Accurate classification of lung lesions at necropsy is crucial for guiding the diagnostic process and ensuring effective management of porcine respiratory diseases. Post-mortem inspection of the lungs during slaughter also provides valuable insights into disease occurrence, offering useful feedback on the efficacy of on-farm prevention and control strategies. However, manual assessment protocols may be impaired by high slaughtering speeds and low inter-rater agreement, which limits continuous data collection and hinders comparability. Artificial intelligence, particularly computer vision (CV), may offer a promising alternative. This study aimed to train and test a commercial CV model for segmenting both anatomical structures and lesions in pig lungs. Overall, 1742 lungs were collected at slaughter, examined macroscopically, and photographed laterally. Two veterinarians with expertise in swine pathology manually annotated the acquired images to outline anatomical (i.e., lung, heart, lung lobes), pathological (i.e., bronchopneumonia, fibrinous pleuropneumonia, chronic pleuritis), and artefactual (i.e., parenchymal laceration, bronchoinhalation of blood) classes, forming the reference dataset for model training and testing. Model performance in segmenting these classes varied by confidence threshold, with sensitivity (36-84 %), positive predictive value (62-93 %) and F1 score (52-78 %) indicating imperfect yet improvable performance. Overall, anatomical structure segmentation outperformed lesion detection, likely due to class imbalance in the training dataset and the complexity of pulmonary pathology. Integrating standardized and real-time detection of lung lesions via digital imaging could improve respiratory health surveillance, thereby enhancing the role of abattoirs as strategic epidemiological observatories.
Recchia et al. (Mon,) studied this question.