DeepCob

DeepCob performs high-throughput, precise phenotyping of maize cobs using convolutional neural networks to enable quantitative analysis in genebank phenomics and crop genetics.

Key Features:

• Mask R-CNN segmentation: Uses a Mask R-CNN model for image segmentation and phenotyping of maize cobs.
• Benchmarking: Was compared against the Felzenszwalb algorithm and Window-based CNNs.
• Convolutional neural networks (CNNs): Leverages CNN architectures for image analysis and feature extraction.
• High-throughput pipeline: Integrates the deep learning model into a pipeline capable of segmenting both maize cobs and rulers within images.
• Phenotypic trait extraction: Provides automated quantitative analysis of eight key phenotypic traits: diameter, length, ellipticity, asymmetry, aspect ratio, and average RGB values for cob color.
• Statistical framework for model updating: Implements a statistical framework to identify optimal training parameters and support iterative model updating.
• Rapid adaptation / minimal retraining: Supports updating the initial Mask R-CNN model with only 10–20 additional images to adapt to new cob image types.
• Performance metric: Demonstrated segmentation accuracy with a correlation coefficient (r) of 0.99.

Scientific Applications:

• Genebank phenomics: Enables large-scale phenotyping of genebank accessions to catalog morphological variation.
• Plant breeding and trait discovery: Provides quantitative cob trait measurements useful for breeding programs and trait association studies.
• Analysis of genebank variation: Was applied to 19,867 maize cobs extracted from 3,449 images representing 2,484 accessions from the Peruvian maize genebank.
• Multivariate clustering of accessions: Facilitates identification of phenotypically homogeneous and heterogeneous accessions via multivariate clustering.
• Studies of crop diversity: Suitable for comparing native landraces and modern varieties to quantify diversity in size, shape, and color.

Methodology:

Computational methods explicitly include Mask R-CNN for segmentation, comparison with the Felzenszwalb algorithm and Window-based CNNs, segmentation of maize cobs and rulers, a statistical framework for iterative model updating, and model updating with 10–20 additional images.