Research Article | Open Access

Volume 2025 |Article ID 100022 | https://doi.org/10.1016/j.plaphe.2025.100022

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

Juan Wang,^1,2,3,5 Si Yang,^2,3,5 Chuanyu Wang,^2,3 Weiliang Wen,^2,3 Ying Zhang,^2,3 Gui Liu,¹ Jingyi Li,⁴ Xinyu Guo ,^2,3 Chunjiang Zhao ^1,2

¹College of Information, Shanghai Ocean University, Shanghai, 201306, China
²Information Technology Research Center, Beijing Academy of Agriculture and Forestry Sciences, Beijing, 100097, China
³Beijing Key Lab of Digital Plants, National Engineering Research Center for Information Technology in Agriculture, Beijing, 100097, China
⁴College of Computer Science and Engineering, Southwest Minzu University, Chengdu, Sichuan, 610225, China
⁵These authors contributed equally to this work.

Abstract

Identifying and segmenting the vitreous and starchy endosperm of maize kernels is essential for texture analysis. However, the complex internal structure of maize kernels presents several challenges. In CT (computed tomography) images, the pixel intensity differences between the vitreous and starchy endosperm regions in maize kernel CT images are not distinct, potentially leading to low segmentation accuracy or oversegmentation. Moreover, the blurred edges between the vitreous and starchy endosperm make segmentation difficult, often resulting in jagged segmentation outcomes. We propose a deep learning-based CT image analysis pipeline to examine the internal structure of maize seeds. First, CT images are acquired using a multislice CT scanner. To improve the efficiency of maize kernel CT imaging, a batch scanning method is used. Individual kernels are accurately segmented from batch-scanned CT images using the Canny algorithm. Second, we modify the conventional architecture for high-quality segmentation of the vitreous and starchy endosperm in maize kernels. The conventional U-Net is modified by integrating the CBAM (convolutional block attention module) mechanism in the encoder and the SE (squeeze-and-excitation attention) mechanism in the decoder, as well as by using the focal-Tversky loss function instead of the Dice loss, and the boundary smoothing term is weighted as an additional loss term, named CSFTU-Net. The experimental results show that the CSFTU-Net model significantly improves the ability of segmenting vitreous and starchy endosperm. Finally, a segmented mask-based method is proposed to extract phenotype parameters of maize kernel texture, including the volume of the kernel (V), volume of the vitreous endosperm (VV), volume of starchy endosperm (SV), and ratios over their respective total kernel volumes (VV/V and SV/V). The proposed pipeline facilitates the nondestructive quantification of the internal structure of maize kernels, offering valuable insights for maize breeding and processing.