52 / 2025-04-13 11:43:05

LDSAM: A high-quality segmentation and scoring system based on deep learning for pig lung disease

Swine Pneumonia, Veterinary X-Ray, Precision Livestock Farming, Computer Vision Diagnostics, SAM Segmentation

3. 智慧牧场与动物福利 > （2）智慧养殖的大数据、机器学习和物联网应用

Detection of swine pneumonia is critical for mitigating substantial economic losses in intensive swine production systems. Slaughterhouse lung assessments serve as a vital surveillance point, as post-mortem examinations provide opportunities for herd-level disease monitoring that are logistically challenging in live livestock inspections. Previously, we developed a digital radiography-based scoring protocol for ex vivo pig lungs and achieved a robust agreement with a modified Madec lung scoring system. However, this method relies on expert visual scoring of X-Ray images, presents limitations in scalability, consistency, and quantitative analysis - factors particularly problematic when processing thousands of pigs daily in modern slaughter facilities. To address these challenges, we developed an AI-driven radiographic analysis and scoring system leveraging computer vision. Firstly, using a customized YOLO to detect pneumonia and then deploying fine-tuned and modified pig lung segmentation algorithm based on the Segment Anything Model (SAM), we segmented the pneumonia and lung regions. Then, pneumonia features such as texture, contour, and area proportion of the affected regions were input into a customized neural network for systematic scoring. Our model achieved a Mean Average Precision of 87.53% in lung pneumonia detection and a mean Intersection over Union  of 96.64% in lung segmentation, demonstrating its accuracy in lung area assessment. The system will analyze individual pig lung X-ray image to quickly and accurately assess pneumonia severity, providing precise pneumonia score for affected pigs. The methodology can establish a framework for merging computer vision into veterinary diagnostics in food animal production systems. Complete experimental results will be present at this conference.