LeafNet

LeafNet automates stomata localization and pavement cell segmentation in bright-field and confocal leaf epidermal images to enable quantitative phenotyping of stomatal and pavement cell morphology.

Key Features:

• Hierarchical deep-learning strategy: Employs a multi-stage hierarchical approach that leverages deep learning for identification and quantification of epidermal features.
• Deep convolutional network for stomata localization: Uses a deep convolutional network to robustly detect stomata in bright-field microscopy images.
• Stomata masking and pavement cell segmentation: Masks detected stomata and segments pavement cells using an efficient region merging method.
• Transfer learning: Applies transfer learning to adapt models to a wide range of species and to confocal images.
• Quantitative phenotype extraction: Accurately quantifies various phenotypes of individual stomata and pavement cells across test images.
• Benchmarking: Demonstrates superior performance compared with other tools including StomataCounter, Cellpose, PlantSeg, and PaCeQuant.
• Batch processing: Supports large-scale image processing in batch mode for high-throughput analysis.
• Extensible architecture: Architecture supports extension of additional functionalities.

Scientific Applications:

• Quantitative phenotyping: Measurement and analysis of stomatal and pavement cell morphological phenotypes from microscopy images.
• Developmental studies: Investigation of morphological variation and developmental regulation of stomata and pavement cells.
• Comparative and cross-species analysis: Application to a wide range of species enabled by transfer learning.
• High-throughput leaf epidermal phenotyping: Large-scale image-based phenotyping using batch processing.

Methodology:

LeafNet implements a hierarchical strategy using a deep convolutional network to localize stomata in bright-field microscopy images; detected stomata are masked and pavement cells are segmented via an efficient region merging method, and transfer learning is used to extend analysis to confocal images and diverse species.