Deep Learning Transforms 3D Imaging of Fruit Tissue Structures

Researchers at KU Leuven have developed a groundbreaking deep learning model that significantly improves the accuracy of analyzing the three-dimensional (3D) microstructure of fruit tissues, specifically apples and pears. Published in the journal Plant Phenomics on July 5, 2025, this study demonstrates how advanced technology can enhance our understanding of plant anatomy and its implications for agriculture and food science.

Traditional microscopy techniques for studying plant tissues often involve extensive preparation and yield limited views of the samples. Although X-ray micro-computed tomography (micro-CT) has emerged as a non-destructive imaging method, it poses challenges in accurately quantifying tissue morphology due to overlapping features and low contrast in images. Previous segmentation methods frequently struggled to differentiate between various types of cells, including parenchyma, vascular tissues, and stone cell clusters.

To address these issues, Pieter Verboven and his team at KU Leuven utilized a comprehensive deep learning framework that automates the labeling and quantification of plant tissue architecture. Their model employs a 3D panoptic segmentation approach, combining the capabilities of the 3D extension of Cellpose and a 3D Residual U-Net architecture. This innovative model effectively performs both instance and semantic segmentation, allowing it to predict gradient fields in three dimensions and classify different tissue types.

The deep learning model was trained on datasets from apples and pears, incorporating synthetic data augmentation techniques such as morphological dilation and erosion, grey-value assignment, and Gaussian noise addition. When benchmarked against a 2D instance segmentation model and a marker-based watershed algorithm, the 3D model achieved impressive results. The model reached an Aggregated Jaccard Index (AJI) of 0.889 for apples and 0.773 for pears, outperforming the 2D model’s scores of 0.861 and 0.732, respectively.

The accuracy of the model was further validated through visual assessments, demonstrating precise detection of vascular bundles in apple varieties such as ‘Kizuri’ and ‘Braeburn’ and realistic segmentation of stone cell clusters in pears like ‘Celina’ and ‘Fred.’ The model achieved a Dice Similarity Coefficient (DSC) of up to 0.90 in some cases.

Despite these advancements, the research team noted that further data augmentation did not enhance performance, suggesting that dataset imbalance and domain shifts might have affected the outcomes. Morphometric analysis confirmed the model’s effectiveness, with vasculature widths ranging from 70 to 780 μm and variable dimensions and sphericity of stone cell clusters (ranging from 0.68 to 0.74).

This deep learning model represents a significant step forward in the field of plant science, offering a non-destructive tool for researchers to explore how microscopic structures affect water, gas, and nutrient transport. By improving the efficiency of “human-in-the-loop” analysis, the model not only reduces manual labor but also enhances tissue characterization accuracy.

In the context of fruit research, the model provides insights into how cellular arrangements influence texture, storability, and susceptibility to physiological disorders such as browning or watercore. More broadly, this technology establishes a scalable framework to study tissue development, ripening, and stress responses across various crops.

The compatibility of this deep learning model with standard X-ray micro-CT instruments positions it as an accessible solution for integrating artificial intelligence into research focused on plant anatomy and food science. As the field continues to evolve, the implications of this work may lead to significant advancements in agricultural practices and food preservation techniques.

Funding for this research was provided by the Research Foundation – Flanders (FWO), under grant number S003421N, as part of the SBO project FoodPhase.