Aerenchyma segmentation in stems of rice seedlings based on Res2-TransUNet

The rice stem, which connects leaves and roots, is essential for nutrient transport and mechanical support. Aerenchyma tissues in the stem comprise air-filled cavities that are closely associated with phenotypic traits such as stem strength, lodging resistance, and stress adaptability. However, quantitative characterization of aerenchyma (e.g., cavity size and spatial distribution) is constrained by low segmentation accuracy, high dataset acquisition cost, and limited robustness of existing approaches. In this study, a low-cost rice seedling cross-sectional image dataset was acquired using a confocal microscopy platform, and a deep learning model, Res2-TransUNet, was developed for joint segmentation of aerenchyma tissues and stems. Res2-TransUNet is constructed by replacing the original backbone of TransUNet with Res2Net to strengthen multi-scale feature perception for complex aerenchyma morphology. On the dataset established in this work, Res2-TransUNet achieves an MIOU of 95.8%, an MPA of 97.71%, and an accuracy of 98.59%, improving MPA by 0.68, 0.23, 9.95, and 26.1 percentage points compared with TransUNet, UNet, DeepLabv3+, and PSPNet, respectively. Based on the predicted masks, CountNonZero was employed to obtain pixel counts of aerenchyma and stem regions, which were then combined with confocal microscope magnification to compute the aerenchyma area, stem cross-sectional area, and cavity-stem ratio (defined as the ratio of aerenchyma area to stem cross-sectional area). Linear regression against manual measurements yields R 2 values greater than 0.976, indicating that the proposed pipeline can provide reliable technical support for phenotypic analysis of internal tissues in rice seedlings.