A User-Friendly Machine Learning Pipeline for Automated Leaf Segmentation in Atriplex lentiformis

Area of Science:

• Plant Science
• Computational Biology
• Agricultural Technology

Background:

• Automated leaf segmentation is crucial for plant phenotyping but requires balancing accuracy, scalability, and usability.
• Existing deep learning models may not be optimized for specialized plant research datasets.

Purpose of the Study:

• To develop and evaluate an end-to-end deep learning pipeline for practical plant phenotyping.
• To compare a fine-tuned Mask R-CNN model against foundation models like Segment Anything Model (SAM).
• To integrate automated sample identification and create a user-friendly application for researchers.

Main Methods:

• Developed a pipeline using a fine-tuned Mask R-CNN segmentation model trained on 176 plant images.
• Compared Mask R-CNN performance against Meta AI's SAM, utilizing Grounded SAM and Leaf-Only SAM post-processing.
• Integrated QR codes for sample identification and benchmarked decoding libraries.
• Deployed the pipeline as a Streamlit web application.

Main Results:

• The fine-tuned Mask R-CNN achieved a high Dice coefficient (0.781) despite a small training dataset.
• Transfer learning on a specialized dataset demonstrated superior performance over a large foundation model for domain-specific tasks.
• QR code integration and decoding proved robust under various real-world imaging conditions.

Conclusions:

• Specialized, fine-tuned deep learning models can outperform general foundation models in specific scientific domains like plant phenotyping.
• The developed pipeline offers an open-source, scalable, and user-friendly framework for plant research.
• Addressing practical deployment challenges enhances the usability and adoption of automated tools by researchers.