Automated scoring of porcine pneumonia: from theory to practice

Background: Scoring lesions at slaughter represents a suitable and efficient tool for estimating the economic impact of porcine enzootic pneumonia (1). However, the systematic assessment of lesions is too demanding in terms of human resources. Considering this, the use of computer vision systems (CVS) offers interesting prospects. Objectives: To report recent experience gained in a high-capacity (800 pigs/hour) abattoir, aiming to fully automate the scoring of enzootic pneumonia-like lesions in slaughtered pigs. Methods: This study was carried out at the Cooperl abattoir in Montfort-sur-Meu (France). Images (8 per pig) were acquired using fixed cameras and analyzed by an ad hoc trained artificial neural network. Scores assigned by the artificial neural network were expressed as percentage of the lung surface and then compared with those given by a skilled operator using the Madec grid (2). Results: The final test was carried out on 1,618 pigs. CVS specificity was 75.55%. Sensitivity exceeded 96% for large lesions (>5 on the Madec grid), it was high for medium sized lesions (82% for values 3-4 on the Madec grid), and much lower for small lesions (54.60% for values 1-2 on the Madec grid). Conclusions: The automatic scoring of lesions is a challenging task, requiring integrated technical solutions and complementary skills, which must be adapted to each slaughterhouse and business setting. Overall, the CVS was rather capable of fulfilling its task, despite challenging logistical conditions, due to the features and speed of the slaughter chain. Looking forward, the CVS could be further improved by expanding the training dataset, balancing the categories to be identified, and enhancing the quality of pictures.