Real-Time Callus Instance Segmentation in Plant Tissue Culture Using Successive Generations of YOLO Architectures

Area of Science:

• Plant Biotechnology
• Computer Vision
• Agricultural Science

Background:

• Callus induction is crucial for plant propagation, metabolite production, and genetic modification.
• Manual monitoring of callus formation is labor-intensive and subjective.
• Automated methods are needed to accurately assess callus development in plant tissue culture.

Purpose of the Study:

• To create the first curated lentil (Lens culinaris) callus dataset for instance segmentation.
• To evaluate the performance of successive YOLO deep learning models for callus segmentation.
• To identify optimal deep learning architectures for precise and efficient callus monitoring.

Main Methods:

• Generated a lentil callus dataset using three genotypes (Firat-87, Cagil, Tigris) and three developmental stages (leaf, green callus, necrosis callus).
• Acquired 122 high-resolution images with 1185 annotations.
• Evaluated YOLOv5, YOLOv7, YOLOv8, and YOLOv11 models using instance segmentation metrics (mAP, Dice, Precision, Recall, IoU) and efficiency metrics (parameters, FLOPs, inference speed).

Main Results:

• Anchor-free YOLOv8 and YOLOv11 models outperformed anchor-based YOLOv5 and YOLOv7 in boundary precision for callus structures.
• YOLOv8 achieved the highest instance segmentation precision (mAP50@0.855).
• YOLOv8 demonstrated real-time inference at 166 FPS with high accuracy and efficiency.

Conclusions:

• The developed lentil callus dataset supports advancements in automated plant tissue culture monitoring.
• Anchor-free deep learning models, particularly YOLOv8, offer significant advantages for accurate and efficient callus instance segmentation.
• This research paves the way for improved precision agriculture and plant breeding through AI-driven analysis.