AI-powered XFruitSeg revolutionizes CT imaging for fruit phenotyping

By combining advanced neural architectures with a novel dataset, this technology achieves high-precision in mapping fruit anatomy, enabling nondestructive, high-resolution analysis of diverse species.

Fruits are among the most important plant organs, serving as critical sources of nutrition and economic value. Their development is governed by intricate genetic and environmental factors, producing remarkable diversity in structure and composition. Identifying the internal traits of fruits is vital for understanding complex genetic and physiological processes. While existing imaging methods such as visible light, hyperspectral, or Raman spectroscopy capture external traits effectively, they fall short in analyzing internal tissues. CT imaging offers nondestructive insights into fruit interiors, detecting features like cavities, bruising, porosity, and internal browning. Yet, accurate segmentation of CT images remains a bottleneck: threshold-based and conventional models often fail when tissues differ subtly in contrast or morphology. These challenges call for specialized computational tools tailored to the unique characteristics of plant fruits.

A study (DOI: 10.1016/j.plaphe.2025.100055) published in Plant Phenomics on 15 May 2025 by Lejun Yu & Qian Liu’s team, Hainan University, provides a highly accurate, robust, and generalizable solution for nondestructive CT image segmentation of diverse fruit tissues, thereby advancing plant phenotyping and biological research.

Using a controlled and reproducible setup, the study benchmarked the proposed XFruitSeg against four strong baselines—UNet, DeepLabv3+, SegFormer, and SegNext—implemented in MMSegmentation and trained under identical conditions on Ubuntu 20.04 with Python 3.8/PyTorch 1.10.0, an Intel Xeon Gold 6130 CPU, and an NVIDIA Tesla V100 (CUDA 11.3). Each model, including a pared-down baseline (XFruitSeg_base) lacking the contour-feature learning (CFL) branch and the composite loss, was optimized per its original paper; XFruitSeg used AdamW (weight decay 0.05, lr 0.0002). Three CT datasets (orange, mangosteen, durian; 1,600 images each) were split 6:2:2 for train/val/test with no overlap. Performance was assessed with pixel accuracy (PA), Dice, and intersection-over-union (IoU); category-wise Dice/IoU and the means (mDice, mIoU) captured both class balance and spatial agreement. Training dynamics over 350 epochs (42,000 iterations) showed rapid loss drops by ~50 epochs and stabilization by ~330 epochs without overfitting; XFruitSeg_base, SegFormer, and DeepLabv3+ exhibited the strongest convergence, with XFruitSeg’s multi-loss curves converging swiftly and synchronously (fastest on orange, slowest on durian, reflecting dataset complexity). In head-to-head tests, XFruitSeg delivered the best overall accuracy across most categories and the smallest parameter count among competitors. On oranges, all models segmented the high-contrast pulp well (>99% Dice; >98% IoU), but XFruitSeg achieved the highest mDice/mIoU (95.21%/91.09%) via notable gains on the harder exocarp and cavity. On low-contrast mangosteen, XFruitSeg improved Dice by 3.17% (exocarp) and 2.55% (cavity) and IoU by 4.59% and 4.43%, respectively, reaching 93.24% mDice and 87.91% mIoU. On structurally complex durian, it excelled on pericarp and pulp (Dice 98.86% and 96.81%; IoU 97.74% and 94.28%) and raised performance on small, scattered cavities, yielding 94.70% mDice and 90.35% mIoU overall. Ablations confirmed RepLKNet plus multiscale skips as a strong backbone (XFruitSeg_base), with the composite loss (LGD + LMS-SSIM) and CFL together best addressing class imbalance and sharpening tissue boundaries.

The ability to nondestructively monitor fruit interiors with such high accuracy has wide-ranging implications. For breeders, XFruitSeg enables precise phenotyping of traits linked to yield, quality, and stress resistance. In postharvest science, it offers a powerful tool to detect hidden bruising, internal cavities, or porosity changes that compromise storage and marketability. Food scientists and supply chains could apply this approach to ensure fruit quality during processing and transport, reducing waste and improving consumer confidence.

###

References

DOI

10.1016/j.plaphe.2025.100055

Original URL

https://doi.org/10.1016/j.plaphe.2025.100055

Funding informantion

This work was supported by the National Key R&D Program of China (2023ZD04073), Sanya Yazhou Bay Science and Technology City (SCKJ-JYRC-2023-25), the National Natural Science Foundation of China (32360116), and the Research Project of the Collaborative Innovation Center of Hainan University (XTCX2022NYB01).

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.