Can AI outperform human evaluators in swine disease detection?

Respiratory diseases in pigs are a major concern in the livestock industry, with Mycoplasma hyopneumoniae being a common pathogen responsible for CVPC. Traditionally, lesion scoring has relied on human evaluators, but this approach is subject to variability and inconsistencies. The research team evaluated a prototype AI system called AI DIAGNOS, developed by HIPRA Laboratories, to determine its accuracy and reliability in scoring lung lesions from slaughterhouse images.

The use of artificial intelligence (AI) in veterinary diagnostics is gaining momentum, offering new possibilities for precision and efficiency. A recent study compares AI-driven computer vision systems (CVS) with human evaluators in scoring swine lung images for cranioventral pulmonary consolidation (CVPC), a key indicator of respiratory infections in pigs.

The study, titled Scoring of Swine Lung Images: A Comparison Between a Computer Vision System and Human Evaluators, was conducted by Valeris-Chacin et al. and published in Veterinary Research (2025). It aims to assess whether AI can provide a viable alternative to manual scoring in slaughter inspections, addressing concerns related to human variability and time constraints.

Role of AI in swine disease detection

Respiratory diseases in pigs are a major concern in the livestock industry, with Mycoplasma hyopneumoniae being a common pathogen responsible for CVPC. Traditionally, lesion scoring has relied on human evaluators, but this approach is subject to variability and inconsistencies. The research team evaluated a prototype AI system called AI DIAGNOS, developed by HIPRA Laboratories, to determine its accuracy and reliability in scoring lung lesions from slaughterhouse images.

The study involved analyzing 1,050 swine lung images captured at slaughter. These images were assessed by five trained human evaluators, and their scores were compared to those assigned by the AI system. The AI DIAGNOS system operates using Amazon SageMaker™, incorporating detection and classification processes powered by convolutional neural networks. By automating lesion assessment, AI aims to standardize scoring, minimize variability, and enhance the efficiency of swine health monitoring.

AI vs. human evaluators: Accuracy and variability

The study found that AI DIAGNOS achieved moderate accuracy (62-71%) in distinguishing between lesioned and non-lesioned lung lobes in most cases. However, its accuracy in classifying lesion severity at a detailed lobe level was lower (24-36%). This discrepancy suggests that while AI excels at identifying affected lungs, it struggles with assigning precise lesion severity scores, a task where human evaluators still maintain an advantage.

Inter-evaluator variability was notable, with intraclass correlation coefficients (ICCs) ranging from 0.29 to 0.6, highlighting inconsistencies in human scoring. Intra-evaluator variability, however, remained low, demonstrating that individual evaluators were relatively consistent in their own assessments. AI, on the other hand, showed perfect consistency, with a repeatability score of 1. This suggests that AI could be instrumental in reducing subjectivity in lung lesion scoring over time.

Future of AI in veterinary diagnostics

While AI DIAGNOS demonstrated promising results, there are still areas for improvement. The study noted that the system tended to assign lower lesion scores compared to human evaluators, potentially due to limitations in differentiating between artifacts such as blood stains and true pathological lesions. Additionally, the inability to palpate lungs, a key aspect of traditional CVPC scoring, poses a challenge for image-based AI analysis.

Despite these limitations, AI-based lung lesion scoring offers several advantages, including time efficiency, scalability, and elimination of human subjectivity. As AI technology advances and training datasets expand, the accuracy of such systems is expected to improve. Future research should focus on refining AI algorithms to enhance their precision in lesion severity classification and integrating complementary diagnostic tools, such as real-time analysis during slaughter inspections.

Conclusion: AI as a supportive tool, not a replacement

AI DIAGNOS can serve as a valuable tool for swine lung lesion scoring, particularly in reducing inter-evaluator variability and streamlining the assessment process. However, AI should not yet replace human evaluators entirely. Instead, a hybrid approach combining AI-driven preliminary assessments with expert validation may offer the best of both worlds.

As the livestock industry moves toward more data-driven decision-making, AI-based diagnostic tools like AI DIAGNOS will likely play an increasingly important role in disease monitoring and animal health management. This study marks a significant step in that direction, paving the way for future enhancements in AI-assisted veterinary diagnostics.