Deep learning-based elaiosome detection in milk thistle seed for efficient high-throughput phenotyping

Area of Science:

• Plant science
• Biotechnology
• Computer vision

Background:

• Milk thistle (Silybum marianum) is valued for silymarin, concentrated in seed elaiosomes.
• Quantifying elaiosome phenotypes is challenging due to difficulties in image-based segmentation.
• Elaiosome traits are crucial for breeding milk thistle cultivars with enhanced medicinal properties.

Purpose of the Study:

• To develop a semi-automated method for detecting and segmenting milk thistle seed elaiosomes.
• To leverage deep learning for accurate elaiosome identification and quantification.
• To facilitate high-throughput phenotyping of elaiosome and other seed morphological traits.

Main Methods:

• Utilized the Detectron2 deep learning algorithm for elaiosome detection and segmentation.
• Employed a semi-automatic image labeling strategy, iteratively retraining the model with predicted results.
• Trained the final model on 6,000 labeled images for robust performance.

Main Results:

• Achieved a 99.9% accuracy in detecting milk thistle seed elaiosomes using Detectron2.
• Successfully generated precise segmentation masks for elaiosomes.
• Demonstrated the utility of elaiosome area analysis for seed phenotyping.

Conclusions:

• The proposed Detectron2-based method is highly effective for milk thistle elaiosome detection and segmentation.
• Semi-automated labeling significantly aids in generating sufficient training data.
• This approach supports high-throughput phenotyping, crucial for breeding superior milk thistle varieties.