Principal Component Analysis (PCA) was used in TDA and FDA parameters, and its combination (TFDA). From PCA model, the Residual Sum Squares (RSS) and Hotelling's T-square (T²) control statistics were used to classify the C and OC images projected on the PCA latent structure. Experimental results demonstrated that the presence of bone fragments within chicken breast fillets led to alterations in the energy-magnitude (avg. amplitude decrease from 81.6 % to 52.6 %, depending on the bone size) and energy-distribution ultrasound parameters (avg. variance decreased from 97.9 % to 70.6 % depending on the bone size). The RSS statistic achieved the best classification performance (accuracy of TDA, FDA and TFDA>95 %) in C and OC images. These results highlight the potential of combining contact ultrasound imaging with multivariate image analysis for the reliable and rapid detection of bone fragments in chicken breasts.