Using latent class modelling to evaluate the performance of a computer vision system for pig carcass contamination

This study evaluates the performance of a computer vision system (CVS) for measuring pig carcass contamination using latent class modelling, a statistical approach that does not depend on a gold standard. Developed by the Danish Technological Institute, the CVS integrates output from various cameras to inspect pig carcasses for presence of faecal contamination. Data from a 16-day period involving 69,215 carcasses were analysed, comparing CVS results with those from official auxiliaries. Descriptive analyses identified four meat inspection findings that were statistically associated with an increased relative risk of positives from the CVS, particularly oil contamination (RR = 4.1, P < 0.001), which the CVS could not differentiate from faecal contamination. Agreement between the CVS and official auxiliary was assessed using Cohen's kappa and prevalence- and bias-adjusted kappa (PABAK), with Cohen's Kappa indicating minimal agreement (κ = 0.17) and PABAK indicating moderate agreement (κ = 0.79). Sensitivity and specificity were estimated using a latent class model fit within a Bayesian framework, without assuming that either the CVS or official auxiliaries were perfect tests. The latent class model showed that the CVS had a median sensitivity of 31.6 % (95 % CI: 27.6 %-39.1 %) and specificity of 97.9 % (95 % CI: 96.1-99.9 %), compared to 22 % (95 % CI: 17.6 %-28.9 %) sensitivity and 99 % (95 % CI: 98.2 %-100 %) specificity for the official auxiliaries. These findings underscore the CVS's strength in detecting true contaminations and official auxiliaries' ability to rule out non-contaminations. This study demonstrates the applicability of latent class modelling for evaluating CVS, offering a flexible and reliable framework that addresses the limitations of traditional gold standard methods. The findings support the use CVS technology alongside traditional inspections to enhance food safety, paving the way for future integration of CVS in meat inspection, pending legislative adjustments.