A deep learning-based micro-CT image analysis pipeline for nondestructive quantification of the maize kernel internal...

Area of Science:

• Agricultural Science
• Biotechnology
• Image Analysis

Background:

• Accurate segmentation of maize kernel vitreous and starchy endosperm is crucial for texture analysis but challenging due to low contrast and blurred edges in CT images.
• Existing methods struggle with segmentation accuracy and jagged outcomes, hindering detailed internal structure analysis.

Purpose of the Study:

• To develop a deep learning-based pipeline for high-quality segmentation of maize kernel internal structures from CT images.
• To improve the accuracy and efficiency of analyzing maize kernel texture for breeding and processing applications.

Main Methods:

• A CT image analysis pipeline was developed, starting with batch scanning and Canny algorithm segmentation of individual kernels.
• A modified U-Net architecture (CSFTU-Net) was proposed, incorporating attention mechanisms (CBAM, SE) and a focal-Tversky loss with a boundary smoothing term.
• Phenotype parameters, including volumes of kernel, vitreous endosperm, and starchy endosperm, were extracted using a segmented mask-based method.

Main Results:

• The CSFTU-Net model demonstrated significantly improved segmentation of vitreous and starchy endosperm in maize kernels.
• The pipeline enabled nondestructive quantification of key internal texture parameters (V, VV, SV, VV/V, SV/V).

Conclusions:

• The proposed deep learning pipeline offers a robust and accurate method for analyzing the internal structure of maize kernels.
• This approach provides valuable insights for nondestructive texture quantification, supporting advancements in maize breeding and food processing.