Deep-learning-based gestational sac detection in ultrasound images using modified YOLOv7-E6E model

Area of Science:

• Agricultural technology
• Artificial intelligence in animal husbandry
• Computer vision for livestock management

Background:

• Rising global population and income drive increased meat demand.
• Decreasing farm numbers and farmers threaten meat sufficiency.
• Information and Communications Technology (ICT) offers solutions for livestock farm efficiency.

Purpose of the Study:

• To develop an AI system for accurate sow gestation sac counting from ultrasound images.
• To improve sow productivity through precise early pregnancy detection.
• To enhance the performance of the YOLOv7-E6E model for this specific application.

Main Methods:

• Utilized the YOLOv7-E6E object detection model.
• Modified the activation function from SiLU to a SiLU + Mish multi-activation function.
• Implemented bicubic upsampling instead of nearest neighbor upsampling.
• Applied AutoAugment for data augmentation.

Main Results:

• The original YOLOv7-E6E model achieved 86.3% mean average precision (mAP).
• Individual improvements were noted with multi-activation function (+0.3%), bicubic upsampling (+0.9%), and AutoAugment (+0.9%).
• Simultaneous application of all proposed methods resulted in a significant performance increase to 89.8% mAP (+3.5%).

Conclusions:

• The optimized YOLOv7-E6E model significantly improves the accuracy of sow gestation sac detection.
• These AI-driven enhancements in pregnancy diagnosis can lead to increased livestock farm productivity.
• The study demonstrates the potential of advanced ICT in addressing challenges in modern animal agriculture.