DeepPod

DeepPod quantifies fruit (silique) number in Arabidopsis thaliana from non-destructive images using a CNN-based patch‑based two‑phase deep learning framework.

Key Features:

• Patching 2-phase framework: Implements a patch-based two-phase deep learning approach for image-wide feature extraction with convolutional neural networks (CNNs).
• CNN part classification: Classifies distinct inflorescence parts — tip, base, body of siliques (seed pods), and stem inflorescence — using CNNs.
• No segmentation or bounding boxes required: Performs classification without detailed segmentation or bounding-box annotations, reducing the need for pixel-level labels.
• Post-processing assembly: Joins classified parts belonging to the same silique and accounts for overlapping siliques to enable accurate detection and localization.
• High-throughput, non-destructive input: Operates on non-destructive imaging data to support large-scale phenotyping and automated quantification of fruit number.
• Validation and accuracy: Validated on an independent test dataset of 2,408 images with a reported correlation of R² = 0.90 to manual counts.

Scientific Applications:

• High-throughput plant phenotyping: Automated quantification of silique number in Arabidopsis thaliana for large-scale phenotyping studies.
• Trait measurement for genetic and environmental studies: Provides rapid fruit-number estimates to support analyses of genetic and environmental factors influencing plant development.
• Cross-species imaging studies: Applicable to other plant species where non-destructive imaging is used for phenotypic analysis.

Methodology:

Uses a patch-based two-phase CNN framework to classify inflorescence parts (tip, base, body of siliques, stem inflorescence) without segmentation or bounding boxes, followed by post-processing to join parts of the same silique and resolve overlaps; validated on an independent test set of 2,408 images with R² = 0.90 versus manual counts.