Automated 3D wheat tissue analysis using x-ray CT and deep learning

Abstract

Understanding wheat grain internal structures is critical for improving quality, pest resistance, and breeding efficiency. While X-ray computed tomography (CT) enables non-destructive 3D imaging, existing segmentation methods rely on manual intervention, introducing inefficiency and subjectivity. This study introduces the Residual Depthwise Separable Convolution and Vision Mamba U-Net (RDVM-UNet), an automated framework combining Depthwise Separable Convolution (DSConv) for efficient local feature extraction and Vision Mamba for global contextual modeling. Trained over 200 iterations, the model achieved a mean Intersection over Union (mIoU) of 95.4 % in segmenting wheat tissues (epidermis, embryo, endosperm). Validation across 10 varieties demonstrated robust generalizability (mIoU is 94.78 %) and rapid processing (9.65 s/grain). The framework generated 3D reconstructions, enabling precise quantification of morphological parameters (volume, surface area) critical for analyzing genetic-environmental-morphological relationships. By establishing a non-destructive, high-throughput pipeline, this work advances precision breeding, functional genomics, and trait optimization in cereal crops. RDVM-UNet bridges computational imaging and agricultural science, offering scalable solutions for crop phenotyping and quality enhancement.