Cross-species AI: shifting a convolutional neural network from pigs to lambs to detect pneumonia at slaughter.

Abattoir-based data are widely regarded as suitable tools to estimate farm animals' health and welfare during the entire lifecycles. However, the systematic detection and recording of lesions at postmortem inspection are expensive, time consuming, somewhat biased by inter- and/or intra-observers' variability. Artificial intelligence could solve the above issues, and it could be particularly well-suited for solving repetitive tasks, by automating workflows and improving their efficiency. This study aims to assess whether a CNN, previously trained to score pneumonia in slaughtered pigs, is likewise capable of solving this task in a different animal species (i.e., in lambs). A total of 229 lamb lungs were photographed at postmortem inspection under different field conditions. Picture were evaluated by 5 independent veterinarians with different professional background, who scored each lung as healthy or diseased. The same pictures were scored by the CNN, which highlighted the lung profile, the bent over lobe (if any), and the lesion (if any). Finally, all veterinarians critically rated CNN's assessments. Overall, the CNN was able to solve that task, showing a substantial agreement (Cohen's kappa coefficient between 0.65-0.71) and high level of sensitivity (0.87-0.88), specificity (0.88-0.91), and accuracy (0.87-0.88) when compared to skilled investigators. Shifting CNN to different animal species could facilitate and fasten the adoption of such tools, which could benefit veterinarians and auxiliary staff, mainly where sheep farming is more widespread and economically relevant.