CitrusGAN: AI revolutionizes 3D fruit phenotyping with sparse X-ray imaging

Unlike traditional computed tomography (CT) scanning that requires expensive machines and time-consuming data acquisition, CitrusGAN produces high-quality, high-resolution models in seconds using only six X-ray views. The generated 3D models capture both external and internal fruit structures, enabling accurate measurement of traits such as peel thickness, edible rate, and number of segments.

Citrus is one of the world’s most widely cultivated crops, producing over 150 million tons annually. To improve fruit quality, flavor, and stress resilience, breeders rely on phenotypic analysis—the study of visible and structural traits linked to genetic and environmental factors. However, manual phenotyping is slow and error-prone, and most existing image-based or LiDAR methods can only capture surface features, not internal traits. Conventional CT imaging can reveal interior structures but remains costly and inefficient. Based on these challenges, researchers sought to create a rapid, affordable, and non-destructive 3D reconstruction technique to visualize and analyze both the internal and external morphology of fruit with unprecedented efficiency.

A study (DOI: 10.1016/j.plaphe.2025.100082) published in Plant Phenomics on 26 June 2025 by Yaohui Chen’s team, Huazhong Agricultural University, marks a significant leap toward non-destructive, high-throughput phenotyping, potentially transforming fruit breeding, quality control, and agricultural automation.

In this study, the researchers employed a generative deep learning method to reconstruct three-dimensional (3D) citrus computed tomography (CT) models from sparse-view X-ray images, testing reconstruction quality across different input configurations. They compared models trained on 2, 4, and 6 input views to evaluate the effect of view number on output fidelity. The results revealed that all reconstructed models achieved structural similarity index (SSIM) values above 0.9, indicating high resemblance to real CT data, while the 6-view model reached the highest performance with a 1.2 dB increase in peak signal-to-noise ratio (PSNR) and an SSIM of 0.92. Qualitative analyses showed that the 6-view model effectively reproduced external contours and internal pulp structures with greater clarity, preserving detailed segment boundaries and shape completeness that closely matched the real CT slices. Minor discrepancies were mainly observed in low-density tissues, such as granulated pulp or inner peel layers, which appeared darker and less distinct due to their reduced water content. For phenotypic validation, 77 fruit samples were analyzed to extract eight structural traits—volume, surface area, height, width, length, peel thickness, edible rate, and segment number—from both real and generated CT models. A strong correlation (R² > 0.95) was observed for most parameters, confirming the method’s high quantitative reliability, although peel thickness and edible rate showed greater variation due to subtle boundary blurring. Overall, the findings demonstrate that the deep learning model can efficiently and accurately reconstruct both external and internal fruit morphology from minimal X-ray data, achieving high-throughput precision phenotyping suitable for practical breeding and quality evaluation.

By drastically reducing imaging costs and computation time, CitrusGAN represents a transformative step in agricultural automation. It enables breeders to evaluate hundreds of fruit samples non-destructively, accelerating the selection of desirable genotypes and improving yield and quality traits. Beyond breeding, the technology could be deployed in automated sorting and grading systems, ensuring consistency in commercial fruit production. The method’s ability to visualize internal features also opens new possibilities for detecting hidden defects and monitoring fruit ripeness, providing real-time insights for both research and industry.

###

References

DOI

10.1016/j.plaphe.2025.100082

Original URL

https://doi.org/10.1016/j.plaphe.2025.100082

Funding information

This work is supported by the National Science Foundation of China (32302206) and the Fundamental Research Funds for the Central Universities, China (2662024SZ002).

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.