Panoptic segmentation for complete labeling of fruit microstructure in 3D micro-CT images with deep learning

Area of Science:

• Plant biology
• Biophysics
• Medical imaging

Background:

• Plant organ function relies on 3D tissue morphology for transport.
• Quantifying microstructures like vascular bundles and stone cells is difficult.
• Current methods require extensive sample preparation or are less accurate.

Purpose of the Study:

• To develop an automated 3D deep learning model for plant tissue microstructure characterization.
• To improve the accuracy and speed of analyzing X-ray micro-CT images of fruit tissue.
• To explore training strategies for enhancing segmentation quality.

Main Methods:

• A 3D deep learning panoptic segmentation model was developed, combining semantic and instance segmentation.
• The model was trained and evaluated on X-ray micro-CT images of apple and pear fruit tissue.
• Various training datasets and data augmentation techniques, including synthetic data, were explored.

Main Results:

• The 3D panoptic segmentation model achieved high accuracy (Aggregated Jaccard Index 0.89 for apple, 0.77 for pear).
• It successfully segmented vascular bundles (DSC 0.51 apple, 0.79 pear) and stone cell clusters (DSC 0.81).
• No tested augmentation or dataset strategy improved performance over the standard dataset.

Conclusions:

• The 3D panoptic segmentation model provides a highly automated protocol for plant tissue analysis.
• It enables accurate morphometric quantification from native X-ray micro-CT images without contrast labeling.
• The method significantly accelerates conventional analysis, offering a powerful tool for plant science research.