Comparing computer vision models for detecting chronic pleurisy in pigs

This study evaluated the diagnostic performance of three convolutional neural network models within a computer vision system (CVS) for detecting chronic pleurisy in pig carcasses compared to official meat inspection. Knowledge about the true prevalence of chronic pleurisy is important for the pig producer, because this condition is negatively associated with productivity. Registering chronic pleurisy is no longer considered a priority by the Danish competent authorities, as such, the abattoir mainly uses the information for quality assurance. The performance was evaluated using latent class modelling, an approach which is independent of a gold standard to estimate sensitivity and specificity. Data from 85,413 pig carcasses across 15 slaughter days were analysed using traditional agreement statistics and Bayesian latent class modelling. Agreement between meat inspectors and the different CVS was estimated using Cohen's kappa and prevalence- and bias-adjusted kappa, indicating moderate (κ=0.72-0.77) and near perfect agreement (κ=0.86-0.89), respectively. The CVS models demonstrated superior sensitivity (84.7-90.3 %) compared to the meat inspectors (79.4-83.0 %), while the meat inspectors maintained slightly higher specificity (99.7-99.9 % versus 97.6-98.8 %). The ResNeXt-101 architecture with 1024-pixel resolution (CVS-Complex-HighRes) performed best overall, while aggregating outputs from multiple CVS models improved specificity without compromising sensitivity (CVS-Consensus). Based on the Danish meat inspection codes, no significant associations were found between the CVS's chronic pleurisy detection and other meat inspection findings except minor slaughter defects. The study demonstrates that uniform registrations can be made precisely by this CVS independently in a context, where the competent authority considers registration redundant. These findings support the implementation of CVS technology within risk-based meat inspection frameworks, though establishing performance thresholds and addressing regulatory considerations remain necessary for widespread adoption in commercial settings.