CT image segmentation of meat sheep Loin based on deep learning

Area of Science:

• Veterinary Imaging
• Medical Image Analysis
• Deep Learning Applications

Background:

• Accurate segmentation of sheep internal tissues in Computed Tomography (CT) images is challenging due to indistinct tissue boundaries.
• Traditional image segmentation methods struggle to meet the demands of practical applications in veterinary medicine.

Purpose of the Study:

• To investigate the effectiveness of deep learning models for segmenting sheep loin CT images.
• To compare the performance of various deep learning architectures for this specific task.

Main Methods:

• A dataset of 1471 sheep loin CT images from Australian White and Dolper rams was utilized.
• Six deep learning models, including Fully Convolutional Neural Network (FCN) and five UNet variants (Attention-UNet, Channel-UNet, ResNet34-UNet), were applied.
• 5-fold cross-validation and 10 independent runs were employed for robust evaluation using metrics like accuracy, AVER_HD, MIOU, DICE, and LOSS.

Main Results:

• All evaluated deep learning models demonstrated excellent performance in sheep loin CT image segmentation.
• Attention-UNet achieved the highest accuracy (0.998±0.009), best MIOU (0.90±0.012), and DICE (0.95±0.007) scores.
• Channel-UNet yielded the optimal LOSS value (0.029±0.018), while ResNet34-UNet offered the shortest running time.

Conclusions:

• Deep learning models provide a robust solution for sheep loin CT image segmentation, surpassing traditional methods.
• Attention-UNet is identified as the top-performing model for accuracy and segmentation quality in this study.
• The choice of model may depend on specific application requirements, balancing performance metrics and computational efficiency.